Htld: A Human-in-the-loop Target Detection and Tracking Method for UAV

Yuting Zhu,Chang Wang,Yifeng Niu,Lizhen Wu

DOI: https://doi.org/10.1109/gncc42960.2018.9018702

2018-01-01

Abstract:It is a challenge for a flying UAV to detect and track a moving target in real-time, especially when the target is small without reliable features, blurred caused by motion, or the lighting conditions and the viewing angles change fast. These issues can easily result in tracking failures that are difficult for autonomous algorithms to recover from. In this paper, we introduce human into the loop of a target detection and tracking algorithm, so that the human can monitor the tracking result and correct it whenever the algorithm makes a mistake. In this way, training samples with better qualities can be provided to improve the accuracy of the object detection classifier. Specifically, we propose hTLD (human-in-the-loop TLD) based on the traditional TLD algorithm. When the tracking result is wrong, the human marks a bounding box around the observed target, and the selected box serves as the reference source to add a set of relabeled positive samples as well as negative samples. In addition, an adaptive strategy is used to adjust the bounding box as optimal as possible to separate the target from the noisy background. Experimental results have demonstrated that hTLD is more efficient and reliable than TLD in a scenario of a fixed-winged UAV tracking a ground vehicle.