CN102682302B - Human body posture identification method based on multi-characteristic fusion of key frame - Google Patents
Human body posture identification method based on multi-characteristic fusion of key frame Download PDFInfo
- Publication number
- CN102682302B CN102682302B CN201210063893.5A CN201210063893A CN102682302B CN 102682302 B CN102682302 B CN 102682302B CN 201210063893 A CN201210063893 A CN 201210063893A CN 102682302 B CN102682302 B CN 102682302B
- Authority
- CN
- China
- Prior art keywords
- key frame
- frame
- video
- image
- human body
- 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 human body posture identification method based on multi-characteristic fusion of a key frame, comprising the following steps of: (1) extracting Hu invariant moment characteristics from a video image; calculating a covering rate of an image sequence; extracting the highest covering percentage of the covering rate as a candidate key frame; then calculating a distortion rate of the candidate key frame and extracting the minimum distortion percentage as the key frame; (2) carrying out extraction of a foreground image on the key frame to obtain the foreground image of a moving human body; (3) extracting characteristic information of the key frame, wherein the characteristic information comprises a six-planet model, a six-planet angle and eccentricity; obtaining a multi-characteristic fused image characteristic vector; and (4) utilizing a one-to-one trained classification model, wherein the classification model is a posture classifier based on an SVM (Secure Virtual Machine); and identifying a posture. The human body posture identification method has the advantages of simplified calculation, good stability and good robustness.
Description
Technical field
The present invention relates to a kind of human posture recognition method.
Background technology
In recent years, along with China and other emerging economy Urbanization Construction accelerate, the city management problems such as a series of traffic that the simultaneous urbanization that movement of population increases causes, public security, video monitoring is more and more universal, also more and more higher to the intelligentized demand of video monitoring.People wish to extract from video by intellectual analysis each application in the live and work that more information is applied to people, such as safety monitoring, Smart Home, man-machine interaction, sportsman's supplemental training etc.Security protection industry, in the places such as bank, rail, warehouse, need to realize the detection and tracking of moving target automatically, and abnormal alarm situation, thereby reduces all kinds of losses; In house security system and medical monitoring system, can detect in time become a Buddhist monk or nun middle old man or patient whether abnormal conditions occur; In field in intelligent robotics, the analysis that hope can be to the attitude of people in video, gesture, language, carries out mutual exchange and effect with people; In some trainings such as sports, dance training, system can be analyzed athletic joint kinematic parameter, for improvement of athletic training patterns.
Existing human body attitude identification adopts two kinds of methods conventionally: template matching method and state-space method.Template matching method is that image sequence is converted to one group of static in shape pattern, then in identifying, with pre-stored behavior sample, carry out the motion of people in interpretation of images sequence, the advantage of this algorithm is that computation complexity is low, realizes simple, but responsive to the time interval of behavior, poor robustness.State-space method defines each static posture with state-space model and as a state, between these states, by certain probability, connect, each motion sequence can be seen the once traversal between these static posture different conditions as, and calculate the process of its joint probability, this algorithm can be avoided the problem to time interval modeling, but need training sample large, computation complexity is high.
Key frame refers to quantity of information maximum, a representational pair or multiple image in video sequence, the content outline that can reflect one section of video, to accomplish simultaneously as far as possible succinct, data volume is few, therefore in video analysis process, extract key frame and carry out gesture recognition and improve significantly video analysis efficiency.Conventional key-frame extraction technology roughly can be divided into: the method based on shot boundary, that video is divided into several independent camera lenses according to scene, get the first frame, last frame or the intermediate frame of every arrangement of mirrors head as key frame, this algorithm is simple, quick, but do not consider the complicacy of vision content, cannot represent the video segment that content is longer; Method based on video content analysis, to utilize characteristics of image, the feature difference degree of judgement video frame image and present frame, difference is more key frame, this class algorithm can obtain good effect for the video of different length and content, but for camera motion and video content, change when more violent, the key frame of extraction will be very unstable; Method based on motion analysis, the Typical Representative of this class algorithm has optical flow method, and the calculated amount of such algorithm is larger, strong for local movable information dependence; Method based on cluster, such algorithm is by cluster analysis, extracts cluster centre as key frame, and its advantage is reflecting video content preferably, but its algorithm is complicated, and poor stability.
Summary of the invention
In order to overcome algorithm complexity, the poor stability of existing human posture recognition method, the deficiency of poor robustness, the invention provides a kind ofly simplify calculating, have good stability, the human posture recognition method of many Fusion Features based on key frame that robustness is good.
The technical solution adopted for the present invention to solve the technical problems is:
A human posture recognition method for many Fusion Features based on key frame, described human posture recognition method comprises the following steps:
(1) video image is extracted to Hu invariant moment features, the coverage rate of sequence of computed images, extracting setting that coverage rate is the highest, to cover percentage be candidate's key frame, the distortion rate of calculated candidate key frame then take that to extract the wherein distortion percentage of minimum be key frame;
(2) key frame is carried out to the extraction of foreground image, obtain after the foreground image of movement human;
(3) extract the characteristic information of key frame, described characteristic information is six star models, six star angle and eccentricities; Obtain the image feature vector of many Fusion Features;
(4) use the man-to-man disaggregated model training, described disaggregated model is the attitude sorter based on SVM, and attitude is identified.
Further, in described step (4), for N class attitude, between any two class samples, design a SVM, therefore just need to design N* (N-1)/2 SVM.
Further again, in described step (4), the training process of disaggregated model is as follows:
(4.1) first the foreground image of human body is carried out to pre-service, for different behavior video segments, first it is carried out to the background training based on code book, by the image difference of video sequence frame and background model, obtain the foreground image of movement human, and carry out morphologic image processing, remove picture noise;
(4.2) extract the feature of the training data that comprises 11 kinds of attitudes, the feature database that forms master sample, use the method that multi-pose merges to carry out posture feature description to movement human, described 11 kinds of attitudes are respectively walking, little jump, jump greatly, sidle, squat down, bend over, creep, push-up, sit-ups and sit down;
(4.3), by the study to master sample, build the multicategory classification model based on SVM;
(4.4) utilize test sample book to verify model, if lower than the accuracy of estimating, adjust training sample, turn back to (4.1), until higher than the accuracy of estimating, the disaggregated model after being trained.
Further, in described step (3), the leaching process of described six star models: extract after the barycenter of human body foreground image, the human body silhouettes point obtaining and the distance between center of mass point, human body outline profile is divided into left and right two parts, calculates respectively the distance of the going up most of its two parts silhouette, the most lower and Far Left point, rightmost point and barycenter; At six star models, obtained, after the axis of the heart point that six point arrive, calculating the angle information between axis and adjacent axis; Calculate the eccentricity of human body silhouette.
In described step (1), calculate the process of coverage rate: first calculate the likeness coefficient of every two frames, the similarity of other frames in present frame and video is greater than in the associated frame set of listing present frame in of likeness coefficient mean value of present frame and other frames; In associated frame set, in frame and video, the ratio of all frame numbers is the coverage rate of present frame, asks for the coverage rate of each two field picture in video, and gets 30% frame that coverage rate is the highest as candidate's key frame;
The process of calculated distortion rate: the estimated probability of the gray-scale value of computed image first, and the gray level average of image, then calculate single order, second order, the third moment of each two field picture, submeter represents average, variance and measure of skewness, by the feature of this tri-vector presentation video, and calculate the average square of all frames in the average square of associated frame set of key frame and video; The last average square of associated frame set of calculated candidate key frame and the objective function maximal value of the average square of all frames, thus the distortion rate of candidate's key frame obtained, get the 50% final key frame as this video of minimum.
Beneficial effect of the present invention is mainly manifested in: (1) adopts the attitude of many Fusion Features to describe operator in algorithm, human body attitude can better be described, strengthen describing the embodiment ability to attitude of operator, can set up more accurate attitude mode and recognition result; (2) in video analysis process, added the concept of key frame, by key frame, described the whole content of video, to the posture analysis of key frame, can make video analysis more efficient.
Accompanying drawing explanation
Fig. 1 is the process flow diagram of the human posture recognition method of the many Fusion Features based on key frame.
Fig. 2 is the process flow diagram of the extraction method of key frame based on video content.
Fig. 3 is six star aspect of model schematic diagram of many characteristic models.
Fig. 4 is six star angle character schematic diagram of many characteristic models.
Fig. 5 is the eccentricity schematic diagram of many characteristic models.
Fig. 6 is the schematic diagram of the SVM that designs one to one.
Embodiment
Below in conjunction with accompanying drawing, the invention will be further described.
With reference to Fig. 1～Fig. 6, a kind of human posture recognition method of the many Fusion Features based on key frame, described human posture recognition method comprises the following steps:
(1) video image is extracted to Hu invariant moment features, the coverage rate of sequence of computed images, extracting setting that coverage rate is the highest, to cover percentage be candidate's key frame, the distortion rate of calculated candidate key frame then take that to extract the wherein distortion percentage of minimum be key frame;
(2) key frame is carried out to the extraction of foreground image, obtain after the foreground image of movement human;
(3) extract the characteristic information of key frame, described characteristic information is six star models, six star angle and eccentricities; Obtain the image feature vector of many Fusion Features;
(4) use the man-to-man disaggregated model training, described disaggregated model is the attitude sorter based on SVM, and attitude is identified.
In the present embodiment, the training process of attitude mode comprises: the study of the expression that many feature extractions of attitude master sample, attitude are described operator and the multicategory classification device based on support vector machine (SVM:Support Vector Machine); The process of gesture recognition comprises: expression and gesture recognition that many feature extractions of content-based key-frame extraction, key frame, attitude are described operator, as shown in Figure 1, wherein content-based key-frame extraction, attitude based on multi-feature fusion represent and the gesture recognition based on SVM is the gordian technique in this method method structure.
Content-based key-frame extraction: complicacy and the stability of considering algorithm, the present invention proposes a kind of Key-frame Extraction Algorithm based on video content analysis, by calculating coverage rate and the distortion rate of sequence of frames of video, selects the maximum keyframe sequence of inclusion information in video sequence.
Because video image can exist certain geometric distortion in shooting process, and geometric distortion will be brought very large impact for image recognition, therefore need a kind of method with Invariant to rotation and scale, in this method with Hu not the characteristics of image of bending moment extract key frame, the Key-frame Extraction Algorithm flow process based on bending moment is not as shown in Figure 2.
(1) Hu bending moment not
Hu not bending moment is to put forward for 1962, has translation, rotation and yardstick indeformable, can in enough squares space, convert and express and analysis image coordinate transform accordingly.Under image discrete state, ask the multistage common Ju He of image center square.When image changes, common square can change, still responsive to rotation but centre distance has translation invariance.Directly with common Ju Huo center square, carry out character representation, can not make feature there is translation, Invariant to rotation and scale simultaneously.If utilize normalization center square, feature not only has translation invariance, but also has constant rate.
Hu utilizes second order and third central moment to construct seven not bending moments, and they can keep translation, convergent-divergent and invariable rotary under consecutive image condition.Video image is extracted to not bending moment of Hu, and the feature that obtains image is described operator.
(2) coverage rate
Calculate after the Hu invariant moment features of image, utilize the coverage rate of the every sub-picture of this feature calculation, extract candidate's key frame of video.Calculate the process of coverage rate, first calculate the likeness coefficient of every two frames, the similarity of other frames in present frame and video is greater than in the associated frame set of listing present frame in of likeness coefficient mean value of present frame and other frames.In associated frame set, in frame and video, the ratio of all frame numbers is the coverage rate of present frame, asks for the coverage rate of each two field picture in video, and gets 30% frame that coverage rate is the highest as candidate's key frame.
(3) distortion rate
Coverage rate has been calculated the probability that each frame can represent the information of other frames, but cannot from candidate's key frame, tell which frame is best key frame, therefore we are by calculating the distortion rate between other frames in each candidate frame and video, and distortion rate is lower is chosen as key frame.
First the estimated probability of the gray-scale value of computed image, and the gray level average of image, then calculate single order, second order, the third moment of each two field picture, submeter represents average, variance and measure of skewness, by the feature of this tri-vector presentation video, and calculate the average square of all frames in the average square of associated frame set of key frame and video.The last average square of associated frame set of calculated candidate key frame and the objective function maximal value of the average square of all frames, thus the distortion rate of candidate's key frame obtained, get the 50% final key frame as this video of minimum.
Attitude based on multi-feature fusion represents: by calculating above after the key frame of video sequence, extract the characteristic information of key frame, and carry out gesture recognition.The characteristic information adopting is herein the human body attitude feature of six star models, six star angles and the many Fusion Features of eccentricity, thereby obtains the image feature vector of many Fusion Features.
With the algorithm that attitude is described and template matches combines of many Fusion Features, the attitude of many Fusion Features is described algorithm can describe out human body attitude comparatively exactly, has overcome the sensitivity of template matching method to the time interval, has strengthened the robustness of method.
(1) six star model
For the different behaviors of human body, in its profile, comprised the abundantest information that can describe the behavior, so this method selects six star models as one of them feature.Six star models extract after the barycenter of human body foreground image, the human body silhouettes point obtaining and the distance between center of mass point.Human body outline profile is divided into left and right two parts, calculates respectively the distance of the going up most of its two parts silhouette, the most lower and the most left (right side) edge point and barycenter, as shown in Figure 3.
(2) six star angles
Second angle that is characterized as six stars in many features, has obtained, after the axis of the heart point that six point arrive, calculating the angle information between axis and adjacent axis at six star models, obtains six angles, as shown in Figure 4.
(3) eccentricity
This method is extracted the eccentricity of human body contour outline as one of feature, calculates human body silhouette eccentricity, thereby has obtained the image feature vector of many Fusion Features, as shown in Figure 5 by formula.
Attitude sorter based on SVM: this method selects the multicategory classification device of support vector to carry out modeling and classification to human body behavior.Support vector machine (Support Vector Machine, SVM) can, the in the situation that of small sample, obtain the good sorter of generalization ability.SVM can be transformed into a linear space by nonlinear feature space by kernel function, then in the feature space after conversion, constructs lineoid, obtains the optimization model between class and class.Due to what extract herein, being the many features that merge, is therefore a nonlinear space, and choice for use radial basis kernel function (Radial Basis Function, RBF) is carried out feature space conversion, finally obtains the optimal classification face between all kinds of.
Because the training sample gathering can be because the error in gatherer process exists certain noise, therefore in support vector machine, define penalty factor and nuclear parameter γ and solved study and the inseparable problem excessively causing due to noise, in training process, by the method for parameter tuning, determine penalty factor and nuclear parameter, thereby obtain an optimal classification face, finally structure obtains SVM multi-class classifier.
By kernel function, proper vector is carried out after higher-dimension conversion, to the training sample study that exercises supervision, with the function in lower dimensional space, calculate sorter.In the present invention, use the multicategory classification device method for designing of (one-versus-one) one to one, for N class attitude, between any two class samples, design a SVM, therefore just need to design N* (N-1)/2 SVM, as shown in Figure 6, there are three kinds of attitude classes, between attitude class 1 and attitude class 2, design a sorter f
1,2(x), design category device f between attitude class 2 and attitude class 3
2,3(x), the sorter f between attitude class 1 and attitude class 3
1,3(x), totally 3 sorters.When a unknown sample is classified, for each sorter, obtain a classification results, finally divide who gets the most votes's classification to be the classification of this unknown sample.
Gathered altogether more than 400 video segment (720*576,30fps), movement human video data and the Weizmann human body behavior public database certainly taken have been comprised, and video data is divided into training data and test data set, its ratio is 2:1, altogether comprised 11 kinds of attitudes, be respectively walking (walk), jump 1(jump1), jump 2(jump2), sidle (sidewalk), squat down (squat), bend over (stoop), creep (crawl), push-up (push-up), sit-ups (sit-up), (sit) sits down.
The training stage of disaggregated model:
(1) first human body prospect is carried out to pre-service, for different behavior video segments, first it is carried out to the background training based on code book, by the image difference of video sequence frame and background model, obtain the foreground image of movement human, and carry out morphologic image processing, remove picture noise;
(2) extract the feature (as shown in Fig. 3,4,5) of the training data that comprises 11 kinds of attitudes, form the feature database of master sample, use the method that multi-pose merges to carry out posture feature description to movement human;
(3), by the study to master sample, build the multicategory classification model based on SVM;
(4) utilize test sample book to verify its correctness to model, lower than the accuracy of estimating, adjust training sample.
To identifying in input video, comprise four actions, have 214 two field pictures, algorithm extracts 32 key frames, realizes the identification of attitude sequence, and as shown in Figure 8, detailed process is as follows:
(1) image is extracted to Hu invariant moment features, for calculating the coverage rate of video image sequence, extract coverage rate the highest 30% be candidate's key frame, the distortion rate of calculated candidate key frame then, take extract wherein 50% be key frame;
(2) key frame is carried out to the extraction of foreground image, obtain after the foreground image of movement human;
(3) in algorithm, use multi-pose to merge human body attitude and describe operator;
(4) with the disaggregated model training, attitude is identified.
Claims (3)
1. a human posture recognition method for the many Fusion Features based on key frame, is characterized in that: described human posture recognition method comprises the following steps:
(1) video image is extracted to Hu invariant moment features, the coverage rate of sequence of computed images, extracting setting that coverage rate is the highest, to cover percentage be candidate's key frame, the distortion rate of calculated candidate key frame then take that to extract the wherein distortion percentage of minimum be key frame; In described step (1), calculate the process of coverage rate: first calculate the likeness coefficient of every two frames, the similarity of other frames in present frame and video is greater than in the associated frame set of listing present frame in of likeness coefficient mean value of present frame and other frames; In associated frame set, in frame and video, the ratio of all frame numbers is the coverage rate of present frame, asks for the coverage rate of each two field picture in video, and gets 30% frame that coverage rate is the highest as candidate's key frame;
The process of calculated distortion rate: the estimated probability of the gray-scale value of computed image first, and the gray level average of image, then calculate single order, second order, the third moment of each two field picture, submeter represents average, variance and measure of skewness, by the feature of this tri-vector presentation video, and calculate the average square of all frames in the average square of associated frame set of key frame and video; The last average square of associated frame set of calculated candidate key frame and the objective function maximal value of the average square of all frames, thus the distortion rate of candidate's key frame obtained, get the 50% final key frame as this video of minimum;
(2) key frame is carried out to the extraction of foreground image, obtain after the foreground image of movement human;
(3) extract the characteristic information of key frame, described characteristic information is six star models, six star angle and eccentricities; Obtain the image feature vector of many Fusion Features; In described step (3), the leaching process of described six star models: extract after the barycenter of human body foreground image, the human body silhouettes point obtaining and the distance between center of mass point, human body outline profile is divided into left and right two parts, calculates respectively the distance of the going up most of its two parts silhouette, the most lower and Far Left point, rightmost point and barycenter; At six star models, obtained, after the axis of the heart point that six point arrive, calculating the angle information between axis and adjacent axis; Calculate the eccentricity of human body silhouette;
(4) use the man-to-man disaggregated model training, described disaggregated model is the attitude sorter based on SVM, and attitude is identified.
2. the human posture recognition method of the many Fusion Features based on key frame as claimed in claim 1, it is characterized in that: in described step (4), for N class attitude, between any two class samples, design a SVM, therefore just need to design N* (N-1)/2 SVM.
3. the human posture recognition method of the many Fusion Features based on key frame as claimed in claim 1 or 2, is characterized in that: in described step (4), the training process of disaggregated model is as follows:
(4.1) first the foreground image of human body is carried out to pre-service, for different behavior video segments, first it is carried out to the background training based on code book, by the image difference of video sequence frame and background model, obtain the foreground image of movement human, and carry out morphologic image processing, remove picture noise;
(4.2) extract the feature of the training data that comprises 11 kinds of attitudes, the feature database that forms master sample, use the method that multi-pose merges to carry out posture feature description to movement human, described 11 kinds of attitudes are respectively walking, little jump, jump greatly, sidle, squat down, bend over, creep, push-up, sit-ups and sit down;
(4.3), by the study to master sample, build the multicategory classification model based on SVM;
(4.4) utilize test sample book to verify model, if lower than the accuracy of estimating, adjust training sample, turn back to (4.1), until higher than the accuracy of estimating, the disaggregated model after being trained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210063893.5A CN102682302B (en) | 2012-03-12 | 2012-03-12 | Human body posture identification method based on multi-characteristic fusion of key frame |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210063893.5A CN102682302B (en) | 2012-03-12 | 2012-03-12 | Human body posture identification method based on multi-characteristic fusion of key frame |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102682302A CN102682302A (en) | 2012-09-19 |
| CN102682302B true CN102682302B (en) | 2014-03-26 |
Family
ID=46814197
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201210063893.5A Active CN102682302B (en) | 2012-03-12 | 2012-03-12 | Human body posture identification method based on multi-characteristic fusion of key frame |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102682302B (en) |
Families Citing this family (40)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103218831B (en) * | 2013-04-21 | 2015-11-18 | 北京航空航天大学 | A kind of video frequency motion target classifying identification method based on profile constraint |
| CN105164696A (en) * | 2013-05-03 | 2015-12-16 | 诺基亚技术有限公司 | A method and technical equipment for people identification |
| JP6359343B2 (en) * | 2013-07-01 | 2018-07-18 | キヤノンメディカルシステムズ株式会社 | Motion information processing apparatus and method |
| CN104679779B (en) * | 2013-11-29 | 2019-02-01 | 华为技术有限公司 | The method and apparatus of visual classification |
| CN104331712B (en) * | 2014-11-24 | 2017-08-25 | 齐齐哈尔格林环保科技开发有限公司 | A kind of alga cells classification of images method |
| CN105184767B (en) * | 2015-07-22 | 2018-04-06 | 北京工业大学 | A kind of movement human posture method for measuring similarity |
| CN105184257B (en) * | 2015-09-08 | 2018-08-07 | 北京航空航天大学 | Object detection method and device |
| CN106295532B (en) * | 2016-08-01 | 2019-09-24 | 河海大学 | A kind of human motion recognition method in video image |
| CN106599907B (en) * | 2016-11-29 | 2019-11-29 | 北京航空航天大学 | The dynamic scene classification method and device of multiple features fusion |
| CN107239728B (en) * | 2017-01-04 | 2021-02-02 | 赛灵思电子科技(北京)有限公司 | Unmanned aerial vehicle interaction device and method based on deep learning attitude estimation |
| CN106980815A (en) * | 2017-02-07 | 2017-07-25 | 王俊 | Facial paralysis objective evaluation method under being supervised based on H B rank scores |
| CN107087211B (en) * | 2017-03-30 | 2020-12-11 | 北京奇艺世纪科技有限公司 | Method and device for detecting lens of host |
| CN107203753B (en) * | 2017-05-25 | 2020-09-08 | 西安工业大学 | Action recognition method based on fuzzy neural network and graph model reasoning |
| CN107330414A (en) * | 2017-07-07 | 2017-11-07 | 郑州轻工业学院 | Act of violence monitoring method |
| CN107483887B (en) * | 2017-08-11 | 2020-05-22 | 中国地质大学（武汉） | Early warning detection method for emergency in smart city video monitoring |
| CN109670520B (en) * | 2017-10-13 | 2021-04-09 | 杭州海康威视数字技术股份有限公司 | Target posture recognition method and device and electronic equipment |
| CN107832713B (en) * | 2017-11-13 | 2021-11-16 | 南京邮电大学 | Human body posture recognition method based on OptiTrack |
| CN107798313A (en) * | 2017-11-22 | 2018-03-13 | 杨晓艳 | A kind of human posture recognition method, device, terminal and storage medium |
| CN108229336B (en) * | 2017-12-13 | 2021-06-04 | 北京市商汤科技开发有限公司 | Video recognition and training method and apparatus, electronic device, program, and medium |
| CN108256433B (en) * | 2017-12-22 | 2020-12-25 | 银河水滴科技（北京）有限公司 | Motion attitude assessment method and system |
| CN108681740A (en) * | 2018-04-04 | 2018-10-19 | 儒安科技有限公司 | Vehicle type classification method based on multi-category support vector machines |
| CN110400332B (en) * | 2018-04-25 | 2021-11-05 | 杭州海康威视数字技术股份有限公司 | Target detection tracking method and device and computer equipment |
| CN108615241B (en) * | 2018-04-28 | 2020-10-27 | 四川大学 | Rapid human body posture estimation method based on optical flow |
| CN109190474B (en) * | 2018-08-01 | 2021-07-20 | 南昌大学 | Human body animation key frame extraction method based on gesture significance |
| CN108965920A (en) * | 2018-08-08 | 2018-12-07 | 北京未来媒体科技股份有限公司 | A kind of video content demolition method and device |
| CN109583340B (en) * | 2018-11-15 | 2022-10-14 | 中山大学 | Video target detection method based on deep learning |
| CN109508684B (en) * | 2018-11-21 | 2022-12-27 | 中山大学 | Method for recognizing human behavior in video |
| CN109858406B (en) * | 2019-01-17 | 2023-04-07 | 西北大学 | Key frame extraction method based on joint point information |
| CN110309720A (en) * | 2019-05-27 | 2019-10-08 | 北京奇艺世纪科技有限公司 | Video detecting method, device, electronic equipment and computer-readable medium |
| CN110457999B (en) * | 2019-06-27 | 2022-11-04 | 广东工业大学 | Animal posture behavior estimation and mood recognition method based on deep learning and SVM |
| CN110490901A (en) * | 2019-07-15 | 2019-11-22 | 武汉大学 | The pedestrian detection tracking of anti-attitudes vibration |
| CN111368810B (en) * | 2020-05-26 | 2020-08-25 | 西南交通大学 | Sit-up detection system and method based on human body and skeleton key point identification |
| WO2021248432A1 (en) * | 2020-06-12 | 2021-12-16 | Beijing Didi Infinity Technology And Development Co., Ltd. | Systems and methods for performing motion transfer using a learning model |
| CN111797714B (en) * | 2020-06-16 | 2022-04-26 | 浙江大学 | Multi-view human motion capture method based on key point clustering |
| CN111783650A (en) * | 2020-06-30 | 2020-10-16 | 北京百度网讯科技有限公司 | Model training method, action recognition method, device, equipment and storage medium |
| CN112090053A (en) * | 2020-09-14 | 2020-12-18 | 成都拟合未来科技有限公司 | 3D interactive fitness training method, device, equipment and medium |
| CN112932470B (en) * | 2021-01-27 | 2023-12-29 | 上海萱闱医疗科技有限公司 | Assessment method and device for push-up training, equipment and storage medium |
| CN112926522B (en) * | 2021-03-30 | 2023-11-24 | 广东省科学院智能制造研究所 | Behavior recognition method based on skeleton gesture and space-time diagram convolution network |
| CN115311324A (en) * | 2022-04-15 | 2022-11-08 | 北京航空航天大学杭州创新研究院 | Attitude tracking method and apparatus, electronic device, and computer-readable medium |
| CN116310015A (en) * | 2023-03-15 | 2023-06-23 | 杭州若夕企业管理有限公司 | Computer system, method and medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101576953A (en) * | 2009-06-10 | 2009-11-11 | 北京中星微电子有限公司 | Classification method and device of human body posture |
| CN101794384A (en) * | 2010-03-12 | 2010-08-04 | 浙江大学 | Shooting action identification method based on human body skeleton map extraction and grouping motion diagram inquiry |
| US20110228976A1 (en) * | 2010-03-19 | 2011-09-22 | Microsoft Corporation | Proxy training data for human body tracking |
| CN102289672A (en) * | 2011-06-03 | 2011-12-21 | 天津大学 | Infrared gait identification method adopting double-channel feature fusion |
-
2012
- 2012-03-12 CN CN201210063893.5A patent/CN102682302B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101576953A (en) * | 2009-06-10 | 2009-11-11 | 北京中星微电子有限公司 | Classification method and device of human body posture |
| CN101794384A (en) * | 2010-03-12 | 2010-08-04 | 浙江大学 | Shooting action identification method based on human body skeleton map extraction and grouping motion diagram inquiry |
| US20110228976A1 (en) * | 2010-03-19 | 2011-09-22 | Microsoft Corporation | Proxy training data for human body tracking |
| CN102289672A (en) * | 2011-06-03 | 2011-12-21 | 天津大学 | Infrared gait identification method adopting double-channel feature fusion |
Non-Patent Citations (2)
| Title |
|---|
| 基于支持向量机的多种人体姿态识别;谢非 等;《重庆工学院学报(自然科学)》;20090331;第23卷(第3期);138-143 * |
| 基于支持向量机的视频监控运动对象识别方法研究;孙斌;《现代计算机》;20111031;42-45,52 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102682302A (en) | 2012-09-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102682302B (en) | Human body posture identification method based on multi-characteristic fusion of key frame | |
| CN110147743B (en) | Real-time online pedestrian analysis and counting system and method under complex scene | |
| CN106815566B (en) | Face retrieval method based on multitask convolutional neural network | |
| Song et al. | Tracking revisited using RGBD camera: Unified benchmark and baselines | |
| JP6025845B2 (en) | Object posture search apparatus and method | |
| CN103942577A (en) | Identity identification method based on self-established sample library and composite characters in video monitoring | |
| CN107424161B (en) | Coarse-to-fine indoor scene image layout estimation method | |
| WO2020050111A1 (en) | Motion recognition method and device | |
| CN105956560A (en) | Vehicle model identification method based on pooling multi-scale depth convolution characteristics | |
| CN104850825A (en) | Facial image face score calculating method based on convolutional neural network | |
| CN107944431A (en) | A kind of intelligent identification Method based on motion change | |
| CN106778474A (en) | 3D human body recognition methods and equipment | |
| CN108280421B (en) | Human behavior recognition method based on multi-feature depth motion map | |
| CN106599785A (en) | Method and device for building human body 3D feature identity information database | |
| CN110390289A (en) | Based on the video security protection detection method for censuring understanding | |
| CN107169117A (en) | A kind of manual draw human motion search method based on autocoder and DTW | |
| CN106611158A (en) | Method and equipment for obtaining human body 3D characteristic information | |
| Tong et al. | Cross-view gait recognition based on a restrictive triplet network | |
| Batool et al. | Telemonitoring of daily activities based on multi-sensors data fusion | |
| CN112906520A (en) | Gesture coding-based action recognition method and device | |
| CN111797705A (en) | Action recognition method based on character relation modeling | |
| CN113378691A (en) | Intelligent home management system and method based on real-time user behavior analysis | |
| CN117333908A (en) | Cross-modal pedestrian re-recognition method based on attitude feature alignment | |
| CN103020631A (en) | Human movement identification method based on star model | |
| Liu et al. | Weighted sequence loss based spatial-temporal deep learning framework for human body orientation estimation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |