Matching Objects in Multi-Camera Surveillance without Geometric Constraints

Intelligent visual surveillance by multiple cameras has becoming more and more important in recent years. A novel and effective method to match objects between multiple cameras is proposed in this work, based on the proposed wavelet salient features. Firstly, the background and foreground is separated based on Gaussian Mixture Model (GMM). To achieve a reliable matching objects between cameras, the wavelet salient features consist of both the color feature components and the salient value that are extracted from salient regions, rather than the color feature from the whole object as in the traditional methods. To make the matching more robust against occlusion, we design dynamic learning templates with wavelet salient features which are constructed and updated in one camera, and used to match objects in another camera without any geometry constraints. Experimental results demonstrate the effectiveness of the matching in different views, even when occluded.