Real-time stage-wise object tracking in traffic scenes: an online tracker selection method via deep reinforcement learning

Xiao Lu, Yihong Cao, Sheng Liu, Xuanyu Zhou, Yimin Yang

Abstract: How to ensemble the high representative ability for discriminating the target from its background and high adaptive ability to fast appearance changes, while keeping the real-time performance simultaneously is still an open topic in the field of object tracking, especially in the complex urban traffic scenes. To address this issue, motivated by that existing excellent trackers may have their advantages in tackling different kinds of tracking difficulties respectively, we propose a new real-time stage-wise object tracking method that allows different trackers to complement each other and combines their respective advantages. A tracker selection agent is trained to learn the policy of switching to the most appropriate candidate tracker according to the current tracking environment. To capture the dynamics of tracking environment effectively, we consider the tracker selection problem as a Partially Observable Markov Decision Process problem. A lightweight deep neural network with the recurrent unit is designed for learning the optimal policy accurately and rapidly. We also elaborately collected Traffic Scenes Object Tracking Annotated Dataset (TS-OTAD) for demonstrating the effectiveness of our method. Experimental results conducted on TS-OTAD and OTB-100 demonstrate that our method has superior performance than any of the candidate tracker and has a good trade-off between accuracy and efficiency compared with other state-of-the-art methods. Besides, our stage-wise tracking framework is not limited to any specific tracker, and any excellent tracker can be used as the candidate, which provides a new way for boosting object tracking accuracy and efficiency.