CN109271840A - A kind of video gesture classification method - Google Patents
A kind of video gesture classification method Download PDFInfo
- Publication number
- CN109271840A CN109271840A CN201810826221.2A CN201810826221A CN109271840A CN 109271840 A CN109271840 A CN 109271840A CN 201810826221 A CN201810826221 A CN 201810826221A CN 109271840 A CN109271840 A CN 109271840A
- Authority
- CN
- China
- Prior art keywords
- video
- visible light
- conspicuousness
- feature
- fusion
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
- G06V40/28—Recognition of hand or arm movements, e.g. recognition of deaf sign language
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of video gesture classification methods.Related method includes: that the visible light video and deep video of same target collection are normalized；The conspicuousness for calculating each frame image in visible light video, obtains conspicuousness video；Visible light video, deep video, conspicuousness video after being utilized respectively normalization are trained Three dimensional convolution neural network, and the Three dimensional convolution neural network includes: the input layer set gradually, 18 convolutional layers, global average pond layer and full articulamentum；The be averaged feature of pond layer of the overall situation of the airspace dimension of obtained visible light video, deep video and conspicuousness video is carried out by Fusion Features obtains fusion feature, video gesture classification is carried out according to obtained fusion feature.The targeted problem of the present invention is the extensive gesture identification based on video, and categorical data and the video data volume are all much larger than existing invention.The present invention can identify in real time video, more confidence level and practicability.
Description
Technical field
The invention belongs to deep learning Applied research fields, further relate to deep based on three-dimensional in convolutional neural networks
The dynamic gesture classification method for spending neural network can be applied to human-computer interaction, intelligent driving, intelligence wearing, Entertainment etc. and relate to
And to the place of gesture identification.
Background technique
Gesture is substantially also a kind of language system.Gesture can be divided into static gesture and dynamic gesture, and static gesture is main
Pay close attention to the factors such as shape, profile, the center of gravity of gesture；The dynamic gesture record motion profile of hand, hand brandish the main points such as orientation.Phase
The Gesture Recognition answered is also classified into static identification technology and dynamic recognition technique, and static gesture refers mainly to the side of human body gesture
To, the shape of gesture and texture, the meaning for being applicable to expression is less, the simple scene of identification；Dynamic hand gesture recognition is main
It is motion process and the track of tracking and identification gesture, suitable for the scene of variation and the image of irregular movement, this gesture
Identification can express meaning more accurate and abundant.
Initial Research of Gesture Recognition is based on a kind of dedicated data glove come what is carried out, and computer passes through data
The sensor of gloves obtains the various information of hand.Luo Ying, Cao Jinling et al. are in a kind of patent " intelligence based on data glove
In hand language recognition device ", the movement variation and amplitude in each joint of hand are perceived by the sensor in data glove, to know
The sign language meaning of other deaf-mute, so that deaf-mute be helped to carry out daily exchange with normal person.Zhou Yanming, Huang Changzheng et al. are in patent
Microcontroller, data acquisition module, bluetooth in " a kind of identification device of quickly small gesture " by being arranged on clamping device
Wireless communication module and haptic feedback module realize the quick identification to minor motion, and by haptic feedback module according to
Different scenes Real-time Feedback corresponding actions.But the requirement based on the gesture identification of data glove to hardware is relatively high, data hand
Covering the features such as at high cost, application method is complicated makes it be not easy to promote.Lu Tong, Hu Wei et al. patent " based on finger contours and
In the static gesture identification method of decision tree ", the profile for obtaining hand is split to depth image, the distance calculated between finger is made
For feature vector, feature vector is trained and is classified using decision tree, to obtain the meaning of gesture.But this method is
It is identified based on two-dimensional static gesture, it is impossible to be used in true 3 D stereo scene.Wang Peng in document, " know by Wang Peng ' dynamic gesture
Not Yan Jiu ', Dalian University of Technology, propose in 2013 " a kind of online vision with adaptivity and recapture ability with
Track algorithm, the tracker of algorithm is based partially on particle filter tracking frame, using a kind of supporting vector of online study
Machine is as observing and nursing；Detector portion has chosen descriptor histograms of oriented gradients (HOG) feature as sample characteristics.The party
Method tentatively realizes dynamic hand gesture recognition, but categories of datasets number used by testing is less, only 10 classes, and is specific hand
Gesture.
In conclusion although existing Gesture Recognition achieves certain achievement, but mostly there is certain limitation
Property, due to ignoring the complementarity of different modalities data, to the extensive identification based on video, that there are still accuracy of identification is low, vulnerable to back
The problems such as scape interferes also requires further study in terms of the dynamic hand gesture recognition in real scene based on space-time characteristic.
Summary of the invention
The present invention be directed to limitation of the existing Gesture Recognition in true stereo scene, especially dynamic gestures to know
The other technology problem lower to the discrimination of extensive classification problem, provides a kind of video gesture classification method.
Video gesture classification method of the invention includes:
Step 1, the visible light video and deep video of same target collection are normalized, it is visible after normalization
Light video and deep video have identical height, same widths and equal length；
Step 2, the conspicuousness for calculating each frame image in visible light video, obtains conspicuousness video；
Step 3, the visible light video, deep video, conspicuousness video after being utilized respectively normalization are to Three dimensional convolution nerve
Network is trained, and the Three dimensional convolution neural network includes: the input layer set gradually, 18 convolutional layers, global average pond
Change layer and full articulamentum；
Step 4, it is averaged feature, the overall situation of deep video of pond layer to the overall situation of visible light video obtained in step 3
The be averaged feature progress Fusion Features of pond layer of the overall situation of the airspace dimension of the feature and conspicuousness video of average pond layer obtain
Fusion feature carries out video gesture classification according to obtained fusion feature.
Further, normalized of the present invention includes airspace normalized and time domain normalized, the sky
Domain normalized includes so that visible light video and deep video is had identical height and identical width by scaling or cutting processing
Degree；The time domain normalized includes adaptively choosing the mode work of all video frame numbers after the normalized of airspace
To normalize standard, reduction frame number is carried out to the video beyond normalization standard；Completion is carried out for substandard video, is made
All video frame numbers having the same.
Preferably, Fusion Features are merged using mean value in step 4 of the present invention, splicing is merged or Canonical correlation
(CCA) analysis fusion.
Preferably, fusion feature is inputted in support vector machines in step 4 of the present invention and is classified.
Compared with prior art, the invention has the following advantages that
First, what the present invention was studied is three-dimensional dynamic hand gesture recognition, rather than is identified based on two-dimensional static gesture,
And the special hardwares such as data glove are needed not rely on, therefore closer to real-life real scene, are conducive to application.
Second, the targeted problem of the present invention is the extensive gesture identification based on video, categorical data and video counts
Existing invention is all much larger than according to amount.The method that the present invention uses can identify in real time video, more confidence level and practical
Property.
Third, the present invention have carried out space-time normalized to data first before carrying out gesture identification, have made new sample
This has a high consistency in time and Spatial Dimension, and can be reasonable, reservation gesture video as much as possible
Feature.
4th, the present invention removes the interference of gesture irrelevant factor using Saliency maps picture.In visible light video, back
Color of scape, the colour of skin and performing artist's clothes etc. is all the factor unrelated with gesture meaning, these factors are in the side based on study
The identification of gesture classification can be interfered in method, and the conspicuousness video masking that the present invention uses these irrelevant factors, make
Network can be absorbed in the movement of gesture itself.
5th, it is of the present invention preferably to be learnt based on canonical correlation analysis fusion (CCA) method to same
Feature association degree of one sample under different modalities data, and corresponding weight is designed accordingly, existed by the method for Weighted Fusion
The complementarity for utilizing different modalities data to the full extent obtains the fusion feature that can more comprehensively, more reflect gesture feature.
Detailed description of the invention
Fig. 1 is flow chart of the invention.
Fig. 2 is the network structure of Three dimensional convolution neural network.
Fig. 3 is the effect picture of Saliency maps picture；(a) figure is RGB original image, and (b) figure is conspicuousness effect picture.
Fig. 4 is characterized extraction simulation result diagram；(a) figure is three kinds of feature accuracys rate and loss function line chart, and (b) figure is
Single features recognition accuracy.
Fig. 5 is characterized fusion simulation result diagram.
Specific embodiment
The visible light video and deep video of same target collection of the invention refer to same target recorded obtained by regard
The visible light video and deep video of frequency, wherein visible light video and deep video are that the conventional of this field understands.It can also lead to
Cross visible light video and deep video that the video data for obtaining and being stored in hard disc of computer space obtains same target collection, example
The visible light video (rgb video) and deep video (depth video) such as got by Kinect sensor.
The purpose of normalized of the invention is to make video height having the same, width and length, it can be achieved that the technology
The data normalization processing method of purpose is suitable for the present invention.
Conspicuousness video acquisition of the invention can be obtained according to conspicuousness theory, i.e. each frame in calculating visible light video
Then each frame Saliency maps picture is connected, obtains conspicuousness video by the significance value of image.
Three-dimensional nerve network of the invention can handle the information of time and Spatial Dimension, obtain the space-time characteristic of gesture.
A kind of specific three-dimensional nerve network is: the number of each layer of convolution kernel is respectively 64,64,64,64,64,128,128,128,
128,256,256,256,256,512,512,512,512, a convolutional layer is then in addition added, the use of 1024 sizes is 1*
The convolution kernel of 1*1, by Feature Mapping to higher dimension.This specific network structure is as shown in Fig. 2, include 18
Convolutional layer, 1 overall situation are averaged pond layer and 1 full articulamentum.It is 3*3*3's that size, which can be used, in most of convolutional layer in network
Convolution kernel carries out convolution operation.In the network structure, traditional pond layer is replaced by step-length by 2 convolutional layer, is come with this
Achieve the effect that down-sampling.Then, after 8 residual units and an additional convolutional layer, it is connected to an airspace dimension
The overall situation of degree is averaged pond layer, and the size of pond window is 7*7.Finally, a full articulamentum is connected to, for what will be inputted
Classification is mapped to 249 classes.
Embodiment:
Embodiment processing carries out on Chalearn LAP IsoGD data set.The data set is that Wan et al. exists
Paper " Chalearn looking at people rgb-d isolated and continuous on CVPRW2016
It is proposed in datasets for gesture recognition ", which shares 47933 gesture videos, each video
One gesture motion is made by a volunteer, which contains 249 class gestures altogether.Wherein 35878 videos are as training set,
5784 videos are as verifying collection, and 6271 videos are as test set.
Referring to Fig.1, the video gesture classification method of the embodiment includes:
Step 1, the reading of video data
Using matlab software reading video data, wherein the data include being got from sensors such as Kinect
Light video (rgb video) and deep video (depth video).
Step 2, the space-time normalization of video data
2a) the airspace normalization of video data, the present invention is by scale scaling and cuts out two ways, in step 1
Video is normalized on Spatial Dimension, with identical height and width.
Steps are as follows for a kind of specific airspace normalized:
Scale reduction is carried out to each frame of video data, makes the video of 320*240 (wide * high)；
Each frame of video data is cut, the video of 112*112 is obtained, wherein cut video top left co-ordinate by
Selected coordinate (i, j) determines (i≤208, j≤128) at random；
2b) the time domain normalization of video data, the embodiment are adaptive according to the statistics to all inputting video data frame numbers
Choose 2a with answering) in all video frame numbers progress of the mode (value S) as normalization standard, to the video beyond standard
Sampling carries out completion to substandard video, with identical frame number.
Steps are as follows for a kind of specific time domain normalized:
The video V for exceeding standard for frame number, carries out equally spaced uniform sampling, retains in original video every m frame
One frame deletes remaining m-1 frame.It can be obtained the video of standard frame number, m is obtained by following formula:
M=total/s (1)
It is wherein standard frame number is s, total is the practical frame number of original video.
Video substandard for frame number calculates the ratio between standard frame number and practical frame number:
Then, it for each frame in video, is replicated ratio times, interpolation is after the position of the frame.View at this time
Frequency frame number and the difference of standard frame number are
Diff=s-total*ratio (3)
If diff > 0, it need to select at random diff frame in original total frame and once be replicated, is placed sequentially in
After the position of random frame.So far, the video that frame number is less than standard is completed to expand frame/completion operation.
Step 3 generates conspicuousness video
Conspicuousness video is generated according to visible light video obtained in step 2, the specific steps of which are as follows:
[1] gaussian filtering is carried out to image, obtains filtered image, takes the Lab value of its pixel
WhereinThe pixel value of image after indicating gaussian filtering,Indicate its luminance component of brightness, value range is
[0,100],Indicate that, from green to red range, value range is [- 128,127],It indicates from blue to yellow
Range, value range are [- 128,127].
[2] image is calculated in the mean value in the space LAB
IμIt is mean value pixel value, LμIndicate brightness, aμIt indicates from green to red range, bμIt indicates from blue to yellow
Range.
[3] Euclidean distance is sought, Saliency maps are obtained
Step 4, the training of Three dimensional convolution neural network model
Conspicuousness video obtained in visible light video, deep video obtained in step 2 and step 3 is sent into Fig. 2
Training, iteration 120000 times, obtains corresponding network model in represented Three dimensional convolution neural network model.
Step 5, feature extraction
Respectively using trained model in step 4, feature is carried out to visible light, depth, conspicuousness gesture video and is mentioned
It takes.Specifically, its feature at pool6 layers is extracted.
Step 6, Fusion Features
Q.-S.Sun, S.-G.Zeng, Y.Liu, the paper .A of P.-A.Heng, D.-S.Xia et al. are used in the embodiment
new method of feature fusion and its application in image recognition.Pattern
Method disclosed in Recognition seeks different modalities obtained in step 5 by canonical correlation analysis method (CCA)
The weight of data, realizes its Weighted Fusion, concrete mode be first with the Canonical correlation method to Visible Light Characteristics and
Depth characteristic is merged, and intermediate fusion feature vector is obtained, and recycles intermediate fusion of the Canonical correlation method to obtaining
Feature is merged with significant characteristics vector, obtains finally merging vector.
Step 7, classification
The present invention is classified using SVM classifier.In the specific embodiment, the SVM in the tool box matlab is used
Classifier carries out Classification and Identification to gesture video.Since SVM is accomplished that two classification, and general gesture type can all be greater than 2,
Therefore, the present invention is by trained and consistent two classifier of gesture number of species, to realize more classification.Specifically, by step 6
Obtained in the fusion feature new feature of training data be input in SVM, SVM is trained, then melting test data
It closes feature and is input in trained SVM and classify, obtain classification results.Wherein according to Chalearn LAP IsoGD data
Collect the training data divided and test data and above-mentioned deep neural network is consistent division.
The simulated effect of embodiment:
Emulation 1, the emulation that the saliency data in above-mentioned steps two and step 3 is obtained.
Emulating 1 simulated conditions is carried out under MATLAB R2016a software.
The emulation experiment of saliency data acquisition is carried out referring to Fig. 3.Fig. 3 (a) is visible data sample, and Fig. 3 (b) is it
Corresponding conspicuousness sample.As seen from Figure 3, saliency data has gesture irrelevant contents in elimination visible data dry
The characteristic and visible data for disturbing, reflecting gesture region of variation have good complementarity.
Emulation 2, the emulation to the feature extraction in above-mentioned steps four and step 5.
Emulating 2 simulated conditions is lower under deep learning frame caffe environment carry out.
Feature extraction emulation experiment is carried out referring to Fig. 4.Fig. 4 (a) is the loss decline curve figure of three kinds of modal datas, the song
Line reflects the case where network model fitting data feature, and penalty values are lower to illustrate that fitting degree is higher；Fig. 4 (b) is different moulds
The respective recognition result of state data.
From Fig. 4 (a) and 4 (b) as can be seen that the feature extracting method that the present invention uses can be fitted different modalities well
Data, after enough iteration, penalty values are lower, while the identification on visible light, depth and conspicuousness modal data
Rate also respectively reaches 45.07%, 48.44% and 43.21%.
Emulation 3, the emulation to six Fusion Features of above-mentioned steps.
Emulating 3 simulated conditions is carried out under MATLAB R2016a software.
Emulation experiment is carried out referring to Fig. 5.Fig. 5 is using the fusion method for directly concatenating feature from beginning to end and using based on allusion quotation
The fusion results that type correlation analysis (CCA) obtains.
From fig. 5, it can be seen that the algorithm in the present invention achieves good syncretizing effect.It is merged using Canonical correlation special
The accuracy of sign has reached 61.98%, higher than the identification knot for the fusion method and single features for directly concatenating feature from beginning to end
Fruit.
Claims (4)
1. a kind of video gesture classification method, which is characterized in that method includes:
Step 1, the visible light video and deep video of same target collection are normalized, the visible light view after normalization
Frequency and deep video have identical height, same widths and equal length；
Step 2, the conspicuousness for calculating each frame image in visible light video, obtains conspicuousness video；
Step 3, the visible light video, deep video, conspicuousness video after being utilized respectively normalization are to Three dimensional convolution neural network
It is trained, the Three dimensional convolution neural network includes: the input layer set gradually, 18 convolutional layers, global average pond layer
With full articulamentum；
Step 4, average to the be averaged feature of pond layer, the overall situation of deep video of the overall situation of visible light video obtained in step 3
The overall situation of the airspace dimension of the feature and conspicuousness video of pond layer be averaged pond layer feature carry out Fusion Features merged
Feature carries out video gesture classification according to obtained fusion feature.
2. video gesture classification method as described in claim 1, which is characterized in that the normalized includes airspace normalizing
Change processing and time domain normalized, the airspace normalized include by scaling or cutting processing make visible light video and
Deep video has identical height and same widths；The time domain normalized includes adaptively choosing to normalize through airspace
The mode for all video frame numbers that treated carries out reduction frame number as normalization standard, to the video beyond normalization standard；
Completion is carried out for substandard video, makes all videos frame number having the same.
3. video gesture classification method as described in claim 1, which is characterized in that Fusion Features are using equal in the step 4
Value fusion, splicing fusion or canonical correlation analysis fusion.
4. video gesture classification method as described in claim 1, which is characterized in that input fusion feature in the step 4
Classify in support vector machines.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810826221.2A CN109271840A (en) | 2018-07-25 | 2018-07-25 | A kind of video gesture classification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810826221.2A CN109271840A (en) | 2018-07-25 | 2018-07-25 | A kind of video gesture classification method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109271840A true CN109271840A (en) | 2019-01-25 |
Family
ID=65148327
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810826221.2A Pending CN109271840A (en) | 2018-07-25 | 2018-07-25 | A kind of video gesture classification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109271840A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110309761A (en) * | 2019-06-26 | 2019-10-08 | 深圳市微纳集成电路与***应用研究院 | Continuity gesture identification method based on the Three dimensional convolution neural network with thresholding cycling element |
| CN110309726A (en) * | 2019-06-10 | 2019-10-08 | 济南大学 | A kind of micro- gesture identification method |
| CN110363093A (en) * | 2019-06-19 | 2019-10-22 | 深圳大学 | A kind of driver's action identification method and device |
| CN111476158A (en) * | 2020-04-07 | 2020-07-31 | 金陵科技学院 | Multi-channel physiological signal somatosensory gesture recognition method based on PSO-PCA-SVM |
| CN112818936A (en) * | 2021-03-02 | 2021-05-18 | 成都视海芯图微电子有限公司 | Rapid recognition and classification method and system for continuous gestures |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103455794A (en) * | 2013-08-23 | 2013-12-18 | 济南大学 | Dynamic gesture recognition method based on frame fusion technology |
| CN103903238A (en) * | 2014-03-21 | 2014-07-02 | 西安理工大学 | Method for fusing significant structure and relevant structure of characteristics of image |
| CN104134061A (en) * | 2014-08-15 | 2014-11-05 | 上海理工大学 | Number gesture recognition method for support vector machine based on feature fusion |
| CN106599789A (en) * | 2016-07-29 | 2017-04-26 | 北京市商汤科技开发有限公司 | Video class identification method and device, data processing device and electronic device |
-
2018
- 2018-07-25 CN CN201810826221.2A patent/CN109271840A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103455794A (en) * | 2013-08-23 | 2013-12-18 | 济南大学 | Dynamic gesture recognition method based on frame fusion technology |
| CN103903238A (en) * | 2014-03-21 | 2014-07-02 | 西安理工大学 | Method for fusing significant structure and relevant structure of characteristics of image |
| CN104134061A (en) * | 2014-08-15 | 2014-11-05 | 上海理工大学 | Number gesture recognition method for support vector machine based on feature fusion |
| CN106599789A (en) * | 2016-07-29 | 2017-04-26 | 北京市商汤科技开发有限公司 | Video class identification method and device, data processing device and electronic device |
Non-Patent Citations (3)
| Title |
|---|
| QIGUANG MIAO等: ""Multimodal Gesture Recognition Based on the ResC3D Network"", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION WORKSHOPS》 * |
| YUNAN LI等: ""Large-Scale Gesture Recognition With a Fusion of RGB-D Data Based on Saliency Theory and C3D Model"", 《IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY》 * |
| 杨红玲等: ""基于卷积神经网络的手势识别 "", 《计算机技术与发展》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110309726A (en) * | 2019-06-10 | 2019-10-08 | 济南大学 | A kind of micro- gesture identification method |
| CN110309726B (en) * | 2019-06-10 | 2022-09-13 | 济南大学 | Micro-gesture recognition method |
| CN110363093A (en) * | 2019-06-19 | 2019-10-22 | 深圳大学 | A kind of driver's action identification method and device |
| CN110309761A (en) * | 2019-06-26 | 2019-10-08 | 深圳市微纳集成电路与***应用研究院 | Continuity gesture identification method based on the Three dimensional convolution neural network with thresholding cycling element |
| CN111476158A (en) * | 2020-04-07 | 2020-07-31 | 金陵科技学院 | Multi-channel physiological signal somatosensory gesture recognition method based on PSO-PCA-SVM |
| CN112818936A (en) * | 2021-03-02 | 2021-05-18 | 成都视海芯图微电子有限公司 | Rapid recognition and classification method and system for continuous gestures |
| CN112818936B (en) * | 2021-03-02 | 2022-12-09 | 成都视海芯图微电子有限公司 | Rapid recognition and classification method and system for continuous gestures |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109271840A (en) | A kind of video gesture classification method | |
| CN107748873B (en) | A kind of multimodal method for tracking target merging background information | |
| CN110929578B (en) | Anti-shielding pedestrian detection method based on attention mechanism | |
| CA2934514C (en) | System and method for identifying faces in unconstrained media | |
| CN104834922B (en) | Gesture identification method based on hybrid neural networks | |
| CN104598915B (en) | A kind of gesture identification method and device | |
| CN103824089B (en) | Cascade regression-based face 3D pose recognition method | |
| CN103810503B (en) | Depth study based method for detecting salient regions in natural image | |
| CN102324025B (en) | Human face detection and tracking method based on Gaussian skin color model and feature analysis | |
| CN104966085B (en) | A kind of remote sensing images region of interest area detecting method based on the fusion of more notable features | |
| CN108038466B (en) | Multi-channel human eye closure recognition method based on convolutional neural network | |
| CN104835175B (en) | Object detection method in a kind of nuclear environment of view-based access control model attention mechanism | |
| CN108230338B (en) | Stereo image segmentation method based on convolutional neural network | |
| CN103020992B (en) | A kind of video image conspicuousness detection method based on motion color-associations | |
| Wang et al. | Multifocus image fusion using convolutional neural networks in the discrete wavelet transform domain | |
| CN103714181B (en) | A kind of hierarchical particular persons search method | |
| CN105739702A (en) | Multi-posture fingertip tracking method for natural man-machine interaction | |
| CN105550678A (en) | Human body motion feature extraction method based on global remarkable edge area | |
| CN106778474A (en) | 3D human body recognition methods and equipment | |
| CN108647625A (en) | A kind of expression recognition method and device | |
| CN108197534A (en) | A kind of head part's attitude detecting method, electronic equipment and storage medium | |
| CN113269089B (en) | Real-time gesture recognition method and system based on deep learning | |
| CN103618918A (en) | Method and device for controlling display of smart television | |
| CN104063871B (en) | The image sequence Scene Segmentation of wearable device | |
| CN109190559A (en) | A kind of gesture identification method, gesture identifying device and electronic equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190125 |