Chang Wang , Yuting Zhu , Xupeng Wen , Yifeng Niu , Wu Lizhen
July 2018
Proceedings of the 32nd International BCS Human Computer Interaction Conference (HCI)
Human Computer Interaction Conference
4 - 6 July 2018
It is difficult for a human operator to monitor multiple video streams in real-time surveillance tasks. A variety of machine learning algorithms have been proposed for target tracking, but human intervention is still necessary because the algorithms have limitations to detect and recover from tracking failures. In this paper, we propose a supervisory target tracking method for multi-video surveillance tasks based on hTLD (human-in-the-loop Tracking-Learning-Detection). The human operator can monitor both the original video streams and the calculated confidence values of the tracking results. Human intervention allows to update the training data for TLD whenever a tracking failure is observed, by manually drawing a bounding box or using an eye-tracker to relocate the target. Meanwhile, an interactive on-line learning algorithm is used to learn the relations between the confidence values and the tracking results. In this way, the confidence values help avoid unnecessary intervention caused by false alarms of tracking failures. Experimental results have demonstrated that the proposed method can reduce the task completion time.
This work is licensed under a Creative Commons Attribution 4.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/