Multi-object tracking based on multi-feature fusion with adaptive weights

In this paper, we propose a multi-object tracking method based on multi-feature fusion with adaptive weights. First, we introduce the feature extraction between two objects in two consecutive frames, including centroid distance, intersection area and histogram distance. Then the concept of feature similarity is introduced. In order to achieve accurate tracking, the new method employs a feature similarity function fusing the three features. The corresponding weight of each feature similarity is able to change adaptively according to the previous state of the same object. Moreover, a template matching based on weighted histogram is employed when mild occlusion between two objects exists. Experiment results show that our scheme can effectively track the objects with quick movement even when the object is in occlusion, in comparison with the methods adopting constant weights.