The Activity Analysis Based On The Learning Of Motion Feature

Activity analysis is one of the important researches in video understanding. It aims to analyze the activities of the surveillance scene and obtain the semantic description of the surveillance scene. This thesis focuses on the moving object's activity analysis. To solve the activity analysis, we give a general procedure. Based on the general procedure, we have research on the extraction of motion information, the representation of the activity of moving object and the building of activity analyzer.1. The general procedure for moving object's activity analysis is introduced. Firstly, the motion information of the object is extracted with motion detection and tracking. Secondly, the activity of moving object is represented with the extracted motion information. Finally, an analyzer is used to analyze the activity of moving object and determine the activity.2. In allusion to the problem of moving object detection and tracking, we summarize the current researches and analyze the feature of these schemes. In order to improve robustness of tracking, we give a tracking algorithm that combined the extended Kalman Filter with Mean Shift method.3. On the research of the activity analysis based on Self-Organizing Mapping, we research the current algorithms and introduce the two different learning methods of the trajectory models and the flow vector models respectively. We also give a new method to represent the real-time activity of moving object with the position, the speed of the moving object and predictive motion information which imply the correlation between the motions of object. Several experiments are given to show that the proposed activity representation method is favorable for detect abnormal activity.4. In allusion to the activity analysis based on Bayesian Network, we proposed an adaptive method for detecting abnormal activities in video surveillance. We use a multi-Gaussian distribution to model a moving object's activities. The model parameters are updated automatically to satisfy the object motion properties when every new frame comes, and at same time, this moving object current activity can be recognized by means of its possibility. Several experiments are given to show the proposed method is efficient.