Font Size: a A A

The Gesture Motion Recognition Based On The Visual System

Posted on:2019-01-23Degree:MasterType:Thesis
Country:ChinaCandidate:Z Q BaoFull Text:PDF
GTID:2428330593951621Subject:Electromagnetic field and microwave technology
Abstract/Summary:PDF Full Text Request
The human-computer interaction based on visual system have played a more and more significant role in our daily life,and therefore it has been applied in the field of real-time sign language recognition,smart robot and visual reality.A vital part in human-computer interaction is gesture recognition,in which main process includes segmentation,tracking and recognition.For the real-time gesture recognition has a fairly high demand to recognition accuracy and time,which means the time of recognition should be programmed as short as possible while the high recognition accuracy should be remained.Meanwhile,the spatial and temporal variety of gesture and the influence of complicated application scenes,have made the gesture recognition become a popular and challenging research hotspot.In the hand gesture segmentation and tracking part,a gesture image extraction method which is on the basis of Kalman filter is proposed.The depth image and skeleton information are obtained from Kinect sensor firstly,then the tracking of precise skeleton points is realized by Kalman filter.With regard to the residual arm information divided by depth threshold,an arm-removing method with geometry algorithm is provided to effectively eliminate the interference of arm.Finally,the detection of contours is implemented by “Finding”,then the gesture area is extracted with maximum rectangular frame and the gesture images could be normalized.To verify the efficiency of segmentation,a static gesture dataset with 31 types of Chinese hand gesture language have been collected and constructed,and there are 173,600 samples in this dataset.Two methods are comparatively tested in the gesture recognition section.One of them is the feature extraction-based gesture recognition,it extracts two local binary pattern(LBP)features and a histogram of oriented gradients(HOG)feature.And the classification and recognition could be realized with the machine learning method of support vector machine(SVM).Another recognition method is constructing a convolutional neural network(CNN)model,which allow the direct import of gesture dataset.Through convolution in this method,the feature of gesture image is automatic extracted,the corresponding model is training and the recognition accuracy and time are tested by the test set.Finally,within the analysis of gesture recognition accuracy and time complexity,the comparison of CNN recognition method and three feature extraction methods is accomplished in detail.The experiment results reveal that,the recognition accuracy of CNN gesture recognition model is 96.23%,which is higher than that of feature extracted-HOG+SVM recognition method.For the analysis results of time complexity,the proposed gesture area extraction method consumes 21.05 ms,adding the CNN method-based average recognition time of 0.92 ms,the final average recognition time of CNN gesture recognition system is 21.97 ms.To sum up,in the application of realtime gesture recognition based on Kinect sensor,the utilization of gesture area extraction method and gesture recognition system of CNN recognition model could meet the requirement of real-time.
Keywords/Search Tags:Real-time hand gesture recognition, Kalman filter, Convolution neural network, Arm-removing, Feature extraction
PDF Full Text Request
Related items