End-To-End Visual Target Tracking in Multi-Robot Systems Based on Deep Convolutional Neural Network

Yawen Cui, Bo Zhang, Wenjing Yang, Zhiyuan Wang, Yin Li, Xiaodong Yi, Yuhua Tang; Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 1113-1121

The problem of one-on-one target tracking from a single monocular image acquired from the viewpoint of a follower robot itself is studied in this paper. Previous works mainly depended on locating, onboard sensors with control mechanism, while robot may not carry advanced onboard equipment for localization or GNSS may also fail in GNSS-denied/Indoor environments. In this paper we propose a novel approach based on a deep convolutional neural network called Deep-Track, which trains a supervised image classifier only using images captured by the camera in the follower robot. Specifically, the Deep-Track system can output the estimated velocity of the target as well as the velocity control for the follower, by operating merely on two adjacent frames. In order to verify the effectiveness of Deep-Track, we build up a large-scale dataset in the simulator, in which the performance of the Deep-Track is evaluated and it is shown that a high tracking accuracy is achieved.