Research On Human Behavior Recognition Algorithm Based On Kinect Multi View Feature

After decades of development,human behavior recognition has become one of the important research directions in the field of computer vision and pattern recognition,which has been successfully applied in many aspects,such as video surveillance?machine control?virtual games and so on.With the emergence of Kinect human-computer interaction equipment,the research object of human behavior recognition has gradually changed from RGB image to depth data.Because the traditional RGB image is vulnerable to illumination interference,the depth data with strong robustness to background noise has been widely concerned.In order to avoid the shortcomings of self-occlusion of bone data or misjudgment of similar behavior in depth image,this paper proposes a deep learning algorithm for human behavior recognition based on Kinect multi view features combining Kinect multi-modal data,which uses convolution neural network to automatically extract the spatiotemporal features of behavior sequence.Firstly,aiming at the problem of how to make full use of the spatiotemporal information of behavior sequence,an improved preprocessing algorithm based on depth image and bone data is proposed,which transforms the depth image sequence into depth motion map,transforms the standardized three-dimensional coordinates of bone nodes into a three-dimensional matrix arranged in chronological order,and constructs the characteristic vectors of bone joints to describe the topological shape of human skeleton;Secondly,aiming at the problem of how to train multimodal behavior data effectively,a five channel convolutional neural network(5CCNN)is proposed based on the idea of dual stream CNN architecture,and a single channel CNN structure suitable for MSR Action 3D database is designed,the first three channels are used to train the depth motion map,the second two channels are used to train the three-dimensional matrix of bone node and the feature vector of bone joint,and the recognition results of the five channels are fused in the later stage;Finally,experiments are carried out on MSR Action 3D database to further discuss and analyze the impact of different integrated decision-making methods on recognition performance,propose the most suitable model for the average value method of integrated decision-making,the experimental results show that the recognition accuracy of the algorithm reaches 98.1%,which verifies the effectiveness of 5C-CNN model construction,and by the confusion matrix to comparing with other algorithms,propose the subsequent improvement of the algorithm.