Moving Target Surveillance And Abnormal Human Behavior Analysis In Videos Of Complex Scene

Intelligent video surveillance technology as a crucial part of public security has recently beenan important research topic in the field of computer vision. This thesis focuses on moving targetdetection, moving target tracking and abnormal human behavior analysis in the video imagesequence; and studies the applications of the three key technique under different conditions. Themain work and contributions are summarized below:（1）Based on the global illumination model and spatial likelihood probability model, the thesisproposes an improved Gaussian mixture model (GMM) method. The moving target detectionmethod of the existing background substraction based on GMM is first introduced. Theoreticanalysis and experiments illustrate that this method does not work under the condition of suddencharge happens in the illumination scene. To solve this problem, the global charging relations amongimage pixels intensity are then discovered and an improved method based on the global illuminationmodel and spatial likelihood probability model is proposed. The proposed method is shown to havebetter performance and be robust to the sudden charge in the illumination scene.（2）Mean-shift algorithm are particle filtering are first respectively, studied to be used inmoving target tracking. Then a new target tracking method combined the advantages of the twoalgorithms is employed. In addition, an improved Mean-Shift algorithm which is based on the grayand moving information of target is proposed. The algorithm can improve the robustness of targettracking and solve the problem of inaccurate or lost tracking when the target and background havesimilar gray distribution at some moment.（3）The thesis proposes a synthesized recognition frame which an analyze human targetbehavior in a multi-stage and multi-level way. The thesis first describes abnormal behavior of humanactivity, and the performs features extraction in video sequence for human target, including motionfeatures, Hu movement features and contour features. The thesis extracts the contour features byusing level sets method, which can extract contour more precisely and obtain smooth and closecontour curve. After extraction features, this thesis constructs a frame that can perform a synthesizedrecognition through a way of multistage and multi-level.