The Research And Implement Of Human Action Recognition Algorithm

In recent years,behavior recognition has been widely used in virtual reality,video surveillance,video retrieval and analysis.It has also become one of the hottest topics in the fields of computer vision,pattern recognition and artificial intelligence,which has attracted more and more researchers' interest.Although the research and implementation of behavior recognition is of great application value and theoretical significance,the current theoretical research is not perfect,there are still many problems to be solved,including:(1)there is a great difference in the same type of behavior due to the complex background,the angle of view transformation,the different objects and so on;(2)different types of behavior sequences contain similar trajectories,so that different actions have smaller inter class variations;(3)high dimensional and redundant action sequences increase the time complexity of behavior recognition;(4)extract accurate behavior characteristics and design effective behavior recognition algorithm architecture.Based on human behavior information acquired by action capture device Kinect,this paper makes an in-depth study of human behavior representation and recognition methods.Firstly,the Kinect device and its working principle are introduced,and then the human action data are acquired by Kinect device and processed.Secondly,a dynamic behavior recognition method based on template matching method is implemented.Firstly,feature extraction of the preprocessed bone data;then the data set created in this experiment is introduced to make it have two features: high cohesion and low coupling between classes;then an improved DTW algorithm is proposed.Finally,the performance of the algorithm on the data set is tested and the detailed experimental results are given.Thirdly,a human behavior recognition algorithm based on state space method is proposed.Firstly,a variety of features of the preprocessed bone data are designed and fused;then the two K-Means algorithm is used to cluster the features to extract the set of the key posture of the human behavior;then,the motion data is modeled bymulti-layer HMM in the training stage,and the set of data sets and two are built in the test stage.The performance of the algorithm is tested by open datasets,and compared with the recognition algorithm based on template matching method from two aspects of recognition accuracy and recognition time.Finally,human body gesture recognition is realized by deep learning.Firstly,the basic knowledge of deep learning and virtual confrontation technology are introduced.Then,a small number of labeled data and unlabeled data are semi supervised by virtual confrontation training.At last,the experimental results are tested on the open data set and the effect of the method is analyzed from three aspects.Experimental results show that the three recognition algorithms have good performance in recognition accuracy and recognition speed.Experimental results show that these three recognition algorithms have good performance in terms of recognition accuracy and recognition speed.