Key Technology Research On Motion Capture Based On Videos

Motion capture is the technology to recover human posture parameters such as joint positions and angles from one or multiple visual angles. This technology is related to many research fields such as computer vision, pattern recognition, human kinematics, artificial intelligence and so on. The application fields include intelligent video surveillance, film and animation production, novel human-computer interaction, content based image and video search.This dissertation focused on human action analysis in monocular videos. The method used is based on the template matching of2D shape analysis. Human silhouettes in videos including human motion are relatively easy to extract and contain rich information so that they can be used as the main object to be analyzed. The specific steps are human detection in foreground, low-level feature extraction and motion classification. Background subtraction is used to get moving regions and the false targets are removed; the low-level features are mainly the information of the human body shape; template matching is used to classify actions. The efficiency of the method is verified through applying on public action video set--Weizmann action library.Aiming to initialize the process of silhouette segmentation, this dissertation put forward a method to label body joints automatically. This dissertation also presented an intelligent video monitoring demonstration system, which can distinguish abnormal behaviors in surveillance videos. These abnormal behaviors include falling to the ground, crossing pre-set line, moving away objects: The innovation of this dissertation lies in the two following points:1). Using relative position of hands and face to analyze human actions with self-occlusion. Considering that shape information of human body silhouette can hardly be used to analyze human actions with self-occlusion, this dissertation proposed to get locations of human hands and face through skin color detection, and then to calculate the relative position of hands and face and to determine the pose of human body. This method can be applied to analyze and recognize several common human actions with self-occlusion.2). This dissertation puts forward a method to label joints of human body automatically. In the process of action recognition, the automatic human body parts segmentation is of great importance to failure recovery of human tracking and correct identification of body posture. This dissertation utilized skin color, silhouette shape and the physiology of the human body proportion to make human body segmentation, which can label joints of front human body automatically.