Human Motion In Video Behavior Recognition

Human activity recognition is an important area of computer vision research today. Recognizing human actions from video sequences is a challenging problem in computer vision. Visual analysis of human motion is an emerging and important area of research, involving pattern recognition, image processing, machine vision, artificial intelligence and other disciplines. It can be widely used in many fields such as motion capture, human-computer interaction, surveillance and security, environmental control and monitoring, sport and entertainment analysis, etc. Behavior recognition of human movements in computer vision research is active and difficult in recent years. Most motion research is focuses on a simple gesture, gestures, facial expressions, gait, etc., and the background is relatively simple, and the existing methods face low recognition accuracy or high computational cost.This thesis aims to conduct innovative and exploratory research into some of the more complex actions such as theft, robbery, fighting, and more complex background real scenes from the network. In addition to the expansion of the research, this thesis summarizes and analyzes some of the state-of-the-art methods in recent years and based on its merit and shortcomings, proposes a new descriptor: PM-PEMO spatial-temporal pyramid features, for the understanding and description of video content. This feature not only contains the local information but also contains global information that can better describe the behavior, with a strong anti-interference ability and anti-noise sound and robust. With what we constructed PM-PEMO feature above through some of the advanced machine learning methods: online dictionary learning, sparse principal component analysis, locality-constrained linear coding, and distance metric learning for machine learning to obtain the video feature representations .And then we combine both multi-task margin nearest neighbor and linear support vector by scoring mechanism for classification of the video feature representations, and significantly improve the recognition results.Our method is simulated by MATLAB and made into Applications Software using C, C + +, MFC, OPENCV. Our proposed method is tested on Intelligent Visual Information Processing and Communication Laboratory (IVIPC) video database, web video database, Weizmann video database and KTH video database, and several state-of-the-art methods proposed in recent years are used for comparison. The experimental results demonstrate that the method gains real-time performance and higher accuracy.