Research On Several Problems In Multi-cameras Cooperation Distributed Intelligent Visual Surveillance

The intelligent image sensor is the critical equipment of distributed intelligent visual surveillance system. If using high capacity computers to design the image sensor, the cost would be high. Furthermore, the reliability of common computers is low, so that they won't meet the need of the visual surveillance system's continuous working. Recently, the price of digital signal processor(DSP) has been dropping continually while the computing speed of DSP has been rising sharply, so DSP could be used to design intelligent image sensors aiming at reducing the cost and enhancing the reliability. Compared with high capacity computers, DSP's computing ability is limited, so most of complex algorithms which have obtained fine results on computers can not to be applied on DSP directly. New algorithms should be presented and existing algorithms should be improved in terms of DSP's computing ability. The aim is to reduce the complexity of the algorithms as much as possible on the condition of ensuring the capacity of the algorithms.Based on the hardware platform of network cameras designed with DSP, several fundamental problems in intelligent visual surveillance are researched in this paper, such as background reconstruction, detection and tracking of motion objects, detection and tracking of human faces, camera calibration, cooperative surveillance of multi-cameras. The main point is solving the contradiction between the algorithm complexity and it's capacity. The main productions and innovating points are as follows:(1) One of the most important basis of distributed intelligence visual surveillance is to reconstruct the background quickly and correctly. The paper proposes two fast background reconstruction algorithms. These are the improved algorithm which is based on the hypothesis that background pixels appear most frequently and a background updating method under the instruction of object division. In the first method, the foreground points should be removed at first, and then record the appear frequency of every background point, choose the pixel value which appears most frequently as the background pixel value. The algorithm can reconstruct the background quickly .And in the case of foreground points appear more frequently than background points which caused by the tardy moving of object or the object's continual appearance in an area, it also can reconstruct the background correctly .Another method detects the "false" moving areas by motion division, and only updates these areas .The computation is very small and updating speed is high. Furthermore, the algorithm is not restricted by the appearing frequency of background points.(2) Single camera tracking is the foundation of multi-cameras tracking. The paper proposes a kalman prediction multi-objects tracking algorithm which combines both motion detection and motion search. It utilizes the strategy of tracking belief degree and can deal with the case of shelter, combination and abruption .The algorithm has a strong ability of multi-objects tracking.(3) Face detection and tracking is a very important function of the visual surveillance system. The paper combines both the skin-color model and the distributing characters of the edge direction points in the face region, proposes a real-time face detection algorithm which is based on the Bayes classifier, and also realizes multi-faces tracking by face detection and Kalman filter.(4) Obtaining the object's world coordinates by camera calibration is one of the most important tasks in visual surveillance. In the paper, an online calibration method of camera rotating matrix based on quaternion is proposed. The method is simple and the error between the object's rebuilt world coordinates and real coordinates is small. Additionally, a new method of object's localization with single camera based on human height model is proposed. The 2D coordinates on ground can be obtained with this method. It can meet the need of single camera visual surveillance inside or outside.(5) Cooperation of multi-cameras is an important method of increasing the intelligence of the visual surveillance system. A model of multi-cameras cooperation using coalition mechanism is proposed in the paper. It chooses multi-cameras to track objects using the fast coalition formation algorithm which is based on the match between task and the capability of camera. An algorithm of weighted merging is used to increase the precision of the tracking.