CN102609682B - Feedback pedestrian detection method for region of interest - Google Patents
Feedback pedestrian detection method for region of interest Download PDFInfo
- Publication number
- CN102609682B CN102609682B CN201210009835.4A CN201210009835A CN102609682B CN 102609682 B CN102609682 B CN 102609682B CN 201210009835 A CN201210009835 A CN 201210009835A CN 102609682 B CN102609682 B CN 102609682B
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- line
- image
- classification
- sorter
- 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.)
- Active
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a feedback pedestrian detection method for a region of interest. The method comprises the following steps of: 1, reading a video image frame; 2, scaling the acquired video image into a series of images to be detected with different scales; 3, extracting image characteristics in a detection window; 4, scanning the images to be detected with different scales by using the detection window, simultaneously judging pedestrian candidate regions in a window to be detected by using a trained offline pedestrian classifier, then fusing the pedestrian candidate regions in the images to be detected under different scales, and finally giving a pedestrian classification result in the original images; 5, inputting the offline pedestrian classification result into an online tracking module, and updating and optimizing an online classifier; 6, integrating the offline pedestrian classification result and the online tracking result to give a pedestrian detection result; and 7, feeding the pedestrian detection result back to an offline pedestrian classification module. The method not only can respond to the changes of the environment, but also can reflect the human body region of interest, so the method can be used for robustly detecting the pedestrians in a variable scene.
Description
Technical field
The present invention relates to intelligent monitoring technology field, particularly the pedestrian detection in intelligent video monitoring system.
Background technology
Pedestrian detection is the important branch of object detection, is computer vision field receives much concern in recent years forward position direction and study hotspot.It is gathered around and has wide practical use in various fields such as intelligent monitor system, driver assistance system, motion analysis, senior man-machine interfaces, as a kind of active safety means that ensure automobile, pedestrains safety, become the common study hotspot of paying close attention to of industrial community and research circle.In recent years, by the unremitting research of researchist, a large amount of outstanding algorithms have been emerged in large numbers in pedestrian detection field, and the precision of pedestrian detection is improved again and again, towards practical direction, constantly stride forward.The research of current line people detection is than having made significant headway in the past, but still find, a kind ofly can in actual environment, not carry out the method for large-scale application.This is mainly because pedestrian detection has its singularity as a special case of object detection, pedestrian's diversity for example, street complex background around etc.
From the point of penetration of research, current research method is broadly divided into three classes: the first kind is the method based on feature.These class methods are devoted to find a kind of feature, can ideally describe human body, and are not subject to the impact of background, illumination, posture as far as possible, pedestrian and background area can be separated under various conditions; Equations of The Second Kind is the method based on multiple location.These class methods are devoted to solve the non-rigid of human body and are blocked the impact on outward appearance, the variation of human body is decomposed into the variation at each position, take the mode of defeating in detail, divide and rule, by the detection at each position of human body is completed to the detection to whole human body; The 3rd class is the method based on various visual angles, and these class methods lay particular emphasis on and solve the apparent difference of pedestrian under different visual angles.Current mainstream research direction is from machine learning angle, and from a large amount of training samples, Automatic Extraction feature, sets up manikin, pedestrian detection problem is turned to the problem of a pattern classification.
Although pedestrian detection has obtained very large breakthrough, distance is real extensive practical still very remote, also has the not solution of a lot of problems, needs us to make great efforts untiringly and explores.Major Difficulties has: 1) complicated background.Because mankind's activity is in extensive range, living environment constantly changes, so the change of background in picture is very large, this has just increased difficulty to human detection virtually, and due in image, general human body only accounts for a small part, so the slight change of background all can have a great impact detecting; 2) human posture is different.Because people is articulate, in image, people's posture otherness is very large, has station, sits, and squat etc., and dressing also differs widely, and having longly has shortly, and the polytrope that these have all caused this classification of human body itself, has caused the difficulty to Human Modeling; 3) occlusion issue.Generally, in the middle of image, human body is not to come out completely, likely by object above, is blocked, or overlaps with other people, and these have all caused the difficulty of human detection; 4) detection speed.Human detection is early stage and the basic step of other application, and therefore it will propose very high requirement to the speed of human detection for follow-up work saves time, and can reach real-time requirement as far as possible, otherwise is difficult in the actual system of its application; 5) scale problem.We know that human body has height, has lowly, apart from the distance of video camera, also can cause human body size variation, can cause the range scale of human body in image or video to alter a great deal.So should process scale problem as detecting system of human body.Only have the above problem that solved, pedestrian detection just can reach gratifying effect.The key of pedestrian detection is pedestrian to be detected quickly and accurately.
Summary of the invention
An area-of-interest reaction type pedestrian detection method, can tackle the pedestrian detection in changing environment in robust ground.The present invention is in conjunction with off-line pedestrian classification and on-line tracing method, on the one hand, off-line pedestrian classification results can not only provide initialization information for online pedestrian follows the tracks of, realize the auto-initiation of following the tracks of, and can be used as the training new samples of online classification device, and then upgrade and optimize online pedestrian's sorter, on-line tracing sorter just can be upgraded along with the variation of target and background like this.On the other hand, the result that online pedestrian follows the tracks of is as the area-of-interest of next frame off-line pedestrian classification, thereby effectively reduces sweep limit and the intrinsic dimensionality of detection window.Therefore, the present invention can the dynamic variation of response environment can reflect human region interested again, can not only improve pedestrian detection speed and accuracy, and can successfully manage scene change.The inventive method comprises the following steps:
Step 1, the extraction of positive and negative sample image feature: adopt MIT and INTRA pedestrian's database as training storehouse sample, extract the proper vector of the positive negative sample of figure, the steps include: 1) gradient of computed image in each pixel; 2) by gradient projection in each least unit rectangle, and compute gradient direction; 3) in piece, the gradient in least unit rectangle is normalized; 4) utilize AdaBoost to select the block structure varying in size, the gradient vector cascade in the piece in detection window is formed to last image feature vector, wherein the large young pathbreaker of detection window is identical with the positive sample of off-line pedestrian sorter;
Step 2, off-line pedestrian classifier design and training: off-line pedestrian sorter adopts AdaBoost sorting algorithm to select the block structure proper vector varying in size, using the proper vector of selection as the feature screening as the input of support vector machine feature, training off-line pedestrian sorter;
Step 3, is scaled to a series of testing image by the video frame image obtaining (original image) according to bilinear interpolation method;
Step 4, from scanning left to bottom right the testing image different scale, extracts the histograms of oriented gradients of image with detection window at detection window, its step comprises: the 1) gradient of computed image in each pixel; 2) by gradient projection in each least unit rectangle, and compute gradient direction; 3) in piece, the gradient in least unit rectangle is normalized; 4) utilize AdaBoost to select the block structure varying in size, the gradient vector cascade in the piece in detection window is formed to last image feature vector, wherein the large young pathbreaker of detection window is identical with the positive sample of off-line pedestrian sorter;
Step 5 is judged pedestrian with the off-line pedestrian sorter training in window to be measured, provides pedestrian candidate region, then, by pedestrian candidate region in testing image under non-maximal value inhibition method fusion different scale, finally provides pedestrian's classification results in original image;
Step 6, pedestrian's result of off-line pedestrian classification is input to on-line tracing module, on-line tracing module is carried out to initialization, on-line tracing module is using off-line pedestrian classification results as new positive sample, online classification device is trained again as negative sample in other regions beyond pedestrian, thereby upgrade, optimizes online classification tracking.Then in new scene, provide tracking prediction result;
Step 8, off-line comprehensive analysis of pedestrian's result and tracking results of classifying provided to final pedestrian detection result, the set of off-line pedestrian classification results and the set of tracking results are done to set operation, in conjunction with their classification predicted values separately, finally provide pedestrian detection result simultaneously.By tracking prediction and pedestrian detection result feedback in off-line pedestrian sort module;
Step 9, carries out subsequent treatment to pedestrian detection result;
Accompanying drawing explanation
In order to be illustrated more clearly in the embodiment of the present invention, below the accompanying drawing of required use during the embodiment of the present invention is described is briefly described, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skills, do not paying under creative work prerequisite the more accompanying drawing that can also obtain according to these accompanying drawings.
Accompanying drawing 1 is the process flow diagram that the present invention is directed to the reaction type pedestrian detection of area-of-interest;
Accompanying drawing 2 is process flow diagrams that in the present invention, on-line tracing module online classification device is realized;
Accompanying drawing 3 is process flow diagrams that in the present invention, off-line pedestrian sorter is realized.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is clearly and completely described.Obviously, described example is only the present invention's part embodiment, rather than whole embodiment.Embodiment based in the present invention, those of ordinary skills, not making the every other embodiment obtaining under creative work prerequisite, belong to protection domain of the present invention.
Accompanying drawing 1 is the process flow diagram that the present invention is directed to area-of-interest reaction type pedestrian detection method.The method comprises:
S101, obtain video frame images (original testing image);
Image can be to obtain from the video file of storage herein, also can directly obtain from video camera.Due to illumination and weather condition impact, the brightness of image of obtaining has very large difference, need to carry out luminance proportion to video image (original testing image);
S102, video image is carried out to convergent-divergent processing;
By bilinear interpolation method, original testing image is scaled to the different testing image of a series of large small scale.S
rfor graphical rule stepping, first initialization yardstick starts yardstick S
s=1,, according to image size and detection window size computed image, finish yardstick S
e=min (W
i/ W
n, H
i/ H
n), (W
i, H
iwidth and the height of image, W
n, H
nwidth and the height of detection window), then calculate yardstick progression S
n:
Graphical rule S so
i=[S
s, S
ss
r..., S
n].
S103, with detection window scanning testing image and extract the histograms of oriented gradients of image in detection window;
With detection window from scanning left to bottom right the testing image different scale, at detection window, extract the histograms of oriented gradients of image simultaneously, its step comprises: the 1) gradient of computed image in each pixel, computed image gradient adopts simple one dimension [1,0,1] gradient operator, its effect is best; 2) by gradient projection in each least unit rectangle, and gradient direction is quantified as within the scope of [0-180 °] degree to 9 units; 3) in piece, the gradient in least unit rectangle is normalized, method for normalizing adopts L2-norm; 4) the gradient vector cascade in the piece in detection window is formed to last image feature vector, this proper vector will be off-line pedestrian sorter classification input;
S104, with off-line human body sorter, in detection window, judge pedestrian;
The image direction histogram of gradients of extracting in detection window in S103 is input in off-line human body sorter, and sorter output identification 1 or-1 (pedestrian/non-pedestrian) and predicted value, mark original testing image pedestrian candidate region.Wherein the training sample of off-line pedestrian sorter will adopt popular pedestrian's database at present, and accompanying drawing 3 is realization flow figure of off-line pedestrian sorter, and its specific implementation details will be introduced in the back;
S105, pedestrian candidate region under different scale is merged, mark the band of position of pedestrian in original testing image, providing pedestrian's classification results is pedestrian band of position r
1r
nand predicted value λ={ λ
1λ
n;
In original image, pedestrian can may detect a lot of pedestrian candidate region under different scale, and this merges the pedestrian candidate region under different scale with regard to requiring, and what adopt is that non-maximal value Restrainable algorithms is processed them herein;
S106, online classification are followed the tracks of;
The online classification device tracking pedestrians of classifying in new scene, provides tracking results R
1r
nand predicted value { μ
1μ
n.Wherein on-line tracing module realizes details and will provide in the back introduction, and accompanying drawing 2 is realization flow figure of on-line tracing module;
S107, off-line pedestrian classification and on-line tracing result synthetic determination are provided to final pedestrian detection result P
1p
n
According to off-line classification and Detection results set r={r in new scene
1r
nand predicted value λ={ λ
1λ
nand new scene in tracking results R={R
1r
nand predicted value μ={ μ
1μ
ncomputing and comparative result provide final testing result.For { r
1r
n∩ { R
1r
ntesting result in set is defined as pedestrian region.And whether testing result in R-r set is that the predicted value μ by corresponding is determined in pedestrian region, even μ < 0.5, this testing result is considered as to non-pedestrian region, otherwise, will be considered as pedestrian region.For the testing result in r-R set, have two kinds may, a kind of is that new pedestrian enters into scene and arrived by off-line pedestrian detection of classifier, another kind is off-line sorter flase drop.To this treating method, adopt r-R to gather corresponding λ and first judge, even λ > 0.6, is regarded as pedestrian region; Otherwise, adopt the rectangular area coordinate position of testing result in r-R set to adjudicate, if the edge of rectangular area in image is regarded as the not edge in image, ,Ruo rectangular area, non-pedestrian region, be considered as pedestrian.So just determined final pedestrian detection result P
1p
n, output pedestrian detection result P
1p
nin time, will, tracking prediction and pedestrian detection result feedback in off-line classification and Detection S103, be adjusted graphical rule in off-line pedestrian classification and change progression S
n, determine the scope that in S104, detection window scans simultaneously, these regions are using the area-of-interest of the pedestrian detection as next frame image.
S108, provide pedestrian detection result;
According to the pedestrian detection result P exporting in upper S107
1p
n, in original testing image frame, with different colours, draw pedestrian region, pedestrian detection result in final output video.
Accompanying drawing 2 is on-line tracing module process flow diagrams, and each functions of modules is described below:
S201, extract the off-line pedestrian classification results in old scene, sample is carried out to feature extraction;
According to front n frame (I
1... I
n) pedestrian detection result r
1r
nas the positive sample of on-line tracing module, its non-pedestrian region is around as negative sample, the pixel set { x in sample areas respectively
ibe expressed as a proper vector, adopt feature direction histogram.First the sample point that belongs to pedestrian target is labeled as to positive sample, the sample point that belongs to background is labeled as to negative sample, like this, sample point set { x
ibe labeled as { y
i, for two classification, be exactly y
ivalue is+1 or-1.
S202, design and training cascade classifier;
On-line tracing sorter adopts cascade classifier, method is, every frame all adopts support vector machine as component classifier, thereby obtain the support vector machine classifier of one group of different suitable accuracy, then by AdaBoost method, they are combined into a strong classifier, support vector machine kernel function is selected radial basis function herein
S203, data and the off-line tracking results in new scene obtained in new scene are upgraded optimization cascade classifier;
By testing result r in new scene
1r
nwith tracking results R
1r
ncompare, wherein r={r
1r
nbe offline inspection results set, R={R
1r
nbe tracking results set, and remove the sample in R-r set, element in the new detection sample r-R set of off-line is joined in new Sample Storehouse and upgrades cascade classifier as new samples optimization, finally obtain new sorter.Then in new scene, utilize the new cascade classifier classification tracking pedestrians upgrading, and provide tracking results R
1r
nand tracking sorter predicted value μ
1μ
n
Wherein, in the more new stage of sorter, need to lose the oldest support vector machine classifier, the sorter that simultaneously adds up-to-date training to obtain, before adding new component classifier, need to carry out weighting again to each old component classifier.No matter in weighting again, or during according to the new sorter of new sample training, if the error of this component classifier is greater than 0.5, so just lose this component classifier.
Accompanying drawing 3 is that off-line pedestrian sorter is concrete implementing procedure figure, and this module comprises:
S301, the positive negative sample pedestrian of normalization database;
Select the static pedestrian's database of at present popular MIT and INRIA, these databases have adopted various attitudes, background, the thousands of pictures under wearing clothes and blocking.Because target has a long way to go, positive negative sample must be normalized, and positive negative sample is normalized into the image of 64 * 128 sizes, and this will be also detection window size;
S302, extract positive and negative sample image proper vector;
In the positive and negative sample database of S301 normalization, extract image feature vector, its step comprises: the 1) gradient of computed image in each pixel, and computed image gradient adopts simple one dimension [1,0,1] gradient operator, and its effect is best; 2) gradient projection is calculated to gradient direction in each least unit rectangle, and gradient direction is quantified as to 9 units within the scope of [0-180 °] degree, wherein least unit rectangle size is 4 * 4 pixels; 3) in piece, the gradient in least unit rectangle is normalized, method for normalizing adopts L2-norm; 4) the gradient vector cascade in the piece in detection window is formed to last image feature vector;
S303, utilize cascade classifier to select proper vector;
The image feature vector that S302 is extracted is input in cascade classifier, utilize AdaBoost algorithm to select to vary in size the feature of block size to vector, select optimal characteristics, the proper vector that AdaBoost chooses can not only be selected for concrete pedestrian different yardstick pieces, can also the concrete position of fine reflection human body.Wherein piece is wide and high than can be 1: 1,1: 2,2: 1.Desirable 4/6/8 pixel size of they sizes.Utilize support vector machine training human body sorter, the Selection of kernel function linear kernel function in support vector machine.
The above; only for the common embodiment of the present invention, but protection scope of the present invention is not limited to this, is anyly familiar with in technical scope that those skilled in the art disclose in the present invention; the variation that can expect easily or replacement, within all should being encompassed in protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claim.
Claims (10)
1. the reaction type pedestrian detection method for area-of-interest, it is characterized in that, the method that adopts off-line pedestrian classification and on-line tracing to combine, utilize off-line pedestrian classification results to provide initialization information for online pedestrian follows the tracks of, realize the auto-initiation of following the tracks of, and as the training sample of online classification device, upgrade and optimize online pedestrian's sorter; The result of simultaneously online pedestrian being followed the tracks of is as the area-of-interest of next frame off-line pedestrian classification, and described method step comprises:
1) positive and negative sample image feature extraction;
2) design of off-line pedestrian sorter and training;
3) read video frame image;
4) video frame image obtaining is scaled to the testing image of a series of different scales;
5) within the scope of detection window, extract characteristics of image;
6) with the different testing image of detection window scanning yardstick, with the off-line pedestrian sorter training, in window to be measured, judge pedestrian simultaneously, provide the pedestrian candidate region in different scale testing image, then merge pedestrian candidate region in different scale testing image, finally provide pedestrian's classification results in video frame image;
7) result of off-line pedestrian classification is input to on-line tracing module, upgrades and optimize online classification device;
8) comprehensive off-line classification pedestrian's result and on-line tracing result, provide pedestrian detection result;
9) pedestrian detection result feedback is to off-line pedestrian sort module;
2. method according to claim 1, it is characterized in that, in described image characteristic extracting method, adopt the histograms of oriented gradients of image as feature, select the block structure of variable-size, its histograms of oriented gradients extraction step comprises: the 1) gradient of computed image in each pixel; 2) by gradient projection in each smallest blocks structure, and compute gradient direction; 3) in piece, the gradient in smallest blocks structure is normalized; 4) utilize AdaBoost to select the block structure varying in size, the gradient vector cascade in the piece in detection window is formed to last image feature vector, wherein the large young pathbreaker of detection window is identical with the positive sample of off-line pedestrian sorter;
3. method according to claim 1, is characterized in that step 2) off-line pedestrian sorter adopts linear SVM classifier algorithm, and the feature input using image feature vector as support vector machine, obtains off-line pedestrian sorter by training;
4. method according to claim 1, is characterized in that step 4) adopt bilinear interpolation method original testing image to be scaled to the testing image of a series of different scale;
5. method according to claim 1, it is characterized in that, step 6) the pedestrian candidate district in different scale testing image is carried out to area of space fusion, fusion results is off-line pedestrian classification results, and this off-line pedestrian classification results comprises the predicted value of size, position and the off-line sorter in pedestrian region;
6. method according to claim 1, it is characterized in that, step 7) utilize off-line pedestrian classification results to complete the initialization of on-line tracing module, mark tracing area and extract sampling feature vectors, choose the training sample set in scene, this sample set is to be obtained by set operation by the set of off-line pedestrian classification results and the set of on-line tracing result;
7. method according to claim 1, it is characterized in that, step 8) computing of spatial domain similarity is carried out in the set that set off-line pedestrian classification results being formed and tracking results form, this spatial domain similarity is measured by degree of overlapping, its computing method are business of two set common factors and union, the predicted value of business and off-line sorter is comprehensively analyzed and obtained final pedestrian detection result, and wherein this off-line pedestrian classification results comprises the predicted value of size, position and the off-line sorter in pedestrian region;
8. method according to claim 1, is characterized in that step 9) the pedestrian detection result that feeds back in off-line pedestrian classification is size and the position in tracking prediction region;
9. method according to claim 5, is characterized in that, adopts non-maximal value to suppress method the pedestrian candidate region under different scale is merged;
10. method according to claim 6, is characterized in that, the renewal of online classification device and optimization reach by combination supporting vector machine and AdaBoost algorithm.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201210009835.4A CN102609682B (en) | 2012-01-13 | 2012-01-13 | Feedback pedestrian detection method for region of interest |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201210009835.4A CN102609682B (en) | 2012-01-13 | 2012-01-13 | Feedback pedestrian detection method for region of interest |
Publications (2)
Publication Number | Publication Date |
---|---|
CN102609682A CN102609682A (en) | 2012-07-25 |
CN102609682B true CN102609682B (en) | 2014-02-05 |
Family
ID=46527041
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201210009835.4A Active CN102609682B (en) | 2012-01-13 | 2012-01-13 | Feedback pedestrian detection method for region of interest |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN102609682B (en) |
Families Citing this family (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103679187B (en) * | 2012-09-12 | 2018-07-27 | 深圳先进技术研究院 | Image-recognizing method and system |
CN103177248B (en) * | 2013-04-16 | 2016-03-23 | 浙江大学 | A kind of rapid pedestrian detection method of view-based access control model |
CN103886583A (en) * | 2014-02-14 | 2014-06-25 | 杭州电子科技大学 | Target detection sliding window scanning method based on scene geometric constraint |
CN104376302B (en) * | 2014-11-17 | 2017-12-15 | 哈尔滨工业大学深圳研究生院 | One kind utilizes adjacent inter frame image combination single-frame images inspection optimization method based on vehicle-mounted |
CN104680134B (en) * | 2015-02-04 | 2018-01-12 | 四川长虹电器股份有限公司 | Quick human body detecting method |
CN104778453B (en) * | 2015-04-02 | 2017-12-22 | 杭州电子科技大学 | A kind of night pedestrian detection method based on infrared pedestrian's brightness statistics feature |
CN106204632A (en) * | 2015-04-29 | 2016-12-07 | 株式会社理光 | Method for tracking target based on multiple Classifiers Combination and system |
CN105303159A (en) * | 2015-09-17 | 2016-02-03 | 中国科学院合肥物质科学研究院 | Far-infrared pedestrian detection method based on distinctive features |
WO2017059576A1 (en) * | 2015-10-09 | 2017-04-13 | Beijing Sensetime Technology Development Co., Ltd | Apparatus and method for pedestrian detection |
CN106022279A (en) * | 2016-05-26 | 2016-10-12 | 天津艾思科尔科技有限公司 | Method and system for detecting people wearing a hijab in video images |
CN106446832B (en) * | 2016-09-27 | 2020-01-10 | 成都快眼科技有限公司 | Video-based pedestrian real-time detection method |
CN106845352B (en) * | 2016-12-23 | 2020-09-18 | 北京旷视科技有限公司 | Pedestrian detection method and device |
CN108303747B (en) * | 2017-01-12 | 2023-03-07 | 清华大学 | Inspection apparatus and method of detecting a gun |
CN106845387B (en) * | 2017-01-18 | 2020-04-24 | 合肥师范学院 | Pedestrian detection method based on self-learning |
CN106960183A (en) * | 2017-03-04 | 2017-07-18 | 浙江大学 | A kind of image pedestrian's detection algorithm that decision tree is lifted based on gradient |
CN109544603B (en) * | 2018-11-28 | 2021-07-23 | 上饶师范学院 | Target tracking method based on deep migration learning |
CN110222585B (en) * | 2019-05-15 | 2021-07-27 | 华中科技大学 | Moving target tracking method based on cascade detector |
CN110414313B (en) * | 2019-06-06 | 2024-02-13 | 平安科技(深圳)有限公司 | Abnormal behavior alarming method, device, server and storage medium |
CN111274991A (en) * | 2020-02-12 | 2020-06-12 | 清华大学 | Pedestrian detection and identification system, method, device and computer readable storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101221620A (en) * | 2007-12-20 | 2008-07-16 | 北京中星微电子有限公司 | Human face tracing method |
CN101714213A (en) * | 2009-11-27 | 2010-05-26 | 中国科学技术大学 | Vehicle-mounted pedestrian detection method and system |
CN102201059A (en) * | 2011-05-20 | 2011-09-28 | 北京大学深圳研究生院 | Pedestrian detection method and device |
-
2012
- 2012-01-13 CN CN201210009835.4A patent/CN102609682B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101221620A (en) * | 2007-12-20 | 2008-07-16 | 北京中星微电子有限公司 | Human face tracing method |
CN101714213A (en) * | 2009-11-27 | 2010-05-26 | 中国科学技术大学 | Vehicle-mounted pedestrian detection method and system |
CN102201059A (en) * | 2011-05-20 | 2011-09-28 | 北京大学深圳研究生院 | Pedestrian detection method and device |
Also Published As
Publication number | Publication date |
---|---|
CN102609682A (en) | 2012-07-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN102609682B (en) | Feedback pedestrian detection method for region of interest | |
Timofte et al. | Multi-view traffic sign detection, recognition, and 3D localisation | |
CN102831404B (en) | Gesture detecting method and system | |
US7136524B1 (en) | Robust perceptual color identification | |
Tan et al. | Color model-based real-time learning for road following | |
Prieto et al. | Using self-organising maps in the detection and recognition of road signs | |
CA3164893A1 (en) | Systems for multiclass object detection and alerting and methods therefor | |
Cheng et al. | Environment classification and hierarchical lane detection for structured and unstructured roads | |
Cao et al. | Robust vehicle detection by combining deep features with exemplar classification | |
CN114241548A (en) | Small target detection algorithm based on improved YOLOv5 | |
CN101406390A (en) | Method and apparatus for detecting part of human body and human, and method and apparatus for detecting objects | |
Shahab et al. | How salient is scene text? | |
CN109711332B (en) | Regression algorithm-based face tracking method and application | |
CN110119726A (en) | A kind of vehicle brand multi-angle recognition methods based on YOLOv3 model | |
CN101930611A (en) | Multiple view face tracking | |
CN102110227A (en) | Compound method for classifying multiresolution remote sensing images based on context | |
CN107480585A (en) | Object detection method based on DPM algorithms | |
Liu et al. | Multi-type road marking recognition using adaboost detection and extreme learning machine classification | |
Mao et al. | Rapid vehicle logo region detection based on information theory | |
Jia et al. | A fast and efficient green apple object detection model based on Foveabox | |
Paek et al. | K-lane: Lidar lane dataset and benchmark for urban roads and highways | |
CN102063726B (en) | Moving target classification method and system | |
CN107886060A (en) | Pedestrian's automatic detection and tracking based on video | |
CN110334703B (en) | Ship detection and identification method in day and night image | |
CN102682291A (en) | Scene person counting method, device and system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant |