US 20020149585 A1 Abstract An integrated system generates a model of a three-dimensional object. A scanning laser device scans the three-dimensional object and generates a point cloud. The points of the point cloud each indicate a location of a corresponding point on a surface of the object. A first model is generated, responsive to the point cloud, that generates a first model representing constituent geometric shapes of the object. A data file is generated, responsive to the first model, that can be inputted to a computer-aided design system.
Claims(16) 1. A method of manually separating from a plurality of clouds of points, representing three-dimensional features in a scene, a subset of the points that represents a desired feature in the scene, the method comprising:
selecting all the point clouds that include at least some data points representing the desired feature; and changing a view of the clouds and drawing a polygonal lasso to refine a selected subset of points to be included in a point sub-cloud and repeating the refining as many times as required to obtain the desired sub-cloud. 2. A method for automatically segmenting a scan field of a scene into subsets of points that represent different surfaces in the scene, comprising the steps of:
separating the scan field into a depth grid that includes depth information for scanned points of surfaces in the scene and a normal grid that includes an estimate of a normal to scanned points of the surfaces; convolving the depth information of the depth grid to generate a depth rating image whose values represent a gradient of depth change from one scanned point to another scanned point in the scene; convolving the components of the normal grid to generate a scalar value for each component for each point of the normal grid, for each point of the normal grid, determining from the scalar values for the components of that particular point a gradient of the normal at that point, wherein the gradients determined for the points of the normal grid collectively constitute a normal rating image; converting the depth rating image to a binary depth image using a recursive thresholding technique; converting the normal rating image to a binary normal image using a recursive thresholding technique; combining the binary depth image and the binary normal image to determine a single edge image; and grouping subsets of non-edge points as belonging to corresponding surfaces of the scene. 3. A method as recited in determining the type of geometric primitive that would best first each group of points; and
fitting the geometric primitive to the data points.
4. A method as recited in 5. A method for fitting a point cloud representing a corner, comprising:
determining a fit of three planes to the points of the point cloud and creating the planes for a model; determining the three lines at the intersection of pairs of planes and creating the lines for the model; and determining the vertex point at the intersection of the three planes and creating a vertex point for the model. 6. A method for modeling a three-dimensional scene, comprising:
generating a plurality of points that each represent a point on a surface of the scene; determining a best fit of a cylinder for a group of the points using surface normal estimates and global error minimization. 7. A method for modeling a three-dimensional scene, comprising:
generating a plurality of points that each represent a point on a surface of the scene; determining a best fit of a cylinder for a group of the points using a quadric surface fit and global error minimization. 8. A method for modeling a three-dimensional scene, comprising:
generating a plurality of points that each represent a point on a surface of the scene; determining a best fit of a sphere for a group of the points using a quadric surface fit and global error minimization. 9. A method for modeling a three-dimensional scene, comprising:
generating a plurality of points that each represent a point on a surface of the scene; determining a best fit quadric surface for a group of points; and determining which geometric primitive of a plurality of the family described by the quadric surface best fits the group of points. 10. A method for merging two geometric primitives of the same type to form a single geometric primitive of the type, comprising:
creating a new group of points by combining the points used to originally fit each of the two primitives; and fitting the new geometric primitive using any appropriate fitting technique and the newly generated point group with points from each of the original primitives. 11. A method of registering a first model, consisting of a plurality of points and geometric primitives and having a first coordinate system, with a second model, consisting of a plurality of points and geometric primitives and having a second coordinate system, comprising:
identifying by a user common features of the first and second scenes identifying a transformation between coordinate systems that is responsive to the identification; and transforming the objects of the second model so that they use the first coordinate system. 12. A method of warping, comprising:
selecting one or more models represented by a plurality of point clouds and geometric primitives; specifying constraints on the locations of any number of points or geometric primitives; creating an artificial volume that surrounds the points and geometric primitives in each view and assigning mechanical material characteristics to the surrounding volume; computing a minimum energy configuration for the material in the surrounding volume in which the points or geometric primitives are embedded such that the configuration satisfies all applied constraints; and displacing the points and primitives in accordance with the computed minimum energy configuration of the surrounding volume of material. 13. The method of 14. An apparatus for acquiring three dimensional information from a remote object comprising:
a scanning laser module for measuring position information of the object; a video module for capturing image information from the object; and a processor for rendering a model of the object which includes the position information and the image information. 15. An apparatus as recited in 16. An approach as recited in Description [0001] The present invention relates generally to systems that document the geometry and other attributes of objects in three dimensions and, specifically, to a system that employs a scanning lidar (range finding laser) to quickly and accurately sense the position in three-dimensional space of selected points on the surface of an object to generate a point cloud which represents the sensed positions of the selected points; that recognizes geometric shapes represented by groups of points in the point cloud, and that generates a model that represents these geometric shapes. The model may be transformed into a further model usable by computer-aided design (CAD) tools, including conventional CAD tools. [0002] Mapping the geometry (shape, dimensions and location) and other attributes (e.g., color, texture and reflectance intensity) of complex real objects (whether small components such as small mechanical parts or large objects such as buildings and sites) has conventionally been a tedious and time consuming process. That is, such measurement have traditionally been performed manually. In addition, transforming these measurements into drawings or computer models required manual drafting or input into a CAD system for the production of the drawing or computer models. [0003] Recently innovations have endeavored to simplify this process, but all have fallen short of achieving full integration, automation, precision, speed and range. For example, in the building industry, mapping a structure conventionally requires three basic steps: [0004] 1. Field data gathering [0005] 2. Data reduction and preparation [0006] 3. Drafting and CAD [0007] The field data gathering step is performed by a team of surveyors who manually measure and record dimensions of pertinent components of the structure such as walls, ceilings, beams, columns, doors, windows, fixtures, pipes, conduits and equipment. The surveyors attempt to determine the geometry of the components as well as the relative location of the components in the structure. [0008] The surveyors recorded the data in a field notebook. The field-collected data is then organized and reduced to tables and organized sketches, and a CAD operator or drafter utilizes these tables to generate final drawings or models. [0009] This process is labor intensive, time consuming, and error prone. In addition, using traditional surveying methods, the number of points which can actually be measured is very limited, due to the high cost of acquiring each point in terms of time and effort. Furthermore, if it is desired to acquire color, texture and other attribute information, additional field notes must be taken (e.g., still photographs and video). [0010] Recently, the field step has been somewhat automated by using a laser ranging device built into or mounted on an electronic theodolite. Precision reflection targets (retro reflectors) are placed at the locations of the object for which measurements are desired. Then, the laser ranging device obtains a precise measurement of the distance between the instrument and the target, which the theodolite provides an accurate indication of the horizontal and vertical angle offsets to the point relative to a given coordinate system. The distance and angle data are either recorded automatically on a magnetic device connected to the instrument or are reduced within the instrument to Cartesian coordinates relative to the instrument axes. This procedure is then repeated as many times as necessary to map a desired number of points of the object. The collected coordinates data can then be plotted directly on a CAD system. [0011] Unfortunately, the plot is of little practical use since it does not indicate the object geometry. Moreover, because of the requirement for retro reflectors which must be manually placed, and because of the relatively long time per reading required by the laser range finder, the gathering of sufficient points to describe most objects is very labor intensive, time consuming and error prone. [0012] Another known field gathering data process employs stereo photography and aerial photogrammetry. That is, stereoscopic images are taken of the objects and the resulting stereo photographs are registered either manually or using computerized techniques to reproduce the relative location of the camera picture plane location at the time each photograph was taken. The data reduction and preparation step is performed manually by a specially trained operator. Specifically, with the aid of specially mounted stereoscopic viewing lenses, the operator digitizes the coordinates of a sufficient number of points to allow the definition of the objects using the stereo photographs. Again, the digitized data is input into a CAD system or is manually drawn on paper. [0013] The present invention is an integrated system for generating a model of a three-dimensional object. A scanning laser device scans the three-dimensional object and generates a point cloud. The points of the point cloud each indicate a location of a corresponding point on a surface of the object. A first model is generated, responsive to the point cloud, representing constituent geometric shapes of the object. A data file is generated, responsive to the first model, that can be inputted to a computer-aided design system. [0014] The subject invention further includes a method of controlling the timing of output pulses from a laser for use in a device which requires scanning of the laser output, wherein each output pulse is generated in response to a pump pulse comprising the steps of: monitoring the time delay between the initiation of the pump pulses and the subsequent generation of the associated output pulses; predicting the time delay between the initiation of next pump pulse and the associated output pulse based on the monitored time delays and; initiating the next pump pulse at a time selected to insure the output pulse is-generated at a time to permit proper positioning of the laser output during the scan of the beam. [0015] The present invention further includes a method of manually separating from a plurality of clouds of points, representing three-dimensional features in a scene, a subset of the points that represents a desired feature in the scene, the method comprising: selecting all the point clouds that include at least some data points representing the desired feature; and changing a view of the clouds and drawing a polygonal lasso to refine a selected subset of points to be included in a point sub-cloud and repeating the refining as many times as required to obtain the desired sub-cloud. [0016] The present invention further includes a method for automatically segmenting a scan field of a scene into subsets of points that represent different surfaces in the scene, comprising the steps of: separating the scan field into a depth grid that includes depth information for scanned points of surfaces in the scene and a normal grid that includes an estimate of a normal to scanned points of the surfaces; convolving the depth information of the depth grid to generate a depth rating image whose values represent a gradient of depth change from one scanned point to another scanned point in the scene; convolving the components of the normal grid to generate a scalar value for each component for each point of the normal grid; for each point of the normal grid, determining from the scalar values for the components of that particular point a gradient of the normal at that point, wherein the gradients determined for the points of the normal grid collectively constitute a normal rating image; converting the depth rating image to a binary depth image using a recursive thresholding technique; converting the normal rating image to a binary normal image using a recursive thresholding technique; combining the binary depth image and the binary normal image to determine a single edge image; and grouping subsets of non-edge points as belonging to corresponding surfaces of the scene. [0017] The method can further include the steps of determining the type of geometric primitive that would best first each group of points; fitting the geometric primitive to the data points; and intersecting adjacent planar regions in the scene. [0018] The subject matter further includes a method for fitting a point cloud representing a corner, comprising: determining a fit of three planes to the points of the point cloud and creating the planes for a model; determining the three lines at the intersection of pairs of planes and creating the lines for the model; [0019] and determining the vertex point at the intersection of the three planes and creating a vertex point for the model. [0020] The subject invention further includes a method for modeling a three-dimensional scene, comprising: generating a plurality of points that each represent a point on a surface of the scene; determining a best fit of a cylinder for a group of the points using surface normal estimates and global error minimization. [0021] The subject invention further includes a method for modeling a three-dimensional scene, comprising: generating a plurality of points that each represent a point on a surface of the scene; determining a best fit of a cylinder for a group of the points using a quadric surface fit and global error minimization. [0022] The subject invention further includes a method for modeling a three-dimensional scene, comprising: generating a plurality of points that each represent a point on a surface of the scene; determining a best fit of a sphere for a group of the points using a quadric surface fit and global error minimization. [0023] The subject invention fisher includes a method for modeling a three-dimensional scene, comprising: generating a plurality of points that each represent a point on a surface of the scene; determining a best fit quadric surface for a group of points; and determining which geometric primitive of a plurality of the family described by the quadric surface best fits the group of points. [0024] The subject invention flier includes a method for merging two geometric primitives of the same type to form a single geometric primitive of the type, comprising: creating a new group of points by combining the points used to originally fit each of the two primitives; and fitting the new geometric primitive using any appropriate fitting technique and the newly generated point group with points from each of the original primitives. [0025] The subject invention further includes a method of registering a first model, consisting of a plurality of points and geometric primitives and having a first coordinate system, with a second model, consisting of a plurality of points and geometric primitives and having a second coordinate system, comprising: identifying by a user common features of the first and second scenes; identifying a transformation between coordinate systems that is responsive to the identification; and transforming the objects of the second model so that they use the first coordinate system. [0026] The subject invention further includes a method of warping, comprising: selecting one or more models represented by a plurality of point clouds and geometric primitives; specifying constraints on the locations of any number of points or geometric primitives; creating an artificial volume that surrounds the points and geometric primitives in each view and assigning mechanical material characteristics to the surrounding volume; computing a minimum energy configuration for the material in the surrounding volume in which the points or geometric primitives are embedded such that the configuration satisfies all applied constraints; and displacing the points and primitives in accordance with the computed minimum energy configuration of the surrounding volume of material. In the latter method, the constraints can be specified to eliminate closure errors. [0027] The subject invention further includes an integrated system for generating a model of a three-dimensional scene, comprising: a scanning laser device that scans the three dimensional scene with pulsed laser beam wherein the pulses of light last less than 1 nanosecond with up to 0.2 μJ in each pulse and measures the time delay, with a resolution of 30 psec or less, between each emitted pulse and a corresponding pulse that is returned from a surface of the scene and wherein said scanning laser device further tracks and measures the angular orientation of the beam during the scan; and means for generating a point cloud based upon the measured time delays and angle measurements, the point cloud comprising a plurality of data points that each represents a location of a corresponding point on the surface. [0028] The subject invention further includes a system for calibrating the measuring electronics in a device which requires monitoring the time of flight of the output pulses from a laser comprising: a single mode optical fiber with one end thereof positioned to receive the output pulses of the laser, said single mode optical fiber having a known length; a detector positioned at one of the ends of the fiber for monitoring when the pulses exit the fiber and generating a signal in response thereto, said signal being passed through the measuring electronics; and a processor for calculating a theoretical length of the fiber based on the detection of the pulse exiting the fiber and comparing that calculated length with known length of the fiber to calibrate the measuring electronics. [0029] The optical fiber can include partial reflectors located at each end thereof so that for each laser pulse entering the fiber a train of pulses will exit the fiber and wherein said train of pulses are used to further calibrate the measuring electronics. [0030] The system can further include delay measurement electronics and wherein the train of pulses have a fixed delay therebetween whereby the monitoring of the train of pulses can be used to calibrate the delay electronics. [0031] The system can further include a means for varying the power of the pulses monitored by the detector and wherein said detector functions to generate a signal when the power of the detected light exceeds a predetermined threshold and wherein said processor functions to track the variation in the delay of the generation of the output signal by the detector as a function of the power of the output pulses, said processor further functioning to calibrate the measurement of the delay based on the measured power of successive pulses used for monitoring the time of flight. [0032] The subject invention further includes an apparatus for obtaining position information about surface points of a three dimensional object comprising: a laser for generating an output beam; scanning system for moving the laser beam over the object; monitoring system for automatically measuring the range to the object based on the measurement of the reflection of the laser beam, said monitor system also tracking and measuring the angular position of the laser beam, said monitoring system having a positional accuracy for each point in three dimensional space equal to or better than six millimeters at one standard deviation over a range of up to 100 meters. Each range measurement can be made in under 0.005 seconds. The laser can generate a pulsed output and the energy per pulse can be less than 0.2 micro joules and the average output power of the laser can be less than 1.0 milliwatts. [0033] The subject invention further includes an apparatus for measuring the distance to an object comprising: a laser for generating a beam of output pulses; a monitoring system for measuring the distance to the object based on the reflection of the laser beam, said monitoring system having an accuracy equal to or better than 6 millimeters at one standard deviation over its entire range of up to 100 meters and wherein each measurement can be made in less than 0.005seconds and wherein the laser has an energy per pulse of no more than 0.2 micro joules and an average power of no more than 1 milliwatt. If the object is provided with retro reflectors and where the range of operation is up to one mile. [0034] The subject invention further includes an apparatus for acquiring three dimensional information from a remote object comprising: a scanning laser module for measuring position information of the object; a video module for capturing image information from the object; and a processor for rendering a model of the object which includes the position information and the image information. [0035] The video image information can be collected in a spatially coincident manner with the measurement of position information. The video image information can be collected from points adjacent to the points where position information is obtained. [0036] The subject invention further includes an apparatus for obtaining positional information about surface points of a three dimensional object comprising: a scanning module for measuring three dimensional position information about an object; a video module for capturing and displaying image information from the object; and a processor operating with the scanning and video modules and permitting the use of the image information captured by said video module to aid in targeting the scanning module. The processor can function to specify a portion of the object to be targeted by the scanning module by dragging the image of an outline over the video image of the area to be targeted. [0037] The subject invention further includes an apparatus for obtaining positional information about surface points of a three dimensional object comprising: a scanning module for measuring three dimensional position information about an object; a video module for displaying image information obtained from the scanning module; a processor operating with the scanning and video modules and permitting the use of the image information displayed by said video module to further refine the targeting of the scanning module. [0038] The subject invention further includes an apparatus for obtaining positional information about surface points of a three dimensional object comprising: a scanning module for measuring three dimensional position information about an object, said scanning module including a laser for emitting a beam of visible radiation; and a processor for controlling the scanning module and wherein said laser can be manually positioned so that the visible beam will target the portion of the object to be scanned in response to a control signal from the processor. [0039] The subject invention further includes a system for calibrating the measuring electronics in a device which requires monitoring frequency changes in a light beam generated by a laser used to measure distance to an object, wherein said beam has frequency chirp imposed thereon comprising a single mode optical fiber with one end thereof positioned to receive light from the laser; a detector positioned to receive light which has traveled through and exited the fiber in combination which light from the laser which has not traveled through the fiber, said detector for monitoring the changes in the frequency of the combined beam; and processor for determining the linearity of the chirp on the beam based on uniformity of the frequency changes measured by the detector and using the result to calibrate the measuring electronics. [0040] The fiber can have a known length and includes a partial reflector on said one end and at least a partial reflector on the other end, and wherein light reflected from said one end of the fiber which has not traveled in the fiber is measured by the detector and wherein the processor further functions to calculate a theoretical length of the fiber based on the frequency changes measured by the detector and compares that calculated length with the known length of the fiber to calibrate the measuring electronics. [0041]FIG. 1 is a block diagram of a system in accordance with an embodiment of the invention. [0042]FIG. 1A shows the overall flow of how one may use an embodiment of the invention to scan an object, organize acquired points, fit geometric shapes to the organized point, manipulate the fitted geometric shapes, and display the resulting manipulated geometric shapes. [0043]FIG. 2 is a more detailed block diagram of the system of FIG. 1. [0044]FIGS. 3A and 3B show the physical arrangement of the FDV of the FIG. 1 system, and also shows how the FDV is coupled to the tripod by a fork mount. [0045]FIGS. 4A and 4B show an example coordinate system relative to the FDV of the FIG. 1 system. [0046]FIG. 5 is a block diagram of one embodiment of an FDV in accordance with the invention. [0047]FIG. 6 is a block diagram of the optical transceiver of the FIG. 5 FDV. [0048]FIG. 6A shows a dual mirror arrangement of the scanner shown in FIG. 6. [0049]FIG. 7 is a block diagram which shows an embodiment of the laser. [0050]FIG. 7A is a block diagram of an embodiment of the beam expander shown in FIG. 6. [0051]FIG. 8 shows an embodiment of the duplexer. [0052]FIG. 8A shows a partially-reflecting duplexer. [0053]FIG. 9 shows an embodiment of the window of the FIG. 8 duplexer. [0054]FIG. 10 is a flowchart that shows calculations performed by the FDV DSP. [0055]FIGS. 11A and 11B show a unidirectional scan pattern and a bidirectional scan pattern, respectively. [0056]FIG. 12 is a block diagram of an embodiment of the FDV processor. [0057]FIG. 13 is a block diagram of example circuitry for determining a desired position of an FDV mirror. [0058]FIG. 14 is a block diagram of an example signal conditioning and energy integration circuit of the timing circuit shown in FIG. 12. [0059]FIG. 15 is a detailed block diagram of the system of FIG. 1. [0060]FIGS. 16A and 16B show two windows used to operate the CGP. [0061]FIGS. 17A and 17B show a targeting box and a point cloud. [0062]FIG. 18 shows a point cloud from the surface of a horse sculpture. [0063]FIG. 19 shows the point cloud of FIG. 18 color mapped with the laser return intensities. [0064]FIG. 20 shows a cloud of points from a corner feature. [0065]FIG. 21 shows the cloud of points of FIG. 20 and a polygonal lasso used for manual segmentation. [0066]FIG. 22 shows the cloud of points of FIG. 20 segmented into four subgroups, three subgroups on the surfaces of planes and a subgroup of edge points that are not part of the plane. [0067]FIG. 23 shows the cloud of points of FIG. 20 rendered as a triangulated mesh. [0068]FIG. 24 shows the corner feature of FIG. 20 with planes fit to the groups of cloud points. [0069]FIG. 25 shows a point cloud from the surface of a cylinder. [0070]FIG. 26 shows a cylinder primitive that was fit to the points shown in FIG. 25. [0071]FIG. 27 shows a cloud of points from the surfaces on a piping system. [0072]FIG. 28 shows cylinder primitives that were fit to the points shown in FIG. 27. [0073]FIG. 29 shows the completed piping model, after extending pipes and adding elbows. [0074]FIG. 30 shows the result of a corner fit, giving three planes, three lines, and a vertex. [0075]FIG. 31 shows a cylinder primitive in a scene. [0076]FIG. 32 shows the cylinder form FIG. 31 extended to meet adjacent objects. [0077]FIG. 33 shows a cloud of points from the surface from a variety of objects. [0078]FIG. 34 shows a model containing primitives that were fit to the points shown in FIG. 33. [0079]FIG. 35 shows configuration of a frequency adjustable laser. [0080]FIG. 36 shows block diagram of conventional FM chirp lidar. [0081]FIG. 37 shows block diagram of self-calibrating FM chirp lidar. [0082]FIG. 38 illustrates the relative timing at which a large and a small pulse cross a predetermined threshold. [0083]FIG. 39 illustrates one circuit for measuring pulse energy. [0084]FIG. 40 illustrates another circuit for measuring pulse energy. [0085] A. Overview [0086] 1. Overall System [0087]FIG. 1 is a block diagram that illustrates the invention in its broadest aspect. Referring to FIG. 1, a Field Digital Vision (FDV) module [0088] A Computer Graphics Perception (CGP) module [0089]FIG. 1A shows the overall flow of how one may use an embodiment of the invention to scan an object, organize acquired points, fit geometric shapes to the organized point, manipulate the fitted geometric shapes, and display the resulting manipulated geometric shapes. [0090] 2. FDV Module Overview [0091] Referring to FIG. 2, the FDV [0092] In response to user input relative to the signal that represents the acquired video image, the CGP [0093] In addition, the narrow angle CCD camera of the video system [0094] 3. CGP Module Overview [0095] Referring still to FIG. 2, the CGP [0096] i. FDV Control [0097] The CGP [0098] ii. Model Generation [0099] Each data point in the point cloud [0100] B. Details [0101] 1. FDV Module Detail [0102]FIG. 5 is a block diagram of one embodiment of an FDV [0103] If the laser beam is quasi-CW, always on with either intensity modulation (AM) or wavelength modulation (FM), the distance to the object [0104] In the time-of-flight embodiment, the laser is preferably of the type disclosed in U.S. Pat. Nos. 5,132,977; 5,386,427; and 5,381,431, assigned to Massachusetts Institute of Technology. In particular, the beam generated by such a laser has special properties such as being capable of producing pulse widths less than 1 nsec. [0105] A particular embodiment of the laser which has been used is particularly suitable for precision lidar since: [0106] 1. The short pulsewidth provides high accuracy, since radar theory shows that the accuracy is proportional to the inverse of the pulse time. [0107] 2. The laser itself is physically quite small, especially useful for portable applications. [0108] 3. It has a diffraction limited beam, which implies that the spot size at a distance is not limited by the properties of the light, but only by the quality of the optics used to collimate or focus it. [0109] 4. Since the wavelength is quite short (532 nm), and Rayleigh range is inversely proportional to wavelength, the beam can be kept small over a large distance interval. In fact with a 1 cm exit aperture, the beam will remain less than 6 mm over 50 m. [0110] In one preferred embodiment, the laser beam is directed by the orthogonal scanner mirrors to a laser impingement point on the surface of the object [0111] A system in accordance with the present invention can provide high ranging accuracy at high acquisition rates. For example, at 100 m ranges, a 1 mm accuracy can be achieved on a single shot basis, with anywhere from 1000 to 5000 data points being acquired per second. [0112] In other embodiments, a chirp lidar may be employed. The essential component of a chirp lidar can be modulated with a linear change of wavelength over time. Thus, the wavelength of the light emitting from the laser will be given by λ(t)=k(t−t [0113] Referring to FIG. 36, in a typical FM chirp system, a portion of the light emitted by the laser [0114] Referring to FIG. 37, a huge improvement in accuracy can be realized by adding a system to calibrate every range measurement. A fiber is prepared which has a partial reflector [0115] Referring to FIGS. 3A and 3B, in the embodiment, the FDV [0116] The fork mount system [0117] It should be noted at this point that, while conventional surveying instruments should be leveled prior to operation in order to operate properly, this is not a requirement of the FDV [0118] Referring still to FIGS. 3A and 3B, in one embodiment, two orthogonal mirrors of the FDV [0119] High accuracy and repeatability electronic encoders read the rotational angles of the orthogonal mirrors, and the readings of the mirror rotation angles are precisely timed to coincide with the range reading. Preferably, the system is Class II FDA eye safe. A first embodiment has [0120] The following is a description of the key components of preferred embodiment of the FDV [0121] Referring to FIG. 6, the laser [0122] The laser beam output of the laser [0123]FIG. 7 is a block diagram which shows an embodiment of the laser [0124] The output power of the diode pump [0125] A piece of KTP frequency doubling crystal [0126] Embodiments of the invention which meet FDA Class II eye safe system design specifications are potentially more commercially viable. In order to meet this specification, the maximum energy per pulse that can be transmitted at 532 nm is 0.2 μJ. With this restriction, the average power transmitted is largely dependent on the pulse repetition rate, and is given by the following table
[0127] In one embodiment of the invention, the beam expander [0128]FIG. 7A shows a further embodiment [0129] As the beam focus is changed, the elements should stay sufficiently aligned so as to prevent the beam from changing direction by more than a fraction of 1 mm at 50 m, or this will appear as an error in the placement of the point in space. In order to minimize this beam wander, a linear servo motor [0130] Duplexer [0131] An embodiment of the duplexer [0132] Moreover, as an optical pulse returns from the object [0133] Partially Reflecting Duplexer [0134] Referring now to FIG. 8A, for the 1 mm accuracy embodiment, a partially-reflecting duplexer [0135] Referring now to FIG. 9, in the 6 mm embodiment, improved efficiency can be achieved in collecting the return optical pulse if only the center of the window [0136] Preferably, the laser [0137] Receiver Telescope [0138] Referring again to FIG. 6, after the returning pulse has passed through the duplexer [0139] Moreover, the power returning from the object [0140] Detector [0141] The detector [0142] Scanner [0143] The scanner [0144] Electronics [0145] A. Timing Circuit [0146] Another embodiment of the scanner [0147] Electronics [0148] Ranging Electronics [0149] The function of the ranging electronics is to compute the range from the FDV [0150] Reflectivity Electronics [0151] In many cases, it is useful to know not only the position in space of a point on the object [0152] Digital Signal Processor [0153] A digital signal processor integrated circuit controls all the time critical functions of the FDV—scanner control, laser firing. It also provides fast floating point computation capability for making geometry corrections, calibration corrections, and video lens corrections, and video compression. The digital signal processor is interrupted at regular time intervals, typically about 10 usec. At each of these time intervals, a check is made to see what real time calculations are outstanding. [0154] Scanner Control [0155] The electronics for the scanner are a simple precision PID controller which are driven by a digital signal from the DSP. When driving this system quickly, there is noticeable lag in the ability of the scanner to follow the driving signal. However, the controller circuit does not have an error signal output. An external precision analog differential amplifier provides an error signal (the difference between the command signal and the actual displacement), which is sampled by the DSP at low resolution. The DSP then computes the exact scan position by computing the sum of the command signal and the error signal. The advantage of this method is that it requires only a low resolution A/ID converter and a precision D/A converter, rather than a far more expensive precision A/D. [0156] The digital signal processor generates the trajectories for the analog scanner controller, and makes measurements of the difference between the desired trajectory and the actual position. It predicts the time at which the laser pump is turned on so that the laser will fire at the desired angle. These predictions are made at regular time intervals. FIG. 10 is a flow chart that shows the calculations performed at each time interval. [0157] Trajectory Computation [0158] The user defines areas within the view of the scanner that are to be scanned, and indicates the density of points to sample within the scanned region. There are several scan patterns which can be used, and these require a specific pattern of mirror movement, known as the trajectory. The objective of picking a good trajectory are the conflicting needs of doing the move quickly and accurately. Accurate movement requires minimum torque, which would otherwise deform the apparatus. This limits the speed with which motions can be made. At equal time increments, a calculation is performed to determine the current position of each mirror. The particular calculation used depends upon the type of scanning employed. [0159] Raster Scanning [0160] When the desired scan field is a polygon, one of two raster scanning patterns is used. In the first, scanning is unidirectional (i.e., always proceeds from left to right, or right to left, on parallel lines). FIG. 11A shows such a unidirectional scan pattern. In between scan lines, the scan mirror retraces to the beginning of the next line without making any range measurements. The retrace can proceed quite quickly since no measurements are being made during the retrace. [0161] A slightly more efficient means of raster scanning is bidirectional, in which scanning is also performed on the retrace. FIG. 11B shows such a bidirectional scan pattern. This is not as efficient as it might seem because the retrace time is used for other calculations, and because the resulting scan pattern is not as regular. [0162] Both raster scanning methods require traversing a straight line in the minimum time, starting at zero velocity and ending at zero velocity. The torque applied to the mirror is proportional to the angular acceleration, which must zero at the beginning and end of the scan since the mirror is at rest. It can be shown that the trajectory that makes such a minimum energy move between two points is given by the sum of a straight line and a full cycle of a sin. However, this is closely approximated with much less computation by the minimum degree polynomial, with boundary conditions p(t0)=p0, p′(t0)=0, p″(t0)=0, p(t1)=p1, p′(t1)=0, and p″(t1)=0 which is the fifth order polynomial: p(t)=(p [0163] Spiral Scanning [0164] A disadvantage of raster scanning is that since the speed of the trajectory is varying, the scanning efficiency is not optimal. A spiral pattern can achieve a constant speed trajectory which permits a uniform point distribution. [0165] Seeking [0166] In addition to scanning a range image, the system is capable of performing a number of functions which are common in surveying. The scanner can be made to search for important features, or locations of high reflectivities. This allows the system to perform normal surveying functions by finding a target whose location is approximated identified, and reporting its exact angles and position. [0167] Angle Calibration [0168] The capacitive encoders in the moving coil motors have tremendous repeatability, but relatively poor accuracy. A number of calibration activities need to be continuously performed to ensure system accuracy. [0169] Before use, each scanner is calibrated over its complete range of angles. At a number of discrete temperatures, a map is created and stored of the measurements of apparent angles for thousands of accurately measured points using an external resolver that is traceable to NBS standards. The DSP linearly interpolates between these measured points on every angle measurement. [0170] Preferably, the accuracy of angle measurement is improved by determining scale or offset errors in the encoder during operation. Commercially available scanners can drift significantly with environment changes. This results in a shift in the effective zero and full scale range of the angle measurement, while maintaining the overall shape of the calibration curve obtained by making careful laboratory measurements before the scanner is installed in the system. The environmental effect is reduced by providing a means for determining when the scanner is at a known and repeatable angle. In one preferred embodiment of such a system, two optical references which are fixed with regard to the case of the instrument are aimed at the back of each scanning mirror. There are a variety of mechanisms for providing the optical reference, but in one preferred embodiment, a pair of autocollimators are aimed at a reflective surface on the back of the scanning mirrors and will provide a highly repeatable measurement of when the mirror is normal to the axis of each autocollimator. Each autocollimator gives a reference angle to within approximately 10 grad. Periodically, the scanner is moved under computer control to the position at which the mirror is closes to being normal to the autocollimator axis, and the apparent angle is measured. The measurements are compared with the measurements taken when the scanners were calibrated, and a linear correction is calculated and applied to every subsequent measurement. [0171] In an alternative embodiment, a pair of mechanical stops is provided just past the normal range of motion of the scanning mirror. Periodically, the mirror is driven until it touches a mechanical stop. Then, the scanning mirror is driven with a known current, which corresponds to a known force. The mirror arrives at equilibrium at a very repeatable position, and this is used to calculate a linear correction to the mirror calibration curves. [0172] Range Calibration Fibers [0173] The timing circuits have a certain amount of offset and scale drift with time and temperature, and a provision has been included to compensate for these variations. When an optical pulse is emitted from the laser [0174] The fibers can be manufactured so that the end at which the pulse is launched [0175] Range Walk Calibration [0176] The lidar system measures the range of surfaces by timing the delay between the laser pulse being emitted and returning from the surface. This delay is measured electronically by imposing a sample of the outgoing pulse, and the return pulse on an optically sensitive electronic detector [0177] Part of the timing circuit estimates the energy in each detected pulse. A table of corrections is maintained to improve the range estimates. Two different circuits have been employed to make a measurement of the pulse energy for this purpose. The first is a gated integrator, the gate being open at the beginning of the pulse, and closed at the end. The signal is applied to a comparator [0178] Periodically, the timing circuit is used to measure a sequence of pulses which have been delayed by the single mode fibers [0179] Geometry Calculation [0180] The output of the FDV after a range scan consists of points in spherical coordinates with respect to a coordinate system in the scanner. However, the raw data consists of mirror angles and time intervals. The DSP computes the spherical coordinates of the scanned points by taking into account scanner geometry (mirror thickness, optic axes, mirror offsets, etc.) and all the appropriate calibration adjustments. [0181] Laser Control [0182] Delay Prediction [0183] The digital signal processor is responsible for controlling the firing of the pulsed laser, but it can only do so indirectly. The processor has control of the timing for starting the pump diode, which causes the passive q-switch to fire after saturation has occurred. However there is a variable delay between turning on the pump and having the laser fire. The delay is a function of junction temperature, which in turn is a complex function of ambient temperature and recent history of laser firing. The delay generally ranges between 100-300 usecs. [0184] Fortunately, it is primarily necessary to know the scanning mirror angle at the precise moment the laser fires. After the laser has been fired just a few times, the pump delay does not change quickly if the firing rate does not change quickly. As a result, accuracy of a few microseconds can be achieved by estimating the next pump delay to be the same as that in the previous firing cycle. The digital signal processor measures the pump delay by reading an internal counter when the pump is started and when the laser actually fires, causing an interrupt. Since the interrupt latency is less than a microsecond, this becomes the timing accuracy to which the pump delay can be measured. [0185] A more sophisticated dynamic model of the thermal properties of the laser could lead to slightly enhanced scanning pattern regularity, but is probably equally limited by the time resolution of the processor interrupts. [0186] Firing Control [0187] Given a time vs. angle trajectory for a scanning axis, w(t), a desired angle to fire the laser, and an interrupt interval Dt, the decision to fire the laser amounts to computing the time at which point the pump diode is started. [0188] Computer Control [0189] The FDV is designed to perform under the control of a remote host computer which contains graphical controls for a user to specify areas to be scanned. The remote machine controls the FDV through a bi-directional serial byte stream, which is effected in any of a number of media: Ethernet, EPP parallel port, serial port. A processor in the FDV is assigned the task of decoding messages, and scheduling the required activity. FIG. 12 is a block diagram of the FDV processor. [0190] Host Communications Interface [0191] The host machine acts as a master, sending a well defined message protocol to command the FDV. When actions are completed, the FDV responds with data and status information. Among the actions which can be requested are: [0192] Point the scanner [0193] measure a distance [0194] range scan a box [0195] fire the laser n times [0196] take a video image [0197] Scanner Control [0198] Referring to FIG. 13, in normal operation, each scanner in the dual mirror system requires a 16 to 18 bit digital word to set the desired position, which is applied to a precision digital to analog converter to create a voltage proportional to the desired position. However, there will be some error between the position commanded by the output of this converter and the actual position of the scanner, which is reflected by the output of the position encoder. A precision difference signal is generated, and the difference is measured to 12 bit accuracy. This provides an economic method of making 18 bit position measurements while only using an inexpensive 12 bit converter. [0199] Commercially available galvo scanners have microradian repeatability, but have relatively poor scale and offset performance, particularly over temperature. A calibration mode has been incorporated into the system to permit making measurements at two precise angles, and using the two measured data points the offset and scale drift of the scanner can be calculated. [0200] Two methods have been developed for this purpose: an optical and a mechanical means. In the mechanical method, the scanner shaft is gently placed against one of two mechanical stops, and the current in the scanner controller is adjusted to a specific value, which provides a known force. The position signal is adjusted until there is no position error, and this gives the calibrated position measurement. In the optical method, two autocollimators are aimed at the back of the scanner mirrors, which have also been polished and mirror coated. When the scanner mirrors are exactly aligned with one of the collimators, the output from the split photodetector in the autocollimator is balanced. By placing the scanner in each of these precise angles in turn, an offset and scale correction for the scanner encoder can be calculated. [0201] Timing Circuit [0202] The purpose of the timing circuit is to provide the relative time between the start pulse and the stop pulse, in picoseconds. There are two subsystems in the timing circuit: a signal conditioning and energy integration circuit (an embodiment of which is shown in FIG. 14), and a time interval analyzer. Both communicate directly with the DSP. Initially, systems have been produced with a commercial timing instrument, the Stanford Research Systems SR620 time interval analyzer. The interface to this instrument is through an IEEE488 interface. In a preferred embodiment, the communications interface to the Stanford Research Systems SR620 time interval analyzer is IEEE488. [0203] A custom time interval measurement circuit has been developed which utilizes a separately patented interpolation technology. The circuit employs a clock, typically operating at >100 mhz, which is used to make a coarse count of 10 nsec intervals between stop and start pulses. Additionally, there is an interpolator which divides each 10 nsec coarse count into 1000 smaller increments, providing 10 psec resolution. This system has approximately 5 psec jitter. Differential time measurements can be made with less than 20 psec RMS error, which corresponds to about 3 mm. This circuit communicates with the DSP using a dedicated serial bus, and employs a packet protocol: the DSP arms the circuit by sending a single byte. When the timing circuit completes its task, it sends a sequence of bytes which represent both the time delay between start and stop pulses, and the intensity of each pulse. [0204] Laser Firing [0205] The DSP has three lines for laser control: one starts the laser pump, the second indicates that the laser has fired, and the third indicates that the return pulse from a target has been detected. When the laser fires, the DSP samples the analog pulse amplitude signal. This happens typically within 1 μsec. [0206] Video [0207] For targeting, the user is provided on the host a video representation of the scene from which he can choose a portion to be range scanned. In most cases this will correspond to the scene rendered in ambient illumination. [0208] Capture [0209] One way the video is captured is by using the scanner to aim single sensitive detector across the scene with the laser turned off. This permits acquiring an image which has very accurate spatial alignment with subsequent range scans. However, image capture can be quite slow in comparison to commercially available cameras. [0210] A second approach is to utilize standard commercial CCD video cameras to acquire an image. One CCD camera with a wide angle lens is aligned with the range scanner with as small an offset as possible. A second camera with a 5 degree field of view is placed so that its optic axis is coaxial with the transceiver. Thus, a much smaller field of view is accessible through the scanner, and can be scanned with the same resolution as the transceiver. This allows targeting small or distant objects. [0211] Alignment [0212] The wide angle lens introduces a fish-bowl effect in the image captured by the CCD sitting behind the lens. Straight lines in the world are not straight in the image. This distortion increases with the distance from the center of the lens. This distortion is removed by comparing the image the camera produces when aimed at a carefully designed and printed calibration target image. The difference in the anticipated image and the recorded image provides the information needed to warp subsequently acquired images to eliminate the distortion. [0213] Compression [0214] Each video image is compressed prior to transfer. Currently we are using JPEG standard image compression. It is relatively fast, and creates reasonably small compressed images for communication. Another desirable feature is that the algorithm operates on blocks, which permits us to do interleave image capture, alignment, compression, and transmission in parallel—significantly enhancing throughput. [0215] Point Video [0216] A second camera, with a narrow field of view (e.g., approximately 50) is placed such that it is coaxial with the scanning laser beam. The field of view is adjusted so that the pixel resolution is approximately the same as the voxel resolution of the lidar system. The camera can be operated while the laser is activated. When this is done, a small group of pixels will be illuminated by the laser, and the centroid of these pixels will correspond to the point which would be measured by the lidar. When a video image is captured, it can be mapped onto a surface which is estimated by a lidar scan. [0217] Computer Graphics Perception (CGP) Software [0218] Referring to FIG. 15, the CGP [0219] The CGP [0220] With reference to FIG. 1A, the data acquisition and modeling process divides into the following steps: FDV [0221] With reference to FIGS. 16A and 16B, the foregoing operations may be performed in at least two graphic display windows. One window [0222] Scan Control [0223] Referring to FIG. 15, prior to using the integrated hardware/software system the FDV [0224] Scan control is the process of indicating which portions of the scene that are visible to the scanner are to be scanned. Different parts of the visible scene can be scanned at different densities, since simple geometric objects, such as planes, cylinders and spheres can be accurately modeled with a fairly low number of scan points. Therefore, the region in front of the scanner is often captured in multiple scans, rather than in one high resolution scan. Only regions with high levels of detail need high resolution scans. [0225] Referring to FIG. 17A, one of the means of scan control is the use of a video image [0226] The CGP [0227] Additional scans at different densities can be initiated in the same way, or one can use previously scanned data points rather than the video image to specify new scan regions. If the view of the scanned data is oriented so that it is exactly aligned with the scanner direction, then a scan region can be indicated by methods such as dragging out a rectangular box. When the data is aligned to the scanner in this way most of the 3-D information is difficult to see, therefore, the software can display the points with the intensity of the returned laser light at each point color mapped as described in the next section. The intensity information is often sufficient to identify objects in the data window, so that new scan regions can be defined. Alternatively, the user can model and/or color some of the objects in the scene to help locate regions of interest in the window. Using the data window to define new scan regions avoids any parallax errors, since the view is aligned with the scanner. [0228] Scan control can also be achieved by using the pointing device to move the laser beam and highlight points in the actual scene. Any number of methods could be used to describe the desired scan region by moving the laser beam and identifying points of interest by a user action, such as clicking a mouse button. Methods could include operations such as: indicating a bounding box by moving the laser to diagonally opposite corners of the desired scan regions; indicating the top, bottom, left and right bounds of the scene; indicating a sequence of points that represent the bounding polygon of the scan region; indicating the center of the scan region and using other means, such as dialog boxes to describe the extent of the desired scan region. [0229] Point Acquisition [0230] With reference to FIG. 17B, the data returned by the FDV [0231] Each point returned is displayed in the data window [0232] The ordered grid of points generated is referred to as a scan field. Multiple, possibly overlapping scan fields may be gathered and simultaneously displayed in the manner described above. The data structures within the CGP [0233] Segmentation [0234] Segmentation is the process of grouping together points that were scanned from the surface of the same object. The points from a single object may be a small portion of a scan field, or may occur across multiple scan fields. The segmentation process may be manual, as described below, or automated, as described later in the auto-segmentation section. [0235] Referring to FIG. 20, the first step of the manual segmentation process is to select one or more scan fields [0236] The group of points resulting from this step form a pool of candidate points that can now be trimmed to remove points on other objects. Each point in the pool is initially marked as selected, and the operations described below can be used to toggle the point states between selected and deselected. [0237] Referring to FIG. 21, scan points [0238] In FIG. 22, three new groups of points [0239] Geometry Fitting [0240] In one preferred embodiment, the CGP [0241] As stated above, the CGP [0242] Fitting a plane to a set of points is a simple problem that has many well-known solutions. The extent of the patch used in the CGP [0243] Many standard approaches are available for fitting more complex shapes. In one preferred embodiment, two phases are involved: a parameter estimation phase to get a starting point, and an optimization phase, where the parameters are varied to minimize the overall error. The total error is the sum of the squares of the distance between each scan point and the nearest point on the surface of the object being fit. The optimization phase uses conventional optimization methods to reduce the error between the object, as defined by its parameters, and the data given by the scan points. [0244] A cylinder fitter can convert a cloud of points [0245] A cylinder is described by five parameters: a normalized vector describing the cylinder axis (two independent parameters), the radius, and two additional parameters that are used to locate the line of action of the cylinder axis in space. The length of the resulting cylinder can be determined by projecting the scan points onto the cylinder axis and noting the extreme values of this projection. [0246] Two novel methods for estimating cylinder parameters are implemented in one preferred embodiment. The first way to find initial parameter estimates for a cylinder is to find approximate surface normals, as described in the auto-segmentation section. If all of the normals are set to unit length, then they can all be consider to be vectors from the origin to a point on the surface of the unit sphere. If one uses each normal vector and its to accumulate a group of points on the unit sphere, then one can fit a plane through the resulting group of points. The resulting plane normal is roughly parallel to the cylinder axis. Given the cylinder axis and the plane from the previous step, one can project the scan points onto the plane. The projected points will be well described by a circle in this plane, since the plane is normal to the cylinder axis. A best fit circle can be calculated using the projected points on the plane to give an estimate of the cylinder radius. The center of the circle on the plane can be converted to a 3-D point to give a point on the cylinder axis. [0247] The second way to estimate the cylinder parameters is to fit the set of point to a quadric surface, which is described by the implicit equation: [0248] where p={p [0249] One can then take c [0250] where
[0251] Two unit vectors, u [0252] The unit principle vectors v [0253] There are two solutions to this equation that yield the orthogonal unit principal vectors v [0254] For cylindrical surfaces one of the principal curvatures will be near zero, and the radius of the cylinder is the reciprocal of the absolute value of the nonzero curvature. A method to determine the radius (r) and axis of the cylinder has been described, and only the location of the axis needs to be determined. A unit surface normal can be calculated as {circumflex over (n)}=N [0255] The novel method described above for curvature estimation using the quadric surface formulation is further used in a novel way for automated object type determination. If the points being fit are well represented by a plane then both principal curvatures will be near zero. If the points being fit are from a cylindrical surface then one curvature will be near zero and the other will be nonzero. If the points are from a sphere then both curvatures will be nonzero and their magnitudes will be approximately equal. Combining the automatic detection of object type and the auto-segmentation algorithm, described later, allows the CGP [0256] A further use of the curvature estimation is sphere fitting, which is achieved by using the quadric surface approach to approximate the radius and location of the center point, and then using a four parameter (center point and radius) minimization to reduce the error between the sphere model and the measured points. The novel method described above for finding a point on the axis of a cylinder is also used in the preferred embodiment to find the center of a sphere. [0257] The segmentation techniques disclosed above can be used to create a variety of useful fitting tool based on combinations of the previously described shapes. For instance, a corner, consisting of an intersection of three planes which may or may not be orthogonal, is a very common feature to scan. Knowing that the specified point group contains three intersecting planes, such as the points [0258] Each object stores information about the quality of the fit, so that the user can query the object and examine the mean, standard deviation, and worst errors. Knowing the accuracy of the FDV [0259] In addition to the class of general object fitters, which are given close to no initial information other than the points to fit, there is a class of fitters that can take advantage of foreknowledge about objects in the scene. An area in which such foreknowledge exists is that of the construction industry, where parts used follow standards of dimensions and design. For example, the external diameter of pipes from a particular manufacturer may come in five different sizes: 4″, 5″, 6.5″, 8″, and 10″. This information typically resides in tables that describe relevant attributes of these parts. The cylinder fitter can take advantage of the information in these tables to significantly reduce the solution space to be searched: the fitter need only search for solutions involving cylinders of one of those diameters. Another way to use such table lookups is to have the fitter come up with a general solution, then match against entries in the object tables to find the entry with the closest parameters. For example, a pipe fit by a cylinder with 7.8″ diameter would be matched against the 8″ entry in the table from the example above; the user (or fitter) then has the option of refitting an 8″ cylinder to the pipe, or accepting the 7.8″ cylinder. Yet another use is for the user to manually select a specific entry (or collection of entries) from the table and tell the fitter to use its parameters in the fit, which also reduces the fitter's solution space (which can decrease the time taken). [0260] Modeling [0261] The fitting of geometric primitives, as described in the previous section does not usually complete the modeling process. It is often the case that only a portion of the objects surface is scanned, such as one end of a cylinder or a portion of a wall, and further operations are required to complete the 3-D model. Modeling is the process of completing the construction of the 3-D model given some fitted geometric primitives. [0262] Many common CAD operations such as extension, intersection (mutual extension) and trimming are available in the CGP [0263] Object extensions can be accomplished in several ways. One way is to select the geometric object to be extended and tell it to extend to a subsequently selected object. The nature of the extension is determined by both the type of object to be extended and the second object selected. For example, a cylinder extends the end closer to the second object along its centerline until its end intersects with the infinite plane defined by the second object's geometry (in the case of a planar patch, the infinite plane is that of the patch's plane, and for a cylinder, the infinite plane is that containing the centerline and as orthogonal to the extending cylinder as possible). [0264] Another way is to make use of object handles, which are nodes that the user can grab. These handles are tied to an object's definition (position, orientation, and size) where appropriate, and by moving a handle, the object's definition changes accordingly. Again, taking the cylinder as an example, the same extension described above can be accomplished by grabbing the handle on the end to be extended, and then moving the handle (and extending the cylinder) to the desired position. A handle's motion depends on the part of the object with which it is tied; a handle on the centerline of the cylinder is constrained to move only along that centerline, while a handle on the boundary of a planar patch is constrained to move only inside the plane of the patch. For some objects, handles may be inserted and removed, changing the definition of the shape of the object (for example, handles on a planar patch have a one-to-one correspondence to vertices on the planar patch's boundary). Other handles can provide rotational control over an object. The control of handles is interactive and dynamically updates, so the user can see the intermediate results of the redefinition. [0265] A new operation, called merging, has been developed to allow different portions of a single object surface to be joined to form a single object in the CGP [0266] Using the manual or automatic methods, the user can take a cloud of points [0267] Scene Registration [0268] The initial position of each scan point is described in a local coordinate system whose origin is that of the FDV [0269] A novel process is used to register the scan fields from different FDV [0270] Given a set of point pairs, the registration process searches the set of candidate points for three pairs that are not colinear. Using the three point pairs, one can construct the transformation required to convert the coordinate system in one view to that used in the other view, which in turn can be used to transform the scan points to all share a single coordinate system. For convenience, the process will be described in terms of the first data set remaining fixed while the second data set is transformed to use the coordinate system of the first. The process works equally well if the user fixes the second data set and transforms the first. [0271] The first step is to apply a rigid body translation to the second data set to make the first pair of points coincide in terms of their x, y and z components. The second step is to rotate the second data set about its first point until the lines formed by points one and two in both data sets are colinear. The third step is to rotate the second data set about the line established in the previous step until the planes defined by points one, two and three in both data sets are coplanar. [0272] Once an initial estimate is made one can use all the point pairs and an error minimization method to reduce the sum of the squares of the distances between each point pair. [0273] In order to use the point registration method described above, the CGP [0274] In replacing planes and lines, one can only introduce points that are at location relative to the user specified objects, since the origins of the two data sets are different. For instance, introducing a new point pair in a plane at the location closest to the origin would not result in points that actually match in space, since the origin is arbitrary. However, introducing a point pair at a plane-line intersection will give matching points in the two data sets. Some pairs of objects, like parallel lines, should not be used to introduce new points so an angular tolerance, called ATOL below, is used to ignore poor object pairs. ATOL is initially set to ten degrees but other values can be used to generate fewer or more artificial point pairs as needed. The point pairs are introduced in the following order: [0275] For all plane-line pairs where the angle between the line and plane is greater than ATOL, introduce two new point pairs. The first new point is inserted at the intersection point of the line and plane, and the second point pair is inserted along the line of action at a fixed distance away from the first point, here taken to be the minimum of the line lengths in the two views. [0276] For all pairs of planes and points, introduce a new point pair on the plane is such that the plane normal passes through the new point and specified point. [0277] For all plane pairs whose normals are at least ATOL apart, generate a new line segment along the intersection of the planes and make the line segments length equal to the minimum extent that any plane has along the line. The new line segment has no direction, but has both length and position information. After this step the planes are no longer needed. [0278] For all pairs of lines and points, introduce a new point on the line at the location where it is closest to the specified point. [0279] For all pairs of lines separated by an angle greater than ATOL, introduce four new pairs of points. The new points are the ends of line segments along the original line of action, but centered on the location of closest approach of the two lines. The distance between the new line points is equal to the minimum length of the line segment lengths along that line of action from the two data sets. After this step the lines are no longer needed. [0280] The result of the plane and line replacements as described above is a set of point pairs that retains the direction information associated with the original planes and lines. The augmented set of point pairs can then be used for the registration process that is described above. [0281] After registration of the two scenes, primitives from the two individual views which represent the same physical object can be combined using the merging technique described previously. In particular, matching planar patches representing the same surface can be combined into one extended planar patch. Similarly, pieces of matching cylindrical surfaces can be merged to form a single cylinder. [0282] Warping Data Sets [0283] The registration process described above is a rigid body transformation that does not modify the relative locations of objects within either data set After registration, most of the point, line or plane pairs that were identified will still have small errors, since the minimization process reduces the total mean square error. A novel method is presented that allows the user to force the identified pairs to exactly match by deforming the scene volumes. [0284] As with any measured data, there is some level of error associated with each scan point location. The magnitude of the error associated with a point location will vary with the measuring technology used, but some error will always be present. Since the data under consideration here describes surface features of objects, the data errors will manifest themselves as surface irregularities. For example, a set of points acquired from an actual planar surface may not all lie on a plane, but will have some small scatter away from the real plane location. Calculating a best fit plane through the set of measured points may not give the real plane location or orientation due to the errors in the point data set. [0285] The errors in recovered features, such as planes, cause errors in the relationships between the recovered objects as well. For instance, if data is collected from two planes that have an exactly ninety degree angle between them, the best-fit planes generated from the data points may not be exactly ninety degrees apart. Similarly, cylinders that were parallel in the real scene may result in best-fit cylinders that are not parallel after fitting from the scanned points. These inconsistencies in the recovered features, that occur due to measurement errors, will appear whether the data points are collected from a single scan position or are a union of scans from a variety of different positions. [0286] The lack of fit problem may actually grow if several different sets of scan data are registered using a relative system. If a series of sequential scans are collected, and each scan is registered with respect to some recognizable sequence of data points in a previous scan, then the absolute errors in each scan may grow. If at the end of the sequence of scans the locations of features are exactly known, then one must adjust the scanned data points so that they fit the known locations. In surveying both the 2-D closure problem and the 3-D benchmark matching problems are similar in nature to the problems described above. In the surveying closure application, when one surveys a sequence of 310 locations and arrives back at the starting location one typically finds that through cumulative measurement errors the starting and finishing locations are not at exactly the same location. The closure error, which is the distance between the starting in finishing locations, is distributed using well known surveying techniques throughout the other data points collected such that the first and last end points meet after the correction is made. Similarly, when surveying benchmarks of known location arc introduced into a surveying data set the data set must be adjusted to accommodate the known benchmark locations. Both the closure problem and the benchmark matching problem can be solved by the method described here since they can be described in terms of displacement constraints. [0287] The novel method described here to correct location errors in measured 3-D data sets and distributes the errors throughout the point sets by applying solid mechanics principles to a volume surrounding the data points. The method provides a technique for satisfying a wide variety of displacement constraints on 3-D data sets and also distributes the measurement errors throughout the data sets. The process of deforming the data sets to achieve these goals is called warping. The displacement constraints can be specified in terms of both control points, whose absolute coordinates are known in space and do not move, and tie points, which represent the same location in two or more data sets, but whose absolute location is unknown. One can describe constraints involving more complex objects, for instance, line segments by specifying two points, and planes by specifying three points. In this way an entire family of constraints can be specified and applied to groups of objects such as points, lines and planes. [0288] All constraints include an offset term that may be zero or nonzero. A zero offset between points indicates that the points should share the same final location, but does not prescribe where that location would be. If one of these points were a control point, with a known absolute position, then it would not move and the other point in the constraint would be forced to move to the same location to satisfy the constraint. In all cases, the final location would result from the energy minimization process that is associated with a solid mechanics solution. A non-zero offset between two points indicates that the points are to be a certain distance apart after the warping process is applied. [0289] If the constraint's objects are lines or planes (rather than points), then in addition to the offset one can specify an angle between the objects. Using this family of constraints types, one could specify any number of relations between features in one or more data sets. The invention works whether a single or multiple 3-D data sets are involved, and the constraints can be between objects in the same or different data sets. [0290] The solution of the constraint satisfaction problem begins by registering each of the data sets, as described in the previous section, so that all of the data shares a single coordinate system. Next, solid mechanics theory is applied to a volume surrounding the points in each data set to satisfy the displacement constraints. The warping method operates on one or more volumes that are created so that they surround all the points in a given data set. Each of these volumes is considered to be made out of deformable materials whose properties are specifiable. An isotropic material can be used with an elastic modulus of one and a Poisson's ration of zero. The solid mechanics solution finds the minimum energy displacement pattern that satisfies the specified constrains. [0291] One can picture each of these volumes made out of a flexible material. If one anchored a volume to prevent rigid body motions and then prescribed a new location for a point on the interior, one could envision a distortion of the volume that involves not only the point of interest but also the rest of the volume. In reality, the constraints themselves can be used to anchor multiple volumes together. Mechanics principles allow one to determine the minimum energy deformation of the volume that satisfies the stated constraints, which mimics what would actually happen to a real deformable object subjected to the same constraints. [0292] In a particular embodiment, the warping method uses principles of solid mechanics to deform a volume containing points of interest in order to satisfy a set of constraints applied to the data points. Not only are the constraints satisfied, but the effects of the initial location errors are spread throughout the volumes operated on. [0293] The finite element method is used to apply the principles of solid mechanics to the volumes enclosing the points. The volume is discretized into a set of points or vertices and a set of elements that connect to the vertices. Four node tetrahedral elements are used to discretize the volume. [0294] The first step of the process is to collect the set of constraints that apply to one or more data sets. During this phase, one must identify constraints that are to be satisfied by the warping process. These constraints include the identification of points that represent the same physical location in different data sets (tie points), like the corner of a cube, which are to appear at the same location when the warping process is completed. Some of the tie points may not be points that were scanned in the original data set, but may be constructed from groups of other points. For instance, if one had a series of points that represented three planes intersecting at a corner, then one could fit three planes to the points, and use the resulting corner point as a tie point. The constraints are specified in terms of pairs of objects, such as points, lines and planes, as well as the desired offset and angle between them. The two objects involved in the constraint can be contained in a single data set or can occur in different data sets. Within a single data set, one could specify that lines or planes remain parallel, or that the distance between two points be a specified amount. Between multiple data sets, one could write similar constraints, or indicate that the features seen in two data sets represent the same object. One might also know the actual location of some points very accurately (benchmarks) and constrain points in the data set to lie at the known locations. Using these benchmark points to anchor different points in the data sets enables the closure problem to be solved, since the data sets will be warped so that the measured data points move exactly to the desired control point locations and the errors in the data set will be smoothed over all volumes. [0295] The second step in the warping process is to register all the data sets involved, as described in the previous section. [0296] The third step in the warping process is to select a volume that surrounds the region of interest, and describe the volume in terms of a set of new points. [0297] The region that can be scanned by the FDV 10 is called the view volume and is shaped like a pyramid, with the tip of the pyramid located at the origin of the scanning device. A pyramidal shape can be used to bound the view region for the purpose of warping, and the pyramid is easily described by five points, using the same coordinate system as the data points. These new points do not become part of the data set, but are used in the warping process. The convex hull of these points represents the new volume surface, and should enclose all the data points on the interior. This operation is performed separately for each data set. [0298] The fourth step is to mesh each of the data volumes. Meshing involves filling the volume with finite elements that leave no gaps and that do not overlap. The finite elements span between the points or vertices that have been defined on the volume boundary and those that are involved in constraints on the interior. The points in the data set do not all need to be included in the warping process, only those that are used in constraint specifications and those that define the volume boundary need to be used. The elements in the initial mesh may be of poor quality due to their shape. Long sliver elements, for instance, are known to give poor results in finite element analysis. Therefore, the meshing process is actually iterative. New points are inserted into the mesh, and then old elements are removed and new elements are introduced so that the mesh quality improves. This iterative process continues until one is satisfied with the overall quality of the mesh. In one preferred embodiment, four node tetrahedral elements are used. The initial mesh is constructed by applying a 3-D Delaunay triangulation on the starting set of points. The iterative process identifies poorly shaped elements using an element quality measure, and introduces new points and remeshes the region. The process terminates when all elements meet a minimum quality criteria The preferred implementation uses longest edge bisection to introduce new points that improve the mesh, but other methods can be used. [0299] The fifth step processes the constraints described in step one into a system of linear constraints. In the preferred embodiment, the final system of constraints is linear in terms of the nodal displacements at the vertices of the tetrahedral elements. The desired form of the constraints is: [0300] The matrix C contains constant coefficients. The number of rows of C is equal to the number of constraints in the system. The vector u represents the 3-D displacements of the vertices of the tetrahedral elements. The vector q contains constant coefficients. If the constraints are homogenous then each element of q will be 0. The form of constraint specification given in Equation (6) allows arbitrary linear multipoint (involving more than one vertex) constraints. [0301] The conversion of the constraints specified in step one into the form shown above depends on the type of constraints involved. For two points to be tied together the constraint would be: or [0302] In these equations, p [0303] Other constraints, like the distance between two points, are non-linear in nature. The nonlinear constraints can use the existing geometry of the system as well as small deformation assumptions to produce linear multipoint constraints. For example, to specify the desired distance between two points to be some specified value x, one could determine the vector v [0304] and then specify the desired length of the vector: || [0305] or, using the vector dot product: [0306] Both Equations (11) and (12) are nonlinear in terms of the displacements of the nodes, u or [( [0307] The term on the right hand side of equation (16) is the desired distance between the points minus the current distance between the points. The x, y and z components of n [0308] In Step 6 the final system of linear equations is assembled. There are two parts to this step: first, assembling the element stiffnesses for each of the tetrahedral elements, and second, selecting and applying a constraint handling technique. The calculation and assembly of element stiffnesses follow standard finite element procedures. Using constraints in the form of Equation (6) involves a constraint processing method. The Lagrange Multipliers technique can be used to introduce the effect of the linear constraints, but any other method, such as penalty or transformation techniques, could be used equally effectively. [0309] Using Lagrange Multipliers, one introduces a new variable into the final system of equations for each constraint in the system. One then modifies the static equilibrium equations for the unconstrained system, which are given by: [0310] In Equation (18), K is the system stiffness matrix, assembled from the individual element stiffness contributions, u is the displacement vector that is the solution of the problem, and r is a vector of externally applied loads. In this embodiment of the invention, there are no externally applied loads, so the r vector contains only zeroes. Equation (18) does not include the effect of any constraints, but these can be included using the Lagrange Multipliers technique to give the system of equations:
[0311] In Equation (19) K, C, u, r, and q are as previously defined, and UL is a vector containing the additional Lagrange Multiplier variables that are introduced using this method. The matrix C [0312] If penalty or transformation methods were used instead of Lagrange Multipliers, a system of linear equations different from those shown in Equation (19) would be produced, but the solution of the linear system of equations will give similar values for the displacement vector u. [0313] In Step 7 Equation (19) is solved to give u and UL. There are many methods available to solve large systems of linear equations, and the preferred embodiment uses a symmetric solver with a profile storage scheme. The different types of solvers that could be used give essentially the same results, but optimize speed and memory usage differently. [0314] The preferred embodiment uses a direct solver, but iterative sparse solvers could be used as well. The system of equations shown in Equation (19) is sparse, so significant speed enhancements can be achieved by selecting the proper solver. However, the results of the warping process overall are unaffected by this choice. [0315] In Step 8, one must check if the current displacement satisfies the constraints to a desired level of accuracy. If the current deformed shape violates the offset or angle in any of the constraints collected in Step 1 by more than a user specified tolerance, then steps 5 through 7 must be repeated, starting with the new deformed shape. The linearizations of the shape may change on each iteration since the geometry of the volume changes with the cumulative deformations. When all constraints are satisfied within the given tolerance, then one can proceed to step 9. [0316] Step 9 uses the nodal deformations u calculated in Step 7 to determine the deformation of any point of interest within the volumes. For each point of interest, one must find an finite element that includes the point on its surface or interior. If the point is internal to an element then only one such element exists. If the point is on the surface of an element or along the edge of an element, then several elements could be considered to contain the point. Any of these elements can be selected to determine where the point of interest moves. If the point is shared between elements, then the use of any of the elements to find the point displacement will give the same results. Once an element is identified, the vertex displacements of that element are extracted from u and are used to determine the displacement of any point on the interior using an interpolation process. This procedure uses the finite element shape functions which are linear in the preferred embodiment, and is a common operation in finite element analysis. [0317] Auto-Segmentation [0318] The novel auto-segmentation process, as presented below, involves a similar sequence of operations to the manual modeling process described previously. A point cloud is segmented, geometric primitive objects are fit to the point groups, and then modeling operations, such as extension and intersection are used to complete the model. In this novel process, automation is applied to each of these steps, as well as the entire process, as described below. [0319] It is possible to automatically partition the scan points into groups representing primitive geometrical shapes by using variations of common machine vision techniques. A gridded scan field is stored in a two dimensional array of points, much like a regular bitmap. The scan field differs from a bitmap in that more information is stored at each location than just a color. Each point stores its location in space, from which the distance to the scanner can be calculated, as well as the intensity of the return laser pulse. The depth information calculated from the three dimensional position stored at the points is crucial to the automated segmentation algorithm described here, even though many operations, such as filtering, rating, thresholding and thinning are commonly used image manipulation operations. [0320] The first stage of the auto-segmentation process is to estimate the surface normal at each point in the grid. This can be achieved using many different techniques, the current embodiment of the software fits a plane to the nearest neighbors of the point in the 3 by 3 grid surrounding it. The normal of the resulting plane is taken as the normal at the center point. Each point in the grid has a normal calculated in the same way, except that edge and corner points ignore the missing neighbors in the normal calculation. The normal stored at each point is a three dimensional vector and is normalized to have unit length. [0321] In the second phase two rating images are created by convolving standard edge detection filters over the grid. The first rating image is created by convolving the depth of the grid point with an edge detection filter to identify depth discontinuities, such as those that would occur at an occluded edge. A variety of edge detection filters can be used, but rather than operate on color or intensity the filter operates on the depth information stored at each grid point. [0322] The second rating image is created by convolving the normal with an edge detection filter. The normal rating image is actually composed of 3 subimages created from a convolution with the normal's x, y, and z components. The resulting three values are combined by taking the square root of the sum of the squares to give a per-point scalar value. The second rating image is used to identify normal discontinuities, such as those that would occur at the edge between a wall and a floor. Again, a wide variety of edge detection filters can be used, but the values used are normal coefficients rather than color or intensity. [0323] Once the two rating images have been created they must separately be converted to binary images. Conventional machine vision algorithms, such as recursive thresholding can be used to achieve this task. Each point in the depth and normal rating images contains an estimate of the gradient of the depth and normal respectively. Recursive thresholding can be used to isolate the regions of highest gradient. In the resulting binary images the points in the regions of highest gradient are marked as edge points while the rest of the points are marked as non-edge. [0324] A final binary image is created by marking a point as an edge point if it is marked as an edge point in either or both of the two binary images created by recursive thresholding above. All other points are marked as non-edge. This image contains all edge points that delineate the boundaries between groups of points on different surfaces. [0325] The final step of the point partitioning process is to use a connected components algorithm to collect the points into groups separated by edges. Points are considered to be connected only if they are vertically or horizontally adjacent in the grid, diagonal adjacency is not used. Very simple algorithms can be used to identify the unique groups of non-edge points in the image. Each group of connected points is then cut from the initial point set to form a new group of points. The result of this algorithm is the partitioning of the point set into multiple point groups that each represents a single surface. Each of the new point groups can be fit by a geometric primitive as described in the next section. [0326] Once the scan cloud has been partitioned into groups of scan points that lie on different surfaces, the next step is to fit objects to the desired surfaces. A variety of methods can be used to achieve this task. The current embodiment of the software can perform the object fitting process in two different ways. The first method fits a series of objects to each group of points, and selects the objects that produces the smallest distance errors between the measured points and the fitted object surfaces. The second method uses the quadric surface fit described previously, and resulting principle curvatures, to determine if a plane, cylinder or sphere should be fit to a particular point group. Other variations of these approaches could also be used, such as progressive commitment, where objects are fitted in order from simplest to most complicated, and the process stops whenever the errors associated with the particular fit drop to acceptable levels. [0327] The last stage of auto-segmentation process extends primitive objects, where possible, to create complete object intersections, rather than stopping at can point boundaries. Using the gridded nature of the original data and the edge information from the point partitioning algorithm described above, it is possible to extend and intersect objects. For all edges that result from surface intersections, which are the surface normal discontinuity edges described above, one can extend the objects on either side of the edge to form an intersection. [0328] Model Annotation [0329] In order to compose a semantically rich 3-D model, individual parts in the above geometrical model can be annotated with additional, possibly non-geometric, information, such as material references or part numbers. Tthis information can be entered manually through a special window for displaying object attributes. [0330] The user may click on an individual part in the geometrical model and recover such additional information through other windows. Similarly, the user may request that all parts which meet some selection criteria are to be highlighted. [0331] A novel method is also used for automatic model annotation. This method uses the FDV [0332] Geometry Display and Query [0333] The model is accessible in a variety of ways, including access through the data window [0334] The resulting model can be exported to any of a number of CAD programs for further editing or designing. In the preferred embodiment, the CGP [0335] The following documents form an integral part of this specification: [0336] Modular Decomposition [0337] Summary of Proposed CGP Specification [0338] Cyrax Software Specification [0339] Product Overview [0340] Collection of View Slides [0341] Introduction and Overview Referenced by
Classifications
Rotate |