Research On Human Hand Behavior Analysis Theory Based On Deep Learning

Human hand behavior analysis can be applied to many fields,such as sign language recognition,medical care,virtual reality entertainment and human-computer interaction,and the method based on convolution neural network is the mainstream of current research.With the continuous development of deep learning,more and more network models have been designed,improved and applied.The powerful feature extraction ability of convolutional neural network enables it to play an excellent role in computer vision tasks.Human hand behavior analysis mainly includes hand detection,gesture estimation and hand gesture recognition,among which the hand gesture recognition module plays a decisive role in behavior analysis.In the traditional computer vision task,the research object of human hand behavior analysis often stays in a single RGB image,that is,two-dimensional static gesture.With the wide application of depth camera and the maturity of depth learning algorithm,more and more researchers begin to focus on the research of dynamic gesture based on skeleton.Skeleton-based dynamic gesture has rich spatiotemporal information,and data acquisition is relatively simple,and is not easy to be disturbed by factors such as light and dark,background transformation,etc.Although skeleton information has the above advantages,there are still many difficulties in its research,such as skeleton noise interference,multiangle problem and slow model running speed.In order to improve the performance and real-time of gesture recognition in complex background,after analyzing the characteristics of human hand behavior,this paper selects a variety of ways to extract different aspects of gesture features,and inputs them into the designed neural network model for recognition and classification.Experimental results show that the algorithm successfully solves the interference problems of angle change,dynamic difference and complex environment on gesture analysis and recognition,and has high robustness to the differences of different subjects and geometric deformation of gestures.The experimental results on the open data set SHREC’17 show that the 14-category classification scheme achieves 95.4%accuracy,and the 28-category classification scheme achieves 92.1% accuracy,which is superior to the traditional algorithm.The size of the network model is 2.24 M,which can run in real time on GPU and meet the real-time interactive application conditions.