A novel occlusion-adaptive multi-object tracking method for road surveillance applications

Occlusion is one of the most challenging issues in visual surveillance. In the real road surveillance systems, there are different kinds of objects to be tracked in real time. These objects may occlude each other, which makes their detection quite difficult. This paper proposes a simple but efficient occlusion-adaptive multi-object tracking approach to resolve this issue. Our approach considers three object states, including normal state, occluded state and split state. In the normal state, it tracks the object according to the spatial continuity. In the occluded state, it establishes a constant velocity model to estimate the position of an occluded object. In the split state, it rematches an object with one object before occlusion according to their appearance features. Experimental results show that our approach can correctly track multiple objects under both partial and total occlusion in real time.