Action-Improved Actor-Critic Tracking for Accurate Object Tracking. Issue 1 (April 2021)

• Record Type: Journal Article
• Authors: Hu, Yangrun; Wu, Ruikang; Gu, Yanjun
• Is Part Of: Journal of physics. Volume 1903:Issue 1(2021)
• Publication Date: 2021-04
• Subjects: Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/1903/1/012010
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: Visual object tracking is a fundamental problem, which tracks the target in the whole sequence with the given object in the first frame. Numerous tracking frameworks have been proposed. Actor-Critic Tracking (ACT) is based on RL, which is a popular framework with real-time speed. However, ACT cannot handle scale estimation effectively, since it cannot change the ratio of target size. In this paper, we aim to improve the scale estimation of ACT and propose the Action-improved ACT(AIACT) model. This work is based on the basic framework of ACT. ACT builds the 'Actor' deep neural network to track the target and the 'Critic' network as an inspection module to improve the model. We add a new element to the bounding box of 'Actor' model to change the aspect ratio. The experiment result shows that our model achieves better accuracy than the previous model. Our method achieves 0.646 and 0.912 score of success rate and precision rate on OTB dataset, which means AIACT model can handle the scale variation of the target. Therefore, it will benefit more applications.