The Research Of Key Algorithms On Multi-Camera Moving Object Tracking And Association

Intelligent video surveillance has become one of the most active fields in computer vision research. The purpose is to extract useful information efficiently in a large number of video taken from cameras for automatic detection, tracking and recognition of objects of interest, understanding and analysis of their behavior. This thesis focuses on the related problems encountered in multi-camera tracking system. Based on multi-camera surveillance scene, detection of moving object, tracking of moving object and object association(i.e. consistency judgment) between cameras(overlapping view and non-overlapping view) are researched in details. In addition, propose the improved solutions on the basis of the previous study. The main contents of this thesis are as follows:1. Studied methods of moving object detection. Use three-frame-difference method and time-average-method to build a background image, then use background subtraction method to detect moving object. Frame difference method is adaptable to various illuminations and works in real-time. Background subtraction method can detect intact profile of object. This algorithm combines these two methods together and has all the advantages. Test results indicate that this algorithm can be used to effectively detect objects.2. An improved Camshift(Continuously Adaptive Mean shift) algorithm is introduced. As the Camshift tracking algorithm can’t track well when the color of object and the background is close, use a new kernel function to weight probability of the corresponding pixel which belong to the object in the search window, then obtain the reversed projection image after weighting, which reduces the influence of background color. At the same time, build a more robust object model by fusing hue, saturation and histogram of oriented edge. The experimental results show that the improved algorithm has stronger anti-interference ability, better robustness, higher track accuracy. By fusion of multiple features this algorithm does not increasing the computational complexity and basically satisfies the real-time requirement.3. An object association method based on overlapping fields of view is introduced. For multi-camera have overlapping fields of view, use four matched points are space coplanar but non-collinear to obtain homography, then recover field of view lines of multiple cameras, generate overlapping fields of view through field of view lines. When moving objects occur in the overlapping fields of view of the camera, the object coordinates are mapped to the same camera, calculate the Euclidean distance between the objects, if the Euclidean distance is larger than a predetermined threshold, they are not the same object, and if only one object is less than the established threshold, the object is the same. If multiple objects’ Euclidean distance is less than the threshold, then calculate the similarity of the color histogram, the object has the highest similarity is the same goal. The experimental results can prove that the algorithm proposed in this paper is more simple and accurate than the method based on field of view lines.4. For multi-camera have overlapping fields of view,Hue, histogram of oriented edge, SURF and the shape of object fused by D-S evidence theory to realize object association. The experimental results show this method is more robustness than using single feature.