Overview of Pose-Assisted Multi-Camera Collaboration
Active Object Tracking (AOT) is an important and fundamental skill for a visual intelligent system, e.g. intelligent surveillance. However, it is difficult to learn transferable multi-camera collaboration under the high-dimensional visual observations, especially with the increasing number of cameras. Therefore, we propose a simple yet effective approach, which exploits the poses of other cameras to enhance the tracking policy, namely Pose-Assisted Collaboration Mechanism. Specifically, the pose-assisted mechanism is composed of a switcher and two controllers that are respectively driven by vision and poses. At each step, the switcher decides which controller to use according to the quality of the visual observation, then the selected controller outputs the action based on its own observation (vision or poses). In particular, the pose-based controller will be switched on when the visual observation is imperfect, e.g. the target is occluded or out of the view. It learns to assist tracking by rotating the camera to point to the same area as other cameras whose vision-based controller works well. The empirical results on unseen environments demonstrate that our system, trained in a randomized room, is capable of transferring to unseen environments and outperforms all the baselines.
[Github]
[Arxiv]
[YouTube]
@inproceedings{li2020pose,
title={Pose-Assisted Multi-Camera Collaboration for Active Object Tracking},
author={Li, Jing and Xu, Jing and Zhong, Fangwei and Kong, Xiangyu and Qiao, Yu and Wang, Yizhou},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={34},
number={01},
pages={759--766},
year={2020}
}
We would like to express our thanks for support from the following research grants MOST-2018AAA0102004, NSFC-61625201, NSFC-61527804, Qualcomm University Research Grant. We thank Tingyun Yan and Weichao Qiu for providing assistance with UnrealCV.