Several Issues Studies On Non-overlapping Multi-camera Object Tracking

The non-overlapping multi-camera object tracking is a hot research direction in video surveillance. It aims to achieve object’s relay tracking under different cameras. In the meantime, these cameras have no overlapping visible areas. The non-overlapping multi-camera object tracking system involves multiple research areas, such as multiple objects detection and tracking, object correlation and topology structure among the cameras. So, it is really difficult to build a relatively perfect tracking system. However, the system saves manpower and material resources (for example, the reduction of setting up cameras), so it has quite significant practical application value to do more research on the system.This paper will introduce the basic theory of object detection and tracking first. After that, a real-time tracking algorithm is proposed. At last, we design a whole tracking system which includes object detection, object tracking and object correlation algorithm. The main works and contributions in this paper is as follows:(1) The paper proposed a real-time compressive object tracking algorithm based on adaptive feature selection. Based on compressive tracking, we select the compressive feature and use a difference method to select adaptively at the same time which improves the robustness of tracking. Compared with several tracking algorithm, our algorithm has better performance in real-time and tracking accuracy.(2) The paper designed a non-overlapping multi-camera object tracking system. The system mainly includes two parts:the tracking module which completes the object detection and tracking work in a single camera and the correlation module which completes the object correlation work. In the tracking module, the detection algorithm is VIBE algorithm and the tracking algorithm is bidirectional matching algorithm based on detection results. The experimental results show that the combination of these two algorithms can track the objects effectively. In the correlation module, firstly, we train the network topology structure among the cameras and propose a Gaussian mixture method to estimate the camera time transition probability. The learned topology structure can be used as prior knowledge for object correlation which improves the accuracy of object correlation. Secondly, we use BTF to eliminate the object’s imaging differences between different cameras. Combined with the network topology structure, the final object correlation result can be obtained. At last, we present these two modules’ interface respectively.