An Adaptive 2D Tracking Approach for Person Following Robot
Shenlu Jiang, Ling Li, Meng Hang, Kuc Tae-Yong · 2017
In this paper, we present a 2D appearance vision based tracking approach for human following robot. Generally, existing methods have high cost of computing and requirements of hardware which makes the difficulty of employing the function on service robot. Hence, minimizing the cost of tracking with slight loss of precision can benefit this area. We focus on approach based on 2D image data which reduce tracking into two dimensions and can minimize the cost of computation. Our approach presents a corporate strategy which utilizes central consensus of correspondence for 2D feature points pairwise to track and employ a semi-supervised learning detector to update appearance change. To overcome difficulties from environment change, we set an enhancing process for feature points and background segmentation through depth information. We carefully evaluate our approach with common challenges of visual tracking in static view and deploy a dynamic view a real-world following task. The experiment results illustrate that our tracking approach works against common risks at 2D appearance tracking and properly follows the user obtaining 25 fps performance on mobile platform.