Research On Some Key Techniques Of Surveillance System

The calibration of camera system and object detection, tracking are two key area in the research of surveillance system. With the development of surveillance system, it becomes more complicate and more intelligent. So, the number of cameras in a surveillance system is increasing and the need of analyze content of scene is more and more important that research of the two area is important and meaningful.There are two approaches to calibrate camera system: strong calibration and weak calibration. The goal of strong calibration is to estimate the project function from real 3D world to the camera image plane under certain camera model. While the weak calibration does not directly find the project function in strong calibration but estimate a relative project relationship between cameras. The strong calibration is the base of multiple camera system and the weak calibration is a key to surveillance system.Video object detection is to detect some special target in video which include moving object detection, face detection, pedestrian detection, car detection etc. The subject comprises techniques in computer vision, image process, pattern recognition and machine learning. Vieo object tracking is to track object of interest in the video. As to object detection, object tracking also integrates several subjects and is the key area in computer vison.At first, the thesis proposes an automatic calibration method based on mosaic image in a master-slave camera surveillance system. This method firstly composes a mosaic image by the PTZ camera, and then, calibrates the surveillance system by selecting and matching feature points.In the second, the thesis proposes an object tracking algorithm based on the combination of Online RankBoost and particle filter. The algorithm creates and updates a detector which can discriminate the object and background by online RankBoost. Then the method uses particle filter to predict the moving function of the object.The experiment shows that both of the two algorithm proposed by this thesis are effective and robust.