Human Action Recognition Method Based on Historical Point Cloud Trajectory Characteristics

Human behavior recognition is a research hot spot in the field of computer vision. However, due to the interference of a complex environment and the diversity of behavior itself, it is difficult to extract suitable behavior features for behavior recognition, thus increasing the difficulty of recognition. In this paper, we propose a new behavior feature called the historical point cloud track feature which includes depth information and skeleton information obtained by Kinect, to solve the problem that the existing methods lose time series and spatial information. We reconstruct a three-dimensional point cloud by using depth information and then keep the 3D point cloud near limbs overlapping with skeleton information; then, we obtain the corresponding features. This feature set can well represent the spatial distribution of the point cloud of the main moving parts of the body. In addition, we use the time pyramid to segment the time series at different scales and splice the characteristics of each time period to strengthen the order relationship for the associated behavior. Finally, a support vector machine is used for training and recognition. Several groups of comparative experiments on the UTD-MHAD dataset show that the recognition method of human behavior based on historical point cloud trajectory features is very effective, and the recognition rate reaches 90.23%, which is higher than that of similar comparison algorithms.