The Human Behavior Recognition Research And System Design Using Kinect

Human behavior recognition was a new direction related to image processing, pattern classification, artificial intelligence and other science domains. Now it had been widely used in intelligent surveillance, human-computer interaction, intelligent robot, visual reality, motion analysis and so on. Through taking the Kinect somatosensory equipment launched by Microsoft and surrounding human behavior recognition problem in a limited area, this thesis studied the human behavior representation and recognition methods for entrepreneurs and designed the Kinect experiment system as research platform. The main contributions were as follows:This paper firstly designed experiments to test Kinect depth information awareness and human skeleton identification capability after having know about the structure and performance of Kinect, then determined the experimental conditions in this thesis based on the experimental result analyze. All of these laid the foundation for next research.How to extract the behavior characteristics effectively was the key step in human behavior recognition. According to the Kinect data features, this paper proposed a new human behavior representation method based on human structure information: First, creating the human structure vectors by human skeleton joints, then selecting the combination of angles and module ratios between vectors to establish the body gesture description vector, at the last taking the sequence of body gesture description vector as the behavior representation characteristics. The characteristics were verified by experiments with translation and scaling invariance, also they were easy to solve and information-rich, so they entirely meet the research requirements.Behavior recognition method research was another important part of this paper. The human behavior recognition technology based on BP neural network and DTW algorithm(Dynamic Time Warping) was emphatically studied. For BP neural network, the mainly work concentrated upon determining the BP neural network structural parameters through the experimental analysis, while for the DTW algorithm, selection of characteristics similarity calculation rules and algorithm optimization were more important. At the same time a new behavior search method was designed to achieve real-time behavior recognition based on DTW algorithm. Finally, this paper proposed a human behavior recognition method which can integrated use of human skeleton joints, scene depth and visual information under a specific experimental scene, and had done some experiments to test the method.In order to complete the research tasks, we designed and developed the Kinect experiment system and some other software tools. Through a lot of experimental tests showed that the BP neural network and the DTW algorithm in this paper all yielded better recognition results, which indicated that the behavior representation and recognition methods were feasible. These tasks had a certain application prospects and academic value, also, they would have reference function to the further research and development.