A novel multi-object tracking method based on main-parts model
Hongbin Liu, Faliang Chang · 2017
Multi-object tracking is often hindered by difficulties such as occlusion and illumination change. In this paper, we propose a novel multi-object tracking method based on main-parts model. Main-parts model is formulated by segmenting parts of object and accumulating variations of appearance from previous frames. We assume that parts with weaker appearance variations are main-parts of an object. By using main-parts model, data association between tracklets and detections can be more accurately. Experiments with challenging public datasets verify that the proposed method can improve tracking accuracy.