Moving Object Detection And Tracking For Multi-camera Video Surveillance

With the high-speed development of video monitoring system, video surveillance in dynamic scenes, especially for humans and vehicles, has become one of the most active research topics in computer vision. It has a wide spectrum of promising applications, including access control in special areas, human identification at a distance, crowd flux statistics and congestion analysis, detection of anomalous behaviors, and interactive surveillance using multiple cameras, etc.This thesis focuses on the key technologies of multi-camera intelligent video surveillance and gives more attention to the enhancement of real-time and robust performance. The ultimate goal is to develop a practical multi-camera intelligent surveillance system. The main contribution consists of three parts as follows.Motion detection: Background subtraction is widely used in moving objects detection. After a brief review, an adaptive multi-model fast background subtraction is proposed which can automatically adapt to the arbitrary distribution of background pixel process. Compared with mixture Gaussian model algorithm, the method maintains the merits of multi-model and decrease the algorithm time cost greatly. The thesis also attempts to give a solution of motion detection in moving camera video. Background subtraction and SIFT motion compensation are integrated and give a promising experimental result. The issues of moving shadow and ghost shadow in motion detection are also studied here and given solutions respectively.Occlusions handling: Occlusions is the most difficult point in video object tracking. In the proposed approach, occlusions are detected by analyzing MBB overlapping feature in consecutive frames, objects are described by probabilistic appearance models and tracked through occlusions under framework of CONDENSATION. The merit of this approach is can handle occlusions between objects of various types, such as pedestrian and vehicle.Multi-camera collaboration: Collaborative multi-camera tracking is a hot topic in recent years. A novel approach is proposed which aim to object matching, location and identification in a complicated multi-camera scene. In the approach, head detection and epipolar relationship is used to confirm correspondence of observations between different camera views, and trifocal tensor transfer is used to locate objects in virtual top view. The sequential Bayesian method is applied to real-time label trajectory after Kalman+PDA tracking. Compared with existing approaches, the approach doesn't need camera calibration and coplanar precondition, doesn't exist the deterioration of epipolar transfer, can label the trajectory segments real-time and accurately. A practical multi-camera experimental system in the lab is introduced finally.