In recent years, multi-camera visual surveillance systems have attracted much attention and been widely used in applications such as continuously tracking interested objects, early warning of abnormal events, etc. Using multiple cameras, the area of surveillance is expanded and the occlusion problem which is very challenging for single cameras can be solved to a certain degree. In this thesis, we focus on the problem of object tracking across multiple non-overlapping cameras. The key issues include: (1) Topology estimation, (2) Color transfer across cameras, (3) Object matching and recognition across cameras and (4) Data association across cameras. The main contributions of this thesis include the following: 1) In order to compensate for lack of spatio-temporal cues across cameras due to blind areas, we propose a novel method to recover the topology of a multi-camera system based on N-neighbor accumulated cross-correlation functions, which is robust to noise and makes the peaks of cross-correlation functions sharper. Experimental results demonstrate our method can deal with the conditions of a long time window or large amounts of data while traditional methods fail. Meanwhile, our method avoids solving the problem of establishing correspondences across cameras, making it easy to be implemented. 2) In order to alleviate the influence of illumination variance across cameras, we use color characteristic transfer (CCT) models to deal with observations under different cameras. The CCT models are learned by impose one observation's color characteristics on another. Given CCT models, we can predict the observation of an object under one camera according to its observation under another camera. Different from other color transfer approaches, our method do not rely on a large training dataset with knowing correspondences. 3) To measure the similarity between two objects observed under different cameras, we propose a direction-based stochastic matching (DSM) method to realize object recognition across cameras. It depends on directional cues to deal with changes in viewpoint and a stochastic matching strategy to compensate for small variances in pose. Experimental results demonstrate the robustness of our method to viewpoint changes and pose variances. 4) As the DSM method is time-consuming, it cannot meet the requirement of a real-time multi-camera object tracking system. We propose a new algorithm based on adaptive models to solve the problem of object recognit...
修改评论