CN104731323A - Multi-rotating direction SVM model gesture tracking method based on HOG characteristics - Google Patents
Multi-rotating direction SVM model gesture tracking method based on HOG characteristics Download PDFInfo
- Publication number
- CN104731323A CN104731323A CN201510079498.XA CN201510079498A CN104731323A CN 104731323 A CN104731323 A CN 104731323A CN 201510079498 A CN201510079498 A CN 201510079498A CN 104731323 A CN104731323 A CN 104731323A
- Authority
- CN
- China
- Prior art keywords
- gesture
- svm model
- window
- svm
- frame
- 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.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a multi-rotating direction SVM model gesture tracking method based on HOG characteristics. Images continuously grabbed by a Web camera serve as input images, and the set gesture initial detecting positioning and later tracking positioning are carried out on the frames of the images. A HOG+Multi-SVM detecting module is used for obtaining the set gesture detected and positioned in front several frames of gesture images from the Web camera; under the situation that the image frames include the set gesture, a gesture tracking module of a multi-rotating direction SVM model based on the HOG characteristics is started for tracking positioning; once the tracking module loses the tracking for the set gesture or the gesture is moved out of the shooting range of the Web camera, the HOG+Multi-SVM detecting module is started again till the set gesture can be continuously detected and positioned successfully. The speed for positioning the set gesture can be effectively increased, and meanwhile the positioning precision for the set gesture is also improved.
Description
Technical field
The present invention relates to a kind of gesture tracking method of the many sense of rotation SVM model based on HOG feature, belong to image procossing and mode identification technology.
Background technology
Nowadays, popularizing and developing rapidly of computing machine makes the life of people more and more rely on computing machine, and computing machine is ubiquitous.In people in the past and computer interactive technology, specific input-output device, if any keyboard, mouse, writing pencil, scanner etc., is widely adopted.In recent years, along with increasing substantially of computing power, personal computer (PC) has had the ability of the multiple communication medium such as speech processes, graph and image processing.For improving the ease for use of computing machine and the naturality of man-machine interaction, novel input technology has become a study hotspot field enjoying users and researcher to pay close attention to.Developing by leaps and bounds and new-type input-output device continuing to bring out like the mushrooms after rain of artificial intelligence, make computing machine become more and more intelligent, man-machine interaction also becomes more natural.The tradition that the research of human-computer interaction technology is gone through centered by computing machine is mutual, progressively transfers to now in the novel interaction technique of multiple passage focusing on people, media.The new-type technology such as body posture identification, recognition of face, Expression Recognition, head tracking, eye tracking and gesture identification is all custom for meeting people and the novel human-machine interaction technology occurred.In life, we always use some fixing gesture to be specifically intended to express certain, such as, represent OK, are passerby's direction indication, representative digit etc.These phenomenons all describe the frequency of gesture operation and naturality directly perceived, also make people wish gesture can be used for carrying out with the computing machine of oneself alternately simultaneously.Such as, user can define suitable gesture and controls etc. equipment.
Gesture identification, as multi-mode human-machine interface technology pith, relate to the research field of multiple subject, such as computer vision, artificial intelligence, pattern-recognition etc.The gesture tracking of view-based access control model is the important step of gesture identification method, particularly dynamic hand gesture recognition.The motion amplitude of staff is comparatively large, the change of illumination condition, background difference, in the picture shared region are less and shape constantly changes, and these are gesture tracking and bring great challenge.In addition, gesture tracking algorithm also requires reach real-time and have low consumption.Histograms of oriented gradients (Histogram of Oriented Gradient is called for short HOG) is a kind of Feature Descriptor being used for carrying out object detection in computer vision and image procossing.Support vector machine (Support Vector Machine is called for short SVM) is a kind of machine learning method that can be used as feature space classification.
The method of gesture tracking can utilize diverse ways to realize, mainly comprise the method based on profile, the method based on region, based on the method for template, the method for feature based and the method etc. based on movable information.But these methods can not process the tracing task of gesture under complex background greatly and take into account real-time and accuracy.
Summary of the invention
The object of the invention is to solve conventional sign Rotation in plane, improve the many sense of rotation SVM model of utilization based on HOG feature to the location of conventional sign.
The invention provides a kind of gesture tracking method of the many sense of rotation SVM model based on HOG feature, specifically:
Step 1: obtain picture frame, and carry out pre-service, specifically:
(1) utilize camera to capture camera lens each picture frame front in real time, and bilinear interpolation process is carried out to obtain setting the image of size to the image of present frame;
(2) the SVM model in dictionary corresponding to the correlation parameter of HOG feature, gradient throwing value space and all directions sub-range based on HOG feature is loaded.
Step 2: judge whether successfully locate frame number arrives default frame number threshold value T;
Set a variable x, successfully detect or trace into the frame number containing gesture continuously before being recorded to present frame, the value of initial x is 0.
More successful locating frame number variable x and the frame number threshold value T preset, is performed as follows:
(1) if successful locating frame number variable x does not reach predetermined threshold value T, so, call the detection module based on HOG+Multi-SVM, enter step 3 and perform;
(2) otherwise, call the tracking module based on HOG+Multi-SVM, enter step 5 and perform.
Step 3: based on the detection of HOG+Multi-SVM;
The image obtained step 1 carries out the HOG feature extraction in each metric space, moving window is utilized to travel through all valid windows, and mate with the SVM model in all directions sub-range of training in advance, the result of record matching, and the call number of gesture at the position of picture frame and SVM model is recorded when the match is successful, then forward step 5 to and perform.
Step 4: based on the tracking of HOG+Multi-SVM;
The image that step 1 obtains is extracted HOG feature according to the metric space traversal order of setting, and utilizes ad hoc fashion to be fixed the movement of window, then, mate with the SVM model of a definite sequence with all directions sub-range of training in advance.The result of record matching, and the call number of gesture at the position of picture frame and SVM model is recorded when the match is successful.Enter step 5 to perform.
Step 5: judge set gesture whether on image, and make respective handling, specifically:
(1) when not navigating to set gesture in the picture, clearing process being carried out to the frame number variable x that success is located, thus ensures the detection sub-module of next invocation step 3;
(2) otherwise, the value of frame number variable x success located increases 1, and records the call number of the SVM model of set gesture position size in the picture and coupling.
The invention has the advantages that:
(1) under complex background and the unstable situation of light condition, the method that the present invention proposes can work well, namely has certain robustness and robustness.
(2) shown by great many of experiments, average tracking locating speed can be reduced to 15 milliseconds/frame by the method that the present invention proposes, and can meet real-time requirement;
(3) show by experiment, the method that the present invention proposes also improves the precision of locating set gesture to a certain extent, ensure that the quality of tracking.
Accompanying drawing explanation
Fig. 1 is the overall flow schematic diagram of gesture tracking method of the present invention;
Fig. 2 is based on the schematic flow sheet that HOG+Multi-SVM detects in the present invention;
Fig. 3 is based on the schematic flow sheet that HOG+Multi-SVM follows the tracks of in the present invention;
Fig. 4 is the actual result utilizing the present invention to run in the present invention and the Comparative result not using tracking module.
Embodiment
Below in conjunction with drawings and Examples, the present invention is described in further detail.
The present invention have studied a kind of gesture tracking method of the many sense of rotation SVM model based on HOG feature, this method utilizes the picture frame before the continual crawl mirror of Web camera, then by carrying out whether comprising conventional sign in judgment frame based on the track and localization framework of HOG+Multi-SVM, find out the interval call number of its director and locate its position in frame.The present invention improves the many sense of rotation SVM model of utilization based on HOG feature effectively to the location of conventional sign.Multi-SVM refers to Multi-SVM (Multi.SupportVectorMachine, Multi-SVM), and the present invention refers to the SVM model in the multiple directions sub-range in plane space.
The method proposed in the present invention mainly comprises two processes in general: the initial detecting location of set gesture and the track and localization of postorder.The present invention utilizes and can obtain the former two field picture detection and location of gesture to conventional sign to from Web camera based on HOG+Multi-SVM detection module; And determining that the gesture tracking module started based on many sense of rotation SVM model of HOG feature carries out track and localization under these picture frames contain set gesture situation; Once tracking module shifts out the coverage of Web camera to set gesture tracking loss or gesture, then restart based on HOG+Multi-SVM detection module until successfully continuous detecting conventional sign can be navigated to.
The present invention detects the matching order of the metric space of detected image, the traversal mode of window and many sense of rotation model in framework in conjunction with the space-time characteristic nearby of gesture by research based on HOG+Multi-SVM, proposes a kind of novel gesture tracking method based on HOG and many SVM model.
The specific implementation step of the gesture tracking method of the many sense of rotation SVM model based on HOG feature that the present invention proposes, as Fig. 1, comprises following step:
Step 1: obtain picture frame, and carry out pre-service;
Utilize camera to capture camera lens each picture frame front in real time, and the image that bilinear interpolation obtains suitable dimension is carried out to image.The large I of suitable dimension is set by the user.Load the SVM model in dictionary corresponding to the correlation parameter of HOG feature and gradient throwing value space and all directions sub-range based on HOG feature.
Wherein, the correlation parameter that HOG feature is main comprises:
(1) size (Window Size) of SVM model inspection window;
(2) gradient throwing value space partition zone number (Bins);
(3) size (Block Size) of block in window;
(4) size (Cell Size) of cell factory in block.
It should be noted that in (2), Bins needs to be set as even number.
The all directions sub-range loading the SVM model of FIVE gesture is:
(1)Mod1:[-3,33]； (2)Mod2:[27,63]； (3)Mod3:[57,93]；
(4)Mod4:[87,123]； (5)Mod5:[117,153]； (6)Mod6:[147,183]。
The Linear SVM model in the multiple directions sub-range in plane space is that training in advance obtains, and has carried out index number to SVM model.
Step 2: judge whether successfully locate frame number arrives predetermined value;
Set a variable x, successfully detect or trace into the frame number containing gesture continuously before being recorded to present frame, and at the beginning of the inventive method program is run, be 0 by this initialization of variable.
The value of more successful locating frame number variable x and default frame number threshold value T, be performed as follows:
(1) if successful locating frame number variable x does not reach predetermined threshold value T, so, call the detection module based on HOG+Multi-SVM, enter step 3 and perform;
(2) otherwise, if successful locating frame number variable x reaches predetermined threshold value T, call the tracking module based on HOG+Multi-SVM, enter step 5 and perform.
Step 3: based on the detection of HOG+Multi-SVM;
Image after step 1 processes is carried out to the HOG feature extraction in each metric space, utilize moving window to travel through all valid windows, and mate with the Linear SVM model in all directions sub-range of training in advance.The result of record matching, and the call number of gesture at the position of picture frame and Linear SVM model is recorded when the match is successful.Forward step 5 to perform.
Based on the concrete steps of the check processing of HOG+Multi-SVM, as shown in Figure 2:
Step 3.1: carry out scaling to input picture, the corresponding RGB image of each metric space extracts HOG feature;
Step 3.2: travel through the corresponding HOG feature of each moving window under each metric space, utilize Linear SVM to mate with the SVM model in all directions scope HOG feature; And the every terms of information of this moving window is retained when the match is successful, comprise metric space, SVM model call number, the top left co-ordinate point of window and the director interval of SVM model;
Step 3.3: repeat step 3.2 until all identification dimensions spaces have all traveled through;
Step 3.4: jump to step 5.
Step 4: based on the tracking of HOG+Multi-SVM;
The metric space traversal mode of image after step 1 processes according to setting is extracted HOG feature, and utilize ad hoc fashion to be fixed the movement of window, then, mate with the Linear SVM model of a definite sequence with all directions sub-range of training in advance.The result of record matching, and the call number of gesture at the position of picture frame and Linear SVM model is recorded when the match is successful.Forward step 5 to perform.
Based on the concrete steps of the tracking process of HOG+Multi-SVM, as shown in Figure 3, performing step is:
Step 4.1: guarantee to load relevant dictionary, comprises following three parts:
The posterior probability table that the SVM model of the director subregion of a, all set gestures occurs mutually;
B, all possible graphical rule size and corresponding scaling value;
The window number that horizontal and vertical direction corresponding under each yardstick of c, image may hold respectively.
Step 4.2: according to set gesture place metric space in former frame, generates the yardstick traversal order of present frame.Now, if set gesture place metric space is not the scaling value of regulation in former frame, then the next stage yardstick of this metric space is got.
If set gesture place metric space is K in former frame, then the yardstick traversal order of present frame sorts according to the principle closest to metric space K.According to large young pathbreaker's metric space order sequence of scaling value, the distance of two metric spaces is exactly the distance of sequence in sequence number.
Step 4.3: utilize bilinear interpolation to scale the images to appointment size.
Step 4.4: to the image zooming-out HOG feature after convergent-divergent.
Step 4.5: according to the window top left co-ordinate of set gesture contained in former frame, generate a screw type window traversal order.Wherein, if top left co-ordinate can not fall in place, then need utilization to round after method adjusts this coordinate and carry out this step again.
Step 4.6: director is interval belonging to the SVM model of set gesture contained in former frame, utilizes posterior probability table to search the probability of this sub-range, direction dependence former frame positioning result, and by probable value descending sort SVM Model Matching order.
Step 4.7: carry out the traversal of window and the coupling of SVM model based on step 4.5 and step 4.6.Once the match is successful, carry out the every terms of information of record window, comprise metric space, SVM model call number, the top left co-ordinate point of window and the director interval of SVM model.If the match is successful and the window number threshold value of arrival regulation, so just terminate tracing process, jump to step 5.If mate unsuccessful under current scale space or do not arrive the window number threshold value of regulation, so carry out the traversal of next metric space according to metric space traversal order, go to step 4.2 execution.
Step 5: judge set gesture whether on image, and make respective handling;
The result of the detection and tracking module that step 3 and step 4 are called is differentiated, and makes following action:
(1) when not navigating to set gesture in the picture, so, clearing process is carried out to the frame number variable x that success is located, thus ensures the detection module of the HOG+Multi-SVM of next invocation step 3;
(2) otherwise, the frame number variable x that success is located is carried out from increasing 1, and records the SVM model call number in the sub-range, sense of rotation place of set gesture position size in the picture and coupling.
In the present invention, by carrying out the sense of rotation child partition of detection FIVE gesture to the video file of 500 FIVE gestures and recording its result in a local file.Then, the occurrence law passing through director subregion belonging to the gesture to each frame appearance in these 500 destination files is added up, and generates its posteriority probability statistics table, as shown in table 1:
Table 1: the posterior probability table of sense of rotation child partition SVM model
| Mod1 | Mod2 | Mod3 | Mod4 | Mod5 | Mod6 | |
| Mod1 | 0.12 | 0.012 | 0.034 | 0.021 | 0.02 | 0.01 |
| Mod2 | 0.005 | 0.058 | 0.003 | 0.0001 | 0.0001 | 0.0001 |
| Mod3 | 0.001 | 0.004 | 0.236 | 0.015 | 0.001 | 0.001 |
| Mod4 | 0.001 | 0.0001 | 0.015 | 0.183 | 0.009 | 0.0001 |
| Mod5 | 0.0002 | 0.0001 | 0.001 | 0.008 | 0.167 | 0.004 |
| Mod6 | 0.0001 | 0.0001 | 0.001 | 0.0002 | 0.003 | 0.075 |
First, 100 test video sequence containing set gesture are gathered, about each video sequence comprises 5000 frames.On this data set, carry out the gesture tracking method of the present invention's proposition and the contrast based on HOG+Multi-SVM detection method, as shown in table 2.
Table 2: the performance comparison table of two kinds of methods
| AMR(％) | AFPR(％) | ACLE(px) | AOR | ATPF(ms) | |
| Detection method | 2.3 | 0.9 | 9.8 | 0.86 | 62.13 |
| Tracking | 0.02 | 0.01 | 9.1 | 0.91 | 12.3 |
Wherein, meaning of parameters is as follows:
(1) AMR, average Loss Rate, has showed the situation that set gesture is lost in picture frame;
(2) AFPR, average false drop rate, have expressed the situation other situations being matched to mistakenly set gesture;
(3) ACLE, window center point average error, indicates the positioning error of set gesture;
(4) AOR, average window coverage rate, indicates the situation of place yardstick and precision;
(5) ATPF, every frame average positioning time, shows the speed of location.
From in upper table, be not difficult to obtain, as drawn a conclusion:
(1) tracking of the present invention's proposition is than fast nearly 5 times of simple detection method in speed, namely can reach 15ms/frame.This method that the present invention is proposed can meet the required real-time of gesture tracking.
(2) tracking that proposes of the present invention, compared with simple detection method, no matter in the accuracy rate differentiated, or all increases, wherein the accuracy rate of differentiation is reduced a magnitude in the precision of location.
Fig. 4 be the method that utilizes the present invention to propose with based on the detection method of HOG+Multi-SVM and the actual motion design sketch of tracking TLD (Tracking-Learning-Detection).Wherein, data set is one that extracts from above-mentioned data centralization.This data set considers illumination condition, fast movement, and the plane internal rotation of FIVE gesture turns and gesture such as to shift out at the situation within the scope of camera picked-up.Represent three kinds of methods with three kinds of lines frames respectively in figure, in order to represent removing in each subgraph, three kinds of lines frames are white.The subgraph of Fig. 4 shows respectively:
(1) in subgraph (a), three kinds of localization methods all produce the situation of very well results.Wherein, large two results being respectively method that the present invention proposes and detection method based on HOG+Multi-SVM.Middle little frame is the result that TLD follows the tracks of.
(2) subgraph (b) is the situation investigating three kinds of localization methods when illumination condition is affected.Wherein, the tracking based on TLD follows the tracks of loss; Another two kinds of methods obtain similar situation.
(3) in subgraph (c), the detection method based on HOG+Multi-SVM is located unsuccessfully, as schemed shown in medium and small rectangle frame.The method location that the present invention proposes is more accurate, as shown in large rectangle frame in figure.And follow the tracks of loss based on the tracking of TLD.
(4) subgraph (d) illustrates the situation that set gesture plane internal rotation turns.Wherein, the tracking based on TLD is followed the tracks of and is lost, and the detection method based on HOG+Multi-SVM is similar as shown in rectangle frame in figure with the performance performance that the method that the present invention proposes obtains.
Claims (4)
1., based on a gesture tracking method for many sense of rotation SVM model of HOG feature, it is characterized in that, comprise the following steps:
Step 1: obtain picture frame, and carry out pre-service, specifically:
(1) utilize camera to capture camera lens each picture frame front in real time, bilinear interpolation process is carried out to obtain the image of setting size to the image of present frame;
(2) the SVM model in dictionary corresponding to the correlation parameter of HOG feature, gradient throwing value space and all directions sub-range based on HOG feature is loaded;
Step 2: judge whether successfully locate frame number arrives default frame number threshold value T;
Set a variable x, successfully detect or trace into the frame number containing gesture continuously before being recorded to present frame; Initial x value is 0;
If x does not reach frame number threshold value T, enter step 3 and perform; Otherwise, enter step 5 and perform;
Step 3: based on the detection of HOG+Multi-SVM;
The image of step 1 acquisition is carried out to the HOG feature extraction in each metric space, moving window is utilized to travel through all valid windows, and mate with the SVM model in all directions sub-range of training in advance, the result of record matching, and the call number of gesture at the position of picture frame and SVM model is recorded when the match is successful, then forward step 5 to;
Step 4: based on the tracking of HOG+Multi-SVM;
According to the metric space traversal order of setting, HOG feature is extracted to the image that step 1 obtains, ad hoc fashion is utilized to be fixed the movement of window, and mate with the SVM model of a definite sequence with all directions sub-range of training in advance, the result of record matching, and the call number of gesture at the position of picture frame and SVM model is recorded when the match is successful, then enter step 5 and perform;
Step 5: judge set gesture whether on image, and make respective handling, specifically:
(1) when not navigating to set gesture in the picture, clearing process is carried out to variable x;
(2) otherwise, the value of variable x is increased 1, and records the call number of the SVM model of set gesture position size in the picture and coupling.
2. the gesture tracking method of a kind of many sense of rotation SVM model based on HOG feature according to claim 1, it is characterized in that, the correlation parameter of the HOG feature described in step 1 comprises:
(1) size of SVM model inspection window;
(2) gradient throwing value space partition zone number, requires to be set as even number;
(3) size of block in window;
(4) size of cell factory in block;
The SVM model loaded is divided into six direction sub-range:
(1)[-3,33]； (2)[27,63]； (3)[57,93]；
(4)[87,123]； (5)[117,153]； (6)[147,183]。
3. the gesture tracking method of a kind of many sense of rotation SVM model based on HOG feature according to claim 1, it is characterized in that, the specific implementation step of described step 3 is:
Step 3.1: carry out scaling to input picture, the corresponding RGB image of each metric space extracts HOG feature;
Step 3.2: travel through the corresponding HOG feature of each moving window under each metric space, HOG feature is mated with the SVM model in all directions scope; And the every terms of information of this moving window is retained when the match is successful, comprise metric space, SVM model call number, the top left co-ordinate point of window and the director interval of SVM model;
Step 3.3: repeat step 3.2 until all identification dimensions spaces have all traveled through;
Step 3.4: jump to step 5 and perform.
4. the gesture tracking method of a kind of many sense of rotation SVM model based on HOG feature according to claim 1, it is characterized in that, the specific implementation step of described step 4 is:
Step 4.1: guarantee to load relevant dictionary, comprises following three parts:
The posterior probability table that the SVM model of the director subregion of a, all set gestures occurs mutually;
B, all graphical rule sizes and corresponding scaling value;
The window number that horizontal and vertical direction corresponding under each yardstick of c, image can hold respectively;
Step 4.2: according to set gesture place metric space in former frame, generates the yardstick traversal order of present frame; If former frame gesture place metric space is not the scaling value of regulation, then get the next stage metric space of this metric space; The yardstick traversal order of present frame sorts according to closest to the principle of set gesture place metric space in former frame;
Step 4.3: utilize bilinear interpolation to scale the images to appointment size;
Step 4.4: to the image zooming-out HOG feature after convergent-divergent;
Step 4.5: according to the window top left co-ordinate of set gesture contained in former frame, generate a screw type window traversal order;
Step 4.6: director is interval belonging to the SVM model of set gesture contained in former frame, utilizes posterior probability table to search the probability of this sub-range, direction dependence former frame positioning result, and by probable value descending sort SVM Model Matching order;
Step 4.7: carry out the traversal of window and the coupling of SVM model based on step 4.5 and step 4.6; Once the match is successful, the every terms of information of record window, comprises metric space, SVM model call number, the top left co-ordinate point of window and the director interval of SVM model; If mate unsuccessful and arrive the window number threshold value specified, so just terminate tracing process, jump to step 5, otherwise carry out the traversal of next metric space, go to step 4.2 execution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510079498.XA CN104731323B (en) | 2015-02-13 | 2015-02-13 | A kind of gesture tracking method of many direction of rotation SVM models based on HOG features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510079498.XA CN104731323B (en) | 2015-02-13 | 2015-02-13 | A kind of gesture tracking method of many direction of rotation SVM models based on HOG features |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104731323A true CN104731323A (en) | 2015-06-24 |
| CN104731323B CN104731323B (en) | 2017-07-04 |
Family
ID=53455291
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510079498.XA Active CN104731323B (en) | 2015-02-13 | 2015-02-13 | A kind of gesture tracking method of many direction of rotation SVM models based on HOG features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104731323B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106920251A (en) * | 2016-06-23 | 2017-07-04 | 阿里巴巴集团控股有限公司 | Staff detecting and tracking method and device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101763515A (en) * | 2009-09-23 | 2010-06-30 | 中国科学院自动化研究所 | Real-time gesture interaction method based on computer vision |
| CN102831439A (en) * | 2012-08-15 | 2012-12-19 | 深圳先进技术研究院 | Gesture tracking method and gesture tracking system |
| CN103530613A (en) * | 2013-10-15 | 2014-01-22 | 无锡易视腾科技有限公司 | Target person hand gesture interaction method based on monocular video sequence |
-
2015
- 2015-02-13 CN CN201510079498.XA patent/CN104731323B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101763515A (en) * | 2009-09-23 | 2010-06-30 | 中国科学院自动化研究所 | Real-time gesture interaction method based on computer vision |
| CN102831439A (en) * | 2012-08-15 | 2012-12-19 | 深圳先进技术研究院 | Gesture tracking method and gesture tracking system |
| CN103530613A (en) * | 2013-10-15 | 2014-01-22 | 无锡易视腾科技有限公司 | Target person hand gesture interaction method based on monocular video sequence |
Non-Patent Citations (4)
| Title |
|---|
| D.F.LLORCA ET AL.: "《A multi-class SVM classifier ensemble for automatic hand washing quality assessment》", 《HTTP://WWW.DCS.WARWICK.AC.UK/BMVC2007/PROCEEDINGS/VOLUMES/223/HAND_WASHING_BMCV07.PDF》 * |
| FENG HAN ET AL.: "《A Two-Stage Approach to People and Vehicle Detection with HOG-Based SVM》", 《HTTPS://PDFS.SEMANTICSCHOLAR.ORG/1C76/6D0F4BF8FF443CBE8A487313E77C20ED4166.PDF》 * |
| KAZUTAKA SHIMADA ET AL.: "《A Combined Method Based on SVM and Online Learning with HOG for Hand Shape Recognition》", 《HTTP://AI2-S2-PDFS.S3.AMAZONAWS.COM/877B/85C13208FE65502015DEEE4DD293F2250542.PDF》 * |
| 陈小柏: "《基于视觉的连续手语识别系统的研究》", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106920251A (en) * | 2016-06-23 | 2017-07-04 | 阿里巴巴集团控股有限公司 | Staff detecting and tracking method and device |
| WO2017219875A1 (en) * | 2016-06-23 | 2017-12-28 | 阿里巴巴集团控股有限公司 | Hand detecting and tracking method and device |
| JP2019519049A (en) * | 2016-06-23 | 2019-07-04 | アリババ グループ ホウルディング リミテッド | Hand detection and tracking method and apparatus |
| TWI703507B (en) * | 2016-06-23 | 2020-09-01 | 香港商阿里巴巴集團服務有限公司 | Human hand detection and tracking method and device |
| US10885639B2 (en) | 2016-06-23 | 2021-01-05 | Advanced New Technologies Co., Ltd. | Hand detection and tracking method and device |
| US10885638B2 (en) | 2016-06-23 | 2021-01-05 | Advanced New Technologies Co., Ltd. | Hand detection and tracking method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104731323B (en) | 2017-07-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110458059B (en) | Gesture recognition method and device based on computer vision | |
| CN109344793B (en) | Method, apparatus, device and computer readable storage medium for recognizing handwriting in the air | |
| CN111488791A (en) | On-device classification of fingertip movement patterns as gestures in real time | |
| CN103353935A (en) | 3D dynamic gesture identification method for intelligent home system | |
| CN103593680A (en) | Dynamic hand gesture recognition method based on self incremental learning of hidden Markov model | |
| Huang et al. | Deepfinger: A cascade convolutional neuron network approach to finger key point detection in egocentric vision with mobile camera | |
| Premaratne et al. | Centroid tracking based dynamic hand gesture recognition using discrete Hidden Markov Models | |
| CN111950514B (en) | Depth camera-based aerial handwriting recognition system and method | |
| JP2013105305A (en) | Gesture recognition apparatus, method therefor, and program therefor | |
| Ruan et al. | Dynamic gesture recognition based on improved DTW algorithm | |
| Ding et al. | Recognition of hand-gestures using improved local binary pattern | |
| CN103336967A (en) | Hand motion trail detection method and apparatus | |
| CN105912126A (en) | Method for adaptively adjusting gain, mapped to interface, of gesture movement | |
| CN111105443A (en) | Video group figure motion trajectory tracking method based on feature association | |
| KR101899590B1 (en) | Method and Apparatus for Recognizing Hand Shapes and Gestures | |
| Al-Akam et al. | Local feature extraction from RGB and depth videos for human action recognition | |
| Beg et al. | Text writing in the air | |
| CN108614988A (en) | A kind of motion gesture automatic recognition system under complex background | |
| Liang et al. | Egocentric hand pose estimation and distance recovery in a single RGB image | |
| Lekova et al. | Fingers and gesture recognition with kinect v2 sensor | |
| Hsieh et al. | A real time hand gesture recognition system based on DFT and SVM | |
| CN104731324B (en) | A kind of gesture inner plane rotation detection model generation method based on HOG+SVM frameworks | |
| CN104731323A (en) | Multi-rotating direction SVM model gesture tracking method based on HOG characteristics | |
| Li et al. | A dynamic hand gesture recognition model based on the improved dynamic time warping algorithm | |
| Rubin Bose et al. | In-situ identification and recognition of multi-hand gestures using optimized deep residual network |
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 |