Enhancing Active Visual Tracking Under Distractor Environments

Abstract

Active Visual Tracking (AVT) faces significant challenges in distracting environments characterized by occlusions and confusion. Current methodologies address this challenge through the integration of a mixed multi-agent game and Imitation Learning(IL). However, during the IL phase, if the training data of students generated by the teacher lacks diversity, it can lead to a noticeable degradation in the performance of the student visual tracker. Furthermore, existing works neglect visual occlusion issues from distractors beyond the collision distance. To enhance AVT performance, we introduce a novel method. Firstly, to tackle the limited diversity issue, we propose an intrinsic reward mechanism known as Asymmetric Random Network Distillation (AS-RND). This mechanism fosters target exploration, augmenting the variety of states among trackers and distractors, thereby enriching the heterogeneity of the visual tracker’s training data. Secondly, to address visual occlusion, we present a distractor-occlusion avoidance reward predicated on the positional distribution of the distractors. Lastly, we integrate a classification score map prediction module to bolster the tracker’s discriminative abilities. Experiments show that our approach significantly outperforms previous AVT algorithms in a complex distractor environment.

This work was supported by the Science and Technology Innovation 2030 Major Project under Grant No.2020AAA0104802 and the National Natural Science Foundation of China(Grant No.91948303). Supplementary material is checked from URL.