Multi-scale behavior learning for multi-object tracking
Wancun Liu, Wenyan Tang, Zhang Liguo, Xiaolin Zhang, Jiafu Li · 2017
Multi-object tracking is an important computer vision applications which is useful in many areas like video surveillance and automatic driving. Traditional, appearance and motion features are extracted for the tracking task, which is insufficient achieve excellent tracking results. In this paper, we propose a novel multi-scale behavior learning approach to explore the motion pattern of object's location and size. The multi-object tracking problem is formulated by a generative probability model, which contains a global and a local learning procedure. The global behavior is offline learned by a neural network, while the local behavior is learned by online learning. With these combination, the accurate prediction is made due to robust multiscale features. Experiments on challenging benchmark datasets shows that the proposed method achieved the state-of-the-art performance especially in the ID-switch measurement.