Deep neural networks which widely used in image classification and speech recognition have been successfully applied to model-free object tracking. However, during tracking, it easily falls into over-fitting problem, when the object size is either over-estimated or underestimated during tracking. Besides, the increasingly complicated discriminative model which strengthens the ability to identify object under highly occlusion also raises the opportunity of getting poor samples for training. In this paper, we propose a visual tracking algorithm based on deep discriminative map. The method guides the tracking algorithm by estimating the object's size and shape, and whether it is proper to gather training samples. Our method utilises two neural networks, one focusing on the center of object and one focusing on the object appearance. Experimental result on 13 public challenging tracking sequences shows that our proposed framework is effective and produces state-of-art tracking performance.
展开▼
