Research On Multi-Camera Target Tracking Handoff And System Implementation

Target tracking is always a hot research topic in the computer vision and it has wide applications in many aspects especially in intelligent surveillance. With the development of hardware and software technology, intelligent surveillance has become more and more networking and intelligent, compared with the primary video surveillance. More and more researchers have joined the multi-camera intelligent surveillance research. Target continuous tracking in the multi-camera network is the foundation of intelligent surveillance system and how to handoff the tracking target among the monitor cameras is the core research issue. This thesis mostly studies the tracking target handoff problem among cameras. The target automatic location in the camera can be treated as a target initialization problem; so this thesis firstly studies the target automatic location problem based on this initialization idea and then makes a preliminary research on the multi-camera target handoff problem; at last implements an algorithm platform that can be used to make multi-camera target tracking experiments based on our PTZ camera hardware system.This thesis does some investigation on how to realize the automatic pedestrian target initialization problem based on the idea that target locating is an initialization process, especially when there are only some descriptive features. How to use the fuzzy descriptive features to locate the object that match the features in the video frames is the mainly research topic of the first part of this thesis. An automatic target initialization algorithm based on a target color-blob model is proposed. First, to overcome the descriptive features’ ambiguity, a membership concept which firstly used in the fuzzy logic is used to quantize the descriptive color features and convert them to the target color-blob model; and then with the help of the pedestrian body parts geometric ratios and motion features, the automatic target initialization task can be achieved. Three initialization experiments under different circumstances are given in the thesis. The results indicate that the method proposed here can successfully locate the target automatically when the target color texture feature is simple.A multi-camera target handoff algorithm is proposed in the second part. It is based on the combination of the target background estimation and the topological model of the surveillance environment. A Kalman filter with a motion estimation sampling is used as the tracking method; a target background estimation module based on the ORB features is used to estimate and define the handoff time and position while tracking the target in a camera. When the background estimation module judges that the target enters the region of interest and target motion feature satisfies the requirement, a target handoff process begins:the camera assignment algorithm based on the environment topological model gives the next tracking camera and the camera can move to the handoff position quickly using the PTZ camera’s preset function, then the target will be located as an initialization process, after it is located, the handoff process finishes. The experiment results show that the method can quickly finish the handoff task no matter there are public fields of view among tracking cameras or not.At the last part, this thesis design and implements a multi-camera target tracking algorithm experiment platform based on4PTZ cameras. The platform encapsulates the common modules so it can avoid the cost of the redevelopment and make the subject’s further study more convenient. The tracking algorithm module is designed as a framework structure so that different implementations of a same module can replace each other and compare the results. At last a multi-camera target tracking experiment based on this platform is given; the result indicates that it successfully completes the target tracking and handoff task.