CN103377373A - Image feature generation method and equipment, classifier, system and capture equipment - Google Patents
Image feature generation method and equipment, classifier, system and capture equipment Download PDFInfo
- Publication number
- CN103377373A CN103377373A CN2012101239754A CN201210123975A CN103377373A CN 103377373 A CN103377373 A CN 103377373A CN 2012101239754 A CN2012101239754 A CN 2012101239754A CN 201210123975 A CN201210123975 A CN 201210123975A CN 103377373 A CN103377373 A CN 103377373A
- Authority
- CN
- China
- Prior art keywords
- image
- classification device
- image block
- mean intensity
- object images
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention discloses an image feature generation method and equipment, a classifier, a system and capture equipment. The method for generating image features from an object image comprises the step of image block detection and the step of image feature generation. According to the step of image block detection, at least one image block is detected in the object image, and the image block is different from the surrounding image areas and can be separated from the surrounding image areas. According to the step of image feature generation, one or more image features are generated from at least one image block and are used for forming an image feature pool of the object classifier.
Description
Technical field
The present invention relates to the object detection technique field, be specifically related to a kind of for produce characteristics of image from object images method and apparatus, be used for method, object classification device, object-tracking systems and image capture device from one or more object images and the acquistion of one or more non-object iconology to the object classification device.
Background technology
Object detection technique is widely used in many situations.Use the equipment of object detection to be referred to as the object classification device.The object classification device is used for the object (for example, subject area and/or object images) of recognition image and/or video.Object identifiers available software, hardware, firmware and/or their combination are implemented.
Therein in the example of the camera of application sorter, the object in the captured video of camera is specified by the operator, is perhaps automatically specified.Object can be the subject area that comprises people's face or spend in the frame of video.The object classification device will be followed the tracks of the object in the video.
In the prior art, many object detection technique schemes have been proposed.
In the technical scheme of prior art, specific characteristic pond at first.Feature pool comprises the feature that can be used for characterizing object.Feature in the feature pool produces at random, and is irrelevant with special object.For example, feature can be that class Lis Hartel well known in the art is levied (Haar-like features), HOG feature or LBP feature.Then, by learn, feature in the use characteristic pond and obtain the object classification device with a group objects image and one group of non-object image.This process also is named as study and processes.
Usually, the size of feature pool is very large, generally comprises several ten thousand features.Therefore, the study processing is very consuming time.Usually, will spend several days or a few week learns to obtain the object classification device.
Such object detection technique scheme described in P.Viola and M.Jones article " Rapid Object Detection Using a Boosted Cascade of Simple Features " and the article " Fast and Robust Classification Using Asymmetric AdaBoost and a Detector Cascade " of P.Viola in NIPS 2001 in CVPR 2001, and these articles are hereby incorporated herein by reference.
In some cases, people expect to learn to process and can finish fast.For example, if people thinks the object in the captured video of tracking camera, then he will expect that the object classification device in the camera can obtain by study very fast so that it can be in capturing video tracing object.Such object classification device is referred to as to have the online object classification device of Real time Efficiency in the art.
For example, so online object detection technique scheme described in H.Grabner and H.Bischof article " On-line Boosting and Vision " and H.Grabner, M.Grabner and the article " Real-time Tracking via On-line Boosting " of H.Bischof in BMVC 2006 in CVPR 2006, and these articles are hereby incorporated herein by reference.
Disclose a kind of human face expression recognition system in submission on June 17th, 2004 and in the U.S. Patent application No.2005/0102246 A1 of announcement on May 12nd, 2005, this U.S. Patent application is hereby incorporated herein by reference.
Disclose a kind of object detection systems in submission on March 20th, 2008 and in the U.S. Patent application No.2008/0232681 A1 of announcement on September 25th, 2008, this U.S. Patent application is hereby incorporated herein by reference.
U.S. Patent application No.2005/0102246 A1 and U.S. Patent application No.2008/0232681 A1 disclose the technical scheme of selecting character subset from large-scale feature pool.
In the prior art, by only using the very little character subset of selecting at random from feature pool, rather than whole feature pool improves learning efficiency.But, because the distinguishing ability of the feature of selecting at random usually very a little less than, so the non-constant of detection performance of such object detector.
Summary of the invention
The present inventor has proposed to solve a kind of new technology of at least one problem of the prior art.
An object of the present invention is to provide a kind of technical scheme for produce characteristics of image from object images.
According to a first aspect of the invention, a kind of method for produce characteristics of image from object images is provided, the method comprises: image block (image blob) detecting step, be used for detecting at least one image block in described object images, described at least one image block different with image-region around it and can with its on every side image-region distinguish; And characteristics of image generation step, be used for producing one or more characteristics of image from described at least one image block, wherein, described one or more characteristics of image are used to form the characteristics of image pond for the object classification device.
Preferably, described one or more characteristics of image is HOG feature or LBP feature.
Preferably, described one or more characteristics of image are that the class Lis Hartel is levied, and described characteristics of image generation step comprises: determine the cell size that the class Lis Hartel is levied according to the image block size; Determine coordinate and the type that described class Lis Hartel is levied according to the coordinate at the center of described image block and the relation between described image block and its at least one adjacent area.
Preferably, in described image block detecting step, described at least one image block detects by quick Hessian detecting device (fast-Hessian detector) or Harris detecting device.
Preferably, described characteristics of image generation step comprises: the mean intensity in the image block of determining to detect is as the first mean intensity; Mean intensity in one or more adjacent areas of definite image block that detects each is as the second mean intensity; Select the gray-level difference between wherein said the first mean intensity and described the second mean intensity to be equal to or greater than the adjacent area of predetermined threshold as selected adjacent area from described one or more adjacent areas; And produce one or more characteristics of image from the piece and the selected adjacent area that detect.
According to a second aspect of the invention, provide a kind of for the method from one or more object images and the acquistion of one or more non-object iconology to the object classification device, the method comprises: by using described one or more object images to produce one or more characteristics of image with method according to a first aspect of the invention; Described one or more characteristics of image are added in the characteristics of image pond; And by learning to obtain described object classification device with described one or more object images, described one or more non-object images and described characteristics of image pond.
Preferably, the described study step that obtains described object classification device comprises: the characteristics of image of selecting to can be used for to distinguish described object images and described non-object image from described characteristics of image pond upgrades described object classification device.
According to a third aspect of the invention we, a kind of equipment for produce characteristics of image from object images is provided, this equipment comprises: the image block detecting unit, it is configured in described object images to detect at least one image block, described at least one image block different with image-region around it and can with its on every side image-region distinguish; And the characteristics of image generation unit, it is configured to produce the one or more characteristics of image about described at least one image block, and wherein, described one or more characteristics of image are used to form the characteristics of image pond for the object classification device.
Preferably, described one or more characteristics of image is HOG feature or LBP feature.
Preferably, described one or more characteristics of image are that the class Lis Hartel is levied, and described characteristics of image generation unit comprises: first determines subelement, and it is configured to determine the cell size that the class Lis Hartel is levied according to the image block size; Second determines subelement, and it is configured to determine coordinate and the type that described class Lis Hartel is levied according to the relation between the coordinate at the center of described image block and described image block and its at least one adjacent area.
Preferably, described image block detecting unit is by detecting described at least one image block with quick Hessian detecting device or Harris detecting device.
Preferably, described characteristics of image generation unit comprises: the 3rd determines subelement, and the mean intensity in its image block that is configured to determine to detect is as the first mean intensity; The 4th determines subelement, and it is configured to determine that mean intensity in one or more adjacent areas of the image block that detects each is as the second mean intensity; The chooser unit, it is configured to select the gray-level difference between wherein said the first mean intensity and described the second mean intensity to be equal to or greater than the adjacent area of predetermined threshold as selected adjacent area from described one or more adjacent areas; And the generation subelement, it is configured to produce one or more characteristics of image from the piece that detects and selected adjacent area.
According to a forth aspect of the invention, provide a kind of can be by the object classification device of learning to obtain from one or more object images and one or more non-object image, this object classification device comprises: equipment according to a third aspect of the invention we, and it is configured to by producing one or more characteristics of image with described one or more object images; Unit, pond (pooling), it is configured to described one or more characteristics of image are added in the characteristics of image pond; And unit, it is configured to by learning to obtain described object classification device with described one or more object images, described one or more non-object images and described characteristics of image pond.
Preferably, the described unit characteristics of image that is configured to select from described characteristics of image pond to can be used for to distinguish described object images and described non-object image upgrades described object classification device.
According to a fifth aspect of the invention, a kind of object-tracking systems that can be used for following the tracks of by detection the object in the video is provided, this object-tracking systems comprises: the object determining unit, it is configured to determine the subject area that comprises object in the frame of described video as object images, and determines that from the described frame of described video the zone different from described subject area is as the non-object image; Object classification device according to a forth aspect of the invention, it is configured to by study and uses described object images and described non-object image to be obtained; Wherein, described object classification device is configured to detect the subject area in the frame of back of described video.
Preferably, described object determining unit is configured to further determine that subject area in the frame of the described back that described object classification device is detected is as object images, and determine the zone except described subject area in the frame of described back as the non-object zone, and described object classification device is configured to further be obtained by learning and use object images in the frame of described back and non-object image.
According to a sixth aspect of the invention, provide a kind of image capture device, this image capture device comprises the object-tracking systems according to fifth aspect present invention, is used for following the tracks of the object of the image of catching.
The present inventor finds that in the prior art, the feature in the feature pool is predetermined, rather than produces from object images, so the present invention is a kind of new technical scheme.
By referring to the detailed description of accompanying drawing to exemplary embodiment of the present invention, it is clear that further feature of the present invention and advantage thereof will become.
Description of drawings
The accompanying drawing that is bonded in the instructions and consists of the part of instructions shows embodiments of the invention, and is used from together with its explanation one and explains principle of the present invention.
Fig. 1 is the block diagram of example of hardware configuration that shows the computing system of the embodiment that can be used in the present invention.
Fig. 2 shows the process flow diagram that is used for producing from object images the method for characteristics of image according to the first embodiment of the present invention.
Fig. 3 shows the block diagram that is used for producing from object images the equipment of characteristics of image according to the first embodiment of the present invention.
Fig. 4 shows the process flow diagram of characteristics of image generation step according to a second embodiment of the present invention.
Fig. 5 shows the block diagram of characteristics of image generation unit according to a second embodiment of the present invention.
Fig. 6 shows that the characteristics of image of a third embodiment in accordance with the invention produces the process flow diagram of step.
Fig. 7 shows the block diagram of the characteristics of image generation unit of a third embodiment in accordance with the invention.
Fig. 8 shows the process flow diagram from one or more object images and the acquistion of one or more non-object iconology to the method for object classification device of being used for of a fourth embodiment in accordance with the invention.
Fig. 9 shows the block diagram of the object classification device of a fourth embodiment in accordance with the invention.
Figure 10 shows the block diagram of tracker according to a fifth embodiment of the invention.
Figure 11 shows the block diagram of image capture device according to a sixth embodiment of the invention.
Figure 12 shows the synoptic diagram according to example of the present invention.
Figure 13 shows the type that the class Lis Hartel is levied.
Figure 14 is the synoptic diagram that the example that the class Lis Hartel levies is shown.
Embodiment
Describe various exemplary embodiment of the present invention in detail now with reference to accompanying drawing.It should be noted that: unless specify in addition, the parts of setting forth in these embodiments and positioned opposite, numeral expression formula and the numerical value of step do not limit the scope of the invention.
Below be illustrative to the description only actually of at least one exemplary embodiment, never as any restriction to the present invention and application or use.
May not discuss in detail for the known technology of person of ordinary skill in the relevant, method and apparatus, but in suitable situation, described technology, method and apparatus should be regarded as the part of instructions.
In all examples with discussing shown here, it is exemplary that any occurrence should be construed as merely, rather than as restriction.Therefore, other example of exemplary embodiment can have different values.
It should be noted that: represent similar terms in similar label and the letter accompanying drawing below, therefore, in case be defined in a certain Xiang Zaiyi accompanying drawing, then in accompanying drawing subsequently, do not need it is further discussed.
＜hardware configuration 〉
Fig. 1 is the block diagram of example of hardware configuration that shows the computing system 1000 of the embodiment that can be used in the present invention.
As shown in Figure 1, computing system comprises calculation element 1110.Calculation element 1110 comprises processing unit 1120, system storage 1130, non-dismountable non-volatile memory interface 1140, removable non-volatile memory interface 1150, user's input interface 1160, network interface 1170, video interface 1190 and the output Peripheral Interface 1195 that connects by system bus 1121.
Non-dismountable nonvolatile memory 1141 such as hard disk is connected with non-dismountable non-volatile memory interface 1140.For example, but non-dismountable nonvolatile memory 1141 storage operating systems 1144, application program 1145, other program module 1146 and some routine datas 1147.
The removable non-volatile storer of being connected with CD-ROM drive such as floppy disk 1151 is connected with removable non-volatile memory interface 1150.For example, floppy disk can be inserted in the floppy disk 1151, and CD (CD) can be inserted in the CD-ROM drive 1155.
The input equipment of being connected with keyboard such as mouse 1161 is connected with user's input interface 1160.
Computing system shown in Figure 1 only is indicative, and anything but in order to limit invention, its application, or uses.
The＜the first embodiment 〉
In the online object detection technique of prior art, at first produce feature pool, then select to be used for some features that study obtains sorter from feature pool.In the online object detection technique of prior art, the feature in the feature pool produces at random, and is irrelevant with object images, and the feature that study obtains sorter is to select at random from such feature pool.The present inventor finds, if the feature that study obtains sorter that is used in the feature pool produces from object images, then the performance of sorter will greatly be improved.In addition, if producing for the piece of the feature of learning from object images in the feature pool, then the performance of sorter will further be improved.Therefore, the present inventor has proposed the present invention.The difference of the present invention and prior art scheme is that in the present invention, the feature that is used for the feature pool of object classification device produces from one or more object images.In addition, all features in the feature pool can be used for learning to obtain the object classification device.That is to say, omitted the step of selecting at random feature.
Describe according to the first embodiment of the present invention with reference to Fig. 2 and Fig. 3.
Fig. 2 shows the method 2000 that is used for producing from object images characteristics of image according to the first embodiment of the present invention.
In step S2100, there is the image block detecting step.In the image block detecting step, in object images, detect at least one image block, described at least one image block different with image-region around it and can with its on every side image-region distinguish.For example, described at least one image block comparable around it image-region bright or dark.
Piece can be the point of interest in the image.In computer vision field, different from image-region around it in the piece presentation video and can with the point that distinguishes of image-region and/or zone around it (for example, piece is than image-region is bright or dark on every side), and piece can be detected by the such point and/or the regional vision module that are intended in the detected image.Each that detects " piece " available (x, y, t) expression, wherein, (x, y) is the position of this piece in image, t is the scale (or size) of this piece.
In the prior art, there are many detecting devices.In example of the present invention, in the image block detecting step, described at least one image block can be by for example Hessian or Harris detecting device detect fast.
Because there is in the art the technical scheme for detection of piece, so will not provide its detailed description in this manual.
In step S2200, exist characteristics of image to produce step.Produce in the step at characteristics of image, produce one or more characteristics of image for described at least one image block.
In the prior art, also nobody finds out or expects to produce characteristics of image for the characteristics of image pond of object classification device with these image blocks.
In the present embodiment, described one or more characteristics of image is used to form the characteristics of image pond for the object classification device.
For example, described one or more characteristics of image can be that HOG feature, LBP feature or class Lis Hartel are levied.
Fig. 3 shows the characteristics of image generation equipment 3000 that is used for producing from object images characteristics of image according to the first embodiment of the present invention.
As shown in Figure 3, characteristics of image generation equipment 3000 comprises image block detecting unit 3100 and characteristics of image generation unit 3200.
Image block detecting unit 3100 is configured in object images to detect at least one image block, described at least one image block different with image-region around it and can with its on every side image-region distinguish.
For example, image block detecting unit 3100 can be configured to by detecting described at least one image block with quick Hessian detecting device or Harris detecting device.
Characteristics of image generation unit 3200 is configured to produce one or more characteristics of image for described at least one image block.
In the present embodiment, described one or more characteristics of image is used to form the characteristics of image pond for the object classification device.
For example, described one or more characteristics of image can be that HOG feature or LBP feature or class Lis Hartel are levied.
The＜the second embodiment 〉
Describe according to a second embodiment of the present invention with reference to Fig. 4 and Fig. 5.
The characteristics of image that the second embodiment has further defined among the first embodiment produces step 2200 and characteristics of image generation unit 3200.In a second embodiment, described one or more characteristics of image is that the class Lis Hartel is levied.
Fig. 4 shows the process flow diagram of characteristics of image generation step 4000 according to a second embodiment of the present invention.Characteristics of image produces step 4000 and produces step 2200 corresponding to the characteristics of image among the first embodiment, and comprises step S4100 and S4200.
In step S4100, determine the cell size that the class Lis Hartel is levied according to the image block size.Image block can be one of described at least one image block that detects in image block detecting step 2100 in the first embodiment.
The class Lis Hartel is levied and can be formed by two, three or four equal rectangles.The image block that detects can be one of these rectangles.Therefore, the class Lis Hartel cell size of levying can be determined according to the image block size.For example, the class Lis Hartel cell size of levying can be the twice, three times or four times of image block size.
In step S4200, determine coordinate and the type that the class Lis Hartel is levied according to the coordinate at the center of image block and the relation between image block and its at least one adjacent area.
The piece that detects can form the class Lis Hartel with its at least one adjacent area and levy.Therefore, the class Lis Hartel type of levying can be determined according to the relation between image block and its at least one adjacent area.
By these steps, can produce one or more class Lis Hartels from described at least one image block and levy.
Fig. 5 shows the block diagram of characteristics of image generation unit 5000 according to a second embodiment of the present invention.
Characteristics of image generation unit 5000 is corresponding to the characteristics of image generation unit 3200 among the first embodiment, and can comprise that first determines that subelement 5100 and second determines subelement 5200.
First determines that subelement 5100 can be configured to determine the cell size that the class Lis Hartel is levied according to the image block size.
Second definite subelement 5200 can be configured to determine coordinate and the type that the class Lis Hartel is levied according to the coordinate at the center of image block and the relation between image block and its at least one adjacent area.
Characteristics of image generation unit 5000 can be by determining that with first subelement 5100 and second definite subelement 5200 come to produce one or more class Lis Hartels from described at least one image block and levy.
The＜the three embodiment 〉
With reference to Fig. 6 and Fig. 7 a third embodiment in accordance with the invention is described.
The characteristics of image that the 3rd embodiment has further defined among the first embodiment produces step 2200 and characteristics of image generation unit 3200.
Fig. 6 shows that the characteristics of image of a third embodiment in accordance with the invention produces the process flow diagram of step 6000.Characteristics of image produces step 6000 and produces step 2200 corresponding to the characteristics of image among the first embodiment, and comprises step S6100, S6200, S6300 and S6400.
In step S6100, the mean intensity in the image block of determining to detect is as the first mean intensity.
In step S6200, the mean intensity in each in one or more adjacent areas of definite image block that detects is as the second mean intensity.
In step S6300, selecting wherein from described one or more adjacent areas, the gray-level difference between the first mean intensity and the second mean intensity is equal to or greater than the adjacent area of predetermined threshold as selected adjacent area.
In step S6400, produce one or more characteristics of image from piece and the selected adjacent area that detects.
Fig. 7 shows the block diagram of the characteristics of image generation unit 7000 of a third embodiment in accordance with the invention.
Characteristics of image generation unit 7000 is corresponding to the characteristics of image generation unit 3200 among the first embodiment, and can comprise that the 3rd determines that subelement 7100, the 4th determines subelement 7200, chooser unit 7300 and produce subelement 7400.
Mean intensity in the image block that the 3rd definite subelement 7100 can be configured to determine to detect is as the first mean intensity.
The 4th determines that subelement 7200 can be configured to determine that mean intensity in one or more adjacent areas of the image block that detects each is as the second mean intensity.
Producing subelement 7400 is configured to produce one or more characteristics of image from the piece and the selected adjacent area that detect.
In the 3rd embodiment, be not that piece all adjacent areas on every side all are selected to form feature with this piece.Only select and the visibly different zone of this piece.In this way, can reduce the quantity of feature, and produce the feature that characterizes best object.In this way, the speed of object classification device and precision can be improved, and therefore, the performance of object classification device can be improved.
For example, the 3rd embodiment can make up with the second embodiment.For example, at least one adjacent area among the step S4200 can be the adjacent area of selecting in step S6300, and step S6400 can comprise step S4100 and S4200.For example, second determines that subelement 5200 employed at least one adjacent area can be chooser unit 7300 selected adjacent areas, and generation subelement 7400 can comprise that first determines subelement 5100 and second definite subelement 5200.
The＜the four embodiment 〉
With reference to Fig. 8 and Fig. 9 a fourth embodiment in accordance with the invention is described.
Fig. 8 shows the process flow diagram from one or more object images and the acquistion of one or more non-object iconology to the method 8000 of object classification device of being used for of a fourth embodiment in accordance with the invention.
In step S8100, by using the method according to this invention, produce one or more characteristics of image with described one or more object images.For example, by produce described first or a plurality of characteristics of image with the method described in the first embodiment, the second embodiment, the 3rd embodiment.
In step S8200, described one or more characteristics of image are added in the characteristics of image pond.
In step S8300, obtain the object classification device by study with described one or more object images, described one or more non-object images and described characteristics of image pond.
For example, the study step S8300 that obtains the object classification device can comprise: the characteristics of image of selecting to can be used for to distinguish object images and non-object image from the characteristics of image pond comes the upgating object sorter.
For example, can according to H.Grabner and H.Bischof in CVPR 2006 article " On-line Boosting and Vision " and the online object detection technique scheme described in H.Grabner, M.Grabner and the article " Real-time Tracking via On-line Boosting " of H.Bischof in BMVC 2006 learn to obtain the object classification device.
Fig. 9 shows the block diagram of the object classification device 9000 of a fourth embodiment in accordance with the invention.
Characteristics of image generation equipment 3000 can be that the characteristics of image among the first embodiment produces equipment 3000.Characteristics of image produces equipment 3000 and is configured to produce one or more characteristics of image with described one or more object images.Characteristics of image generation device 3000 can comprise characteristics of image generation unit 5000 or the characteristics of image generation unit 7000 among the 3rd embodiment or their combination among the second embodiment.
Unit, pond (Pooling Unit) 9200 is configured to described one or more characteristics of image are added in the characteristics of image pond.
The characteristics of image that unit 9300 can be configured to select from the characteristics of image pond to can be used for to distinguish object images and non-object image comes the upgating object sorter.
In the present embodiment, being used for the characteristics of image pond that study obtains the object classification device produces from object images.Therefore, performance can be improved as mentioned above like that.
Object images in the present embodiment and non-object image can be the zones in image and/or the image.
The＜the five embodiment 〉
Figure 10 shows the block diagram of tracker 10000 according to a fifth embodiment of the invention.
Object-tracking systems 10000 can be used for the object follow the tracks of in the video by detecting.Object-tracking systems 10000 comprises object determining unit 10100 and object classification device 9000.
Object determining unit 10100 can be configured to determine the subject area that comprises object in the frame of video as object images, and determines that from this frame of this video zone except this subject area is as the non-object image.
For example, object images can be specified by the operator.Scheme as an alternative, subject area can be determined automatically by object-tracking systems 10000.The non-object image also can be specified by the operator.Scheme as an alternative, the non-object image can be determined automatically by object-tracking systems 10000.
For example, the moving window scheme can be used for following the tracks of the subject area in the frame of back of video.For each moving window in the frame of back, object classification device 9000 is determined the confidence value in detected object zone, and the subject area in the frame of the back of video is determined according to these confidence values.How to it will be appreciated by those skilled in the art that by coming the detected object zone with the object classification device, and be not the part that the present invention is concerned about by come the detected object zone with the object classification device how.Therefore, will it be described here.
Object determining unit 10100 can further be configured to determine subject area in the frame of the back that the object classification device is detected as object images and be determined that the zone except subject area in the frame of this back is as the non-object image.In this case, object classification device 9000 can be configured to further obtain by study and with the object images in the frame of back and non-object image.May need not object classification device 9000 further from behind frame produce characteristics of image.Object classification device 9000 can use the characteristics of image pond that produces from the first frame, and wherein the operator has specified object images in this first frame.
In the 5th embodiment, the object classification device can be used for detecting the object in each frame of (tracking) video, and then other zone in the institute's detected object in each frame and each frame can be used for study and obtains the object classification device.So the performance of object classification device can be along with the process of the frame of video and is improved gradually.
The＜the six embodiment 〉
Figure 11 shows the block diagram of image capture device 11000 according to a sixth embodiment of the invention.
For example, but the captured video of operator's specify image capture device 11000 or the object in the image, and image capture device 11000 is by following the tracks of this object with object-tracking systems 10000 wherein.
＜example 〉
With reference to Figure 12, Figure 13 and Figure 14 example of the present invention is described.
In this example, the class Lis Hartel is levied and will be used for explaining the present invention.
Figure 12 shows the synoptic diagram according to example of the present invention.Figure 13 shows the type that the class Lis Hartel is levied.Figure 14 is the synoptic diagram that the example that the class Lis Hartel levies is shown.
Figure 12 shows two frames that camera (image capture device) is captured, that is, and and frame 1 and frame 2.Camera can comprise according to tracker of the present invention.Tracker can comprise according to object classification device of the present invention.
But the zone in operator's designated frame 1 is as object images.For example, regional R1 is designated as object images.Zone R1 comprises a people's portrait.Tracker in the camera determines that automatically some non-object zones are as the non-object image.For example, regional R2 and R3 are confirmed as the non-object image.
The object classification device can find piece in regional R1.For example, find piece BL, piece BL from its around image-region different and can with its around image-region distinguish.In this example, this piece is darker than image-region around it.In other example, this piece comparable around it image-region bright.
In Figure 12, the piece BL that is marked as " * " and the adjacent area that is marked as " 0-7 " have been shown with zoomed-in view.
But object classification device a third embodiment in accordance with the invention removes some adjacent areas.For example, adjacent area 0-7 is corresponding to the described one or more adjacent areas among the second embodiment and/or the 3rd embodiment.For example, the mean intensity among the piece BL that the object classification device is determined to detect is as the first mean intensity, and the mean intensity among definite adjacent area 0-7 each is as the second mean intensity.
Sorter compares the second mean intensity among the first mean intensity and the adjacent area 0-7 each.The object classification device is selected wherein from adjacent area 0-7, and the gray-level difference between the first mean intensity and the second mean intensity is equal to or greater than the adjacent area of predetermined threshold as selected adjacent area.
Predetermined threshold can arrange according to practical application.For example, predetermined threshold can be 10 gray levels.
For example, in Figure 12, first mean intensity in zone 2 or 5 and the gray-level difference between the second mean intensity are less than predetermined threshold.Therefore, zone 2 and 5 is not selected.
The object classification device can or produce one or more class Lis Hartels from piece BL and the adjacent area (such as area 0,1,3,4,6,7) as above selected from piece BL and adjacent area 0-7 and levy.
For example, the object classification device is determined the cell size that the class Lis Hartel is levied according to the size of piece BL, and determines coordinate and the type that the class Lis Hartel is levied according to the coordinate at the center of piece BL and the relation between piece BL and its at least one adjacent area (area 0-7).
Figure 13 shows that the class Lis Hartel levies.The class Lis Hartel is levied by two, three or four equal rectangles and is formed.In Figure 13, (A), class Lis Hartel with two rectangles of (A '), (B) and (B ') expression is levied, (C), class Lis Hartel with three rectangles of (C '), (D) and (D ') expression levies, (E) and the class Lis Hartel of (E ') expression with four rectangles levy.
For example, zone 3 and regional " * " (A) the represented class Lis Hartel that can form among Figure 13 is levied.(A ') represented class Lis Hartel that zone 4 and regional " * " can form among Figure 13 is levied.
(B) represented class Lis Hartel that zone 1 and zone " * " can form among Figure 13 is levied.Zone 6 and regional " * " can represent that (B ') the represented class Lis Hartel among Figure 13 levies.
(C) represented class Lis Hartel that zone 1,6 and zone " * " can form among Figure 13 is levied.
(D) represented class Lis Hartel that zone 3,4 and zone " * " can form among Figure 13 is levied.
(E ') represented class Lis Hartel that zone 1,2,4 and zone " * " or zone 3,5,6 and zone " * " can form among Figure 13 is levied.
Figure 14 shows the example that the class Lis Hartel in the image is levied.Feature among Figure 14 (a) is corresponding to (B) among Figure 13.(b) among Figure 14 is corresponding to (D) among Figure 13.(c) among Figure 14 is corresponding to (E ') among Figure 13.(d) among Figure 14 is corresponding to (E) among Figure 13.Feature among Figure 14 (e) is corresponding to (B ') among Figure 13.(f) among Figure 14 is corresponding to (A) among Figure 13.(g) among Figure 14 is corresponding to (B) among Figure 13.Feature among Figure 14 (h) is corresponding to (B) among Figure 13.
The object classification device is levied the class Lis Hartel that produces and is added the class Lis Hartel to and levy in the pond.
The object classification device obtains by learning and using regional R1 (object images), regional R2 and R3 (non-object image) and class Lis Hartel to levy the pond.For example, levy the class Lis Hartel of selecting to can be used for to distinguish object images and non-object image in the pond from the class Lis Hartel and levy the upgating object sorter.
Object-tracking systems in the camera detects regional R1 ' in the frame 2 with the object classification device.
The object classification device can further be learnt to obtain by the regional R1 ' from frame 2 and other zone.
By this way, sorter can by in detected object and by study obtain.
Use
May realize in many ways method of the present invention, equipment, object classification device, object-tracking systems and image capture device.For example, can realize them by any combination of software, hardware, firmware or software, hardware, firmware.The said sequence that is used for the step of described method only is in order to describe, and the step of method of the present invention is not limited to above specifically described order, unless otherwise specify.In addition, in certain embodiments, can be the program that is recorded in the recording medium with the invention process also, these programs comprise for the machine readable instructions that realizes the method according to this invention.Thereby the present invention also covers the recording medium that storage is used for the program of executive basis method of the present invention.
Although by example specific embodiments more of the present invention are had been described in detail, it should be appreciated by those skilled in the art that above example only is in order to describe, rather than in order to limit the scope of the invention.It should be appreciated by those skilled in the art, can in situation about not departing from the scope of the present invention with spirit, above embodiment be made amendment.Scope of the present invention is limited by claims.
Claims (17)
1. one kind is used for comprising from the method for object images generation characteristics of image:
The image block detecting step, it is used for detecting at least one image block in object images, described at least one image block different with image-region around it and can from its on every side image-region distinguish; With
Characteristics of image produces step, and it is used for producing one or more characteristics of image from described at least one image block, and wherein, described one or more characteristics of image are used to form the characteristics of image pond for the object classification device.
2. method according to claim 1, wherein, described one or more characteristics of image are HOG feature or LBP feature.
3. method according to claim 1, described one or more characteristics of image are that the class Lis Hartel is levied, and described characteristics of image produces step and comprises:
Determine the cell size that the class Lis Hartel is levied according to the image block size;
Determine coordinate and the type that described class Lis Hartel is levied according to the coordinate at the center of described image block and the relation between described image block and its at least one adjacent area.
4. method according to claim 1, in described image block detecting step, described at least one image block is detected by quick Hessian detecting device or Harris detecting device.
5. the described method of any one according to claim 1-4, described characteristics of image produce step and comprise:
Mean intensity in the image block of determining to detect is as the first mean intensity;
Mean intensity in one or more adjacent areas of definite image block that detects each is as the second mean intensity;
Select the gray-level difference between wherein said the first mean intensity and described the second mean intensity to be equal to or greater than the adjacent area of predetermined threshold as selected adjacent area from described one or more adjacent areas; With
Produce one or more characteristics of image from the piece and the selected adjacent area that detect.
6. one kind is used for comprising from one or more object images and the acquistion of one or more non-object iconology to the method for object classification device:
By any one the described method in using according to claim 1-5, use described one or more object images to produce one or more characteristics of image;
Described one or more characteristics of image are added in the characteristics of image pond; With
By learning to obtain described object classification device with described one or more object images, described one or more non-object images and described characteristics of image pond.
7. method according to claim 6, wherein, the step that described study obtains described object classification device comprises:
The characteristics of image of selecting to can be used for to distinguish described object images and described non-object image from described characteristics of image pond upgrades described object classification device.
8. one kind is used for comprising from the equipment of object images generation characteristics of image:
The image block detecting unit, it is configured in described object images to detect at least one image block, described at least one image block different with image-region around it and can with its on every side image-region distinguish; With
The characteristics of image generation unit, it is configured to produce one or more characteristics of image for described at least one image block, and wherein, described one or more characteristics of image are used to form the characteristics of image pond for the object classification device.
9. equipment according to claim 8, wherein, described one or more characteristics of image are HOG feature or LBP feature.
10. equipment according to claim 8, wherein, described one or more characteristics of image are that the class Lis Hartel is levied, and described characteristics of image generation unit comprises:
First determines subelement, and it is configured to determine the cell size that the class Lis Hartel is levied according to the image block size;
Second determines subelement, and it is configured to determine coordinate and the type that described class Lis Hartel is levied according to the relation between the coordinate at the center of described image block and described image block and its at least one adjacent area.
11. equipment according to claim 8, wherein, described image block detecting unit is by detecting described at least one image block with quick Hessian detecting device or Harris detecting device.
12. the described equipment of any one according to claim 8-11, wherein, described characteristics of image generation unit comprises:
The 3rd determines subelement, and the mean intensity in its image block that is configured to determine to detect is as the first mean intensity;
The 4th determines subelement, and it is configured to determine that mean intensity in one or more adjacent areas of the image block that detects each is as the second mean intensity;
The chooser unit, it is configured to select the gray-level difference between wherein said the first mean intensity and described the second mean intensity to be equal to or greater than the adjacent area of predetermined threshold as selected adjacent area from described one or more adjacent areas; With
Produce subelement, it is configured to produce one or more characteristics of image from the piece that detects and selected adjacent area.
13. an object classification device, it can be by learning to be obtained from one or more object images and one or more non-object image, and this object classification device comprises:
The described equipment of according to claim 8-12 any one, it is configured to produce one or more characteristics of image with described one or more object images;
The unit, pond, it is configured to described one or more characteristics of image are added in the characteristics of image pond; With
Unit, it is configured to by learning to obtain described object classification device with described one or more object images, described one or more non-object images and described characteristics of image pond.
14. object classification device according to claim 13, wherein, described unit is configured to select and can upgrade described object classification device for the characteristics of image of distinguishing described object images and described non-object image from described characteristics of image pond.
15. an object-tracking systems, it can be used for following the tracks of by detection the object of video, and this object-tracking systems comprises:
The object determining unit, it is configured to determine the subject area that comprises object in the frame of described video as object images, and determines that from the described frame of described video zone except described subject area is as the non-object image;
According to claim 13, or 14 described object classification devices, it is configured to by study and uses described object images and described non-object image to be obtained;
Wherein, described object classification device is configured to detect the subject area in the frame of back of described video.
16. object-tracking systems according to claim 15,
Wherein, described object determining unit is configured to further determine that subject area in the frame of the described back that described object classification device is detected is as object images, and determine the zone except described subject area in the frame of described back as the non-object image, and
Described object classification device is configured to by study and uses object images in the frame of described back and non-object image and further being obtained.
17. an image capture device, it comprises according to claim 15 or 16 described object-tracking systems, is used for following the tracks of the object of the image of catching.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2012101239754A CN103377373A (en) | 2012-04-25 | 2012-04-25 | Image feature generation method and equipment, classifier, system and capture equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2012101239754A CN103377373A (en) | 2012-04-25 | 2012-04-25 | Image feature generation method and equipment, classifier, system and capture equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103377373A true CN103377373A (en) | 2013-10-30 |
Family
ID=49462457
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2012101239754A Pending CN103377373A (en) | 2012-04-25 | 2012-04-25 | Image feature generation method and equipment, classifier, system and capture equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103377373A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104655638A (en) * | 2013-11-25 | 2015-05-27 | 杭州美盛红外光电技术有限公司 | Analytical comparison device and analytical comparison method |
| CN106503724A (en) * | 2015-09-04 | 2017-03-15 | 佳能株式会社 | Grader generating means, defective/zero defect determining device and method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2006134981A1 (en) * | 2005-06-15 | 2006-12-21 | Matsushita Electric Industrial Co., Ltd. | Object detecting device and its learning device |
| US20100027895A1 (en) * | 2006-08-31 | 2010-02-04 | Osaka Prefecture University Public Corporation | Image recognition method, image recognition device, and image recognition program |
| CN101877064A (en) * | 2009-04-30 | 2010-11-03 | 索尼株式会社 | Image classification method and image classification device |
-
2012
- 2012-04-25 CN CN2012101239754A patent/CN103377373A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2006134981A1 (en) * | 2005-06-15 | 2006-12-21 | Matsushita Electric Industrial Co., Ltd. | Object detecting device and its learning device |
| US20100027895A1 (en) * | 2006-08-31 | 2010-02-04 | Osaka Prefecture University Public Corporation | Image recognition method, image recognition device, and image recognition program |
| CN101877064A (en) * | 2009-04-30 | 2010-11-03 | 索尼株式会社 | Image classification method and image classification device |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104655638A (en) * | 2013-11-25 | 2015-05-27 | 杭州美盛红外光电技术有限公司 | Analytical comparison device and analytical comparison method |
| CN106503724A (en) * | 2015-09-04 | 2017-03-15 | 佳能株式会社 | Grader generating means, defective/zero defect determining device and method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Kadir et al. | A comparative study between LBP and Haar-like features for Face Detection using OpenCV | |
| US20150278642A1 (en) | Neural network image representation | |
| Favaretto et al. | Using group behaviors to detect hofstede cultural dimensions | |
| CN106295532B (en) | A kind of human motion recognition method in video image | |
| Qazi et al. | Human action recognition using SIFT and HOG method | |
| Du et al. | Abnormal event detection in crowded scenes based on structural multi-scale motion interrelated patterns | |
| Arunnehru et al. | Motion intensity code for action recognition in video using PCA and SVM | |
| Nabi et al. | Temporal poselets for collective activity detection and recognition | |
| CN108234821A (en) | Detect the methods, devices and systems of the action in video | |
| Pawar et al. | Scene understanding: A survey to see the world at a single glance | |
| Swamy et al. | Indian sign language interpreter with android implementation | |
| US20120314939A1 (en) | Recognizing apparatus and method, program, and recording medium | |
| CN103377373A (en) | Image feature generation method and equipment, classifier, system and capture equipment | |
| Zhang et al. | Fuzzy segmentation and recognition of continuous human activities | |
| Mei et al. | Training more discriminative multi-class classifiers for hand detection | |
| Carvajal et al. | Multi-action recognition via stochastic modelling of optical flow and gradients | |
| Al-Shibli et al. | Face recognition using local graph structure and support vector machine (LGS-SVM) | |
| Nguyen | Salient object detection via objectness proposals | |
| Loos et al. | Detection and identification of chimpanzee faces in the wild | |
| Xu et al. | Crowd density estimation based on improved Harris & OPTICS Algorithm | |
| CN105760881A (en) | Facial modeling detection method based on Haar classifier method | |
| Kumar et al. | Skin based occlusion detection and face recognition using machine learning techniques | |
| Kachurka et al. | From human eye fixation to human-like autonomous artificial vision | |
| KR101791514B1 (en) | Apparatus and Method for Learning on the basis of Adaboost Algorithm | |
| Gao et al. | Real-time multi-view face detection based on optical flow segmentation for guiding the robot |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20131030 |