Object Inter-camera Tracking with Non-overlapping Views: A New Dynamic Approach

Disjoint inter-camera object tracking is the task of tracking objects across video-surveillance cameras that have non-overlapping views. Unlike the closely related task of single-camera tracking, disjoint inter-camera tracking is difficult due to the gaps in observation as an object moves between camera views. To overcome this problem, appearance profiles of the objects seen in each camera are built and used for matching/tracking across different cameras. This paper proposes a new method that uses multiple features that are dynamically weighed for matching moving objects (people in our case) across cameras. In particular, the Zernike moment shape descriptor has been used together with blob histogram and other features to describe a moving object. Weighting emphasis is given to the better features, based on their stability, reliability and their time in the system (how recent they are). This weighting is used both during appearance aggregation and object comparison. Our experiments with real videos have shown the success of our proposed method even in difficult situations where the cameras used are different in terms of brand, quality and resolution.