A multi-hypothesis approach for salient object tracking in visual surveillance

In this paper, a multiple-object tracking method for visual surveillance applications is presented. Moving objects are detected by adaptive background subtraction and tracked by using a multi-hypothesis testing approach. Object matching between frames is done based on proximity and appearance similarity. A new confidence measure is assigned to each possible match. This information is arranged into a graph structure where vertices represent blobs in consecutive frames and edges represent match confidence values. This graph is later used to prune and refine trajectories to obtain the salient object trajectories. Occlusions are handled through position prediction using Kalman filter and robust color similarity measures. Proposed framework is able to handle imperfections in moving object detection such as spurious objects, fragmentation, shadow, clutter and occlusions.