TrackDQN: Visual Tracking via Deep Reinforcement Learning

Abstract

Deep learning is a powerful method for visual object tracking, with excellent tracking accuracy and efficiency. However, the fast motion and larger deformation make a huge impact both on accuracy and efficiency of trackers, which makes more trackers less robustness with deep learning. In this paper, we make a attempt of applying deep reinforcement learning to improve robustness for Visual Tracking. The motivation is that deep Q-network has witnessed a success in solving various problems with computer vision and achieved promising performance on both running time and accuracy. In order to solve visual tracking problem scenario, we propose our Deep Q-Network for Visual Tracking(TrackDQN), which is inspired by deep Q-network and firstly customized from general deep reinforcement learning framework. Specifically, the tracker with TrackDQN can significantly improve the reliability and accuracy, meanwhile maintain higher tracking speed. Experimental results on the OTB benchmarks with fast motion and larger deformation demonstrate that our TrackDQN tracker has a comparable performance over state-of-the-art methods. Our method can also improve robustness and make a improvement of about 1.3% on accuracy with higher speed.