Human-in-the-loop Multi-task Tracking Improved by Interactive Learning

It is a challenge to track a moving target with deformation, occlusion or motion blur in a video stream. Various machine learning algorithms have been proposed for target tracking such as TLD (Tracking-Learning-Detection) and MTT (Multi-task Tracking). However, these algorithms still have limitations to detect and recover from tracking failures. In this paper, we propose hMTT that introduces human into the tracking loop of MTT. Specifically, the tasks are represented and separated as related particles and outlier particles. We design a human-guided mechanism that keeps the related particles as transferable knowledge and replaces the outlier particles with more reliable ones whenever a tracking drift or a tracking failure happens. Meanwhile, the multi-task learning module keeps updating the shared features stored as templates in a dictionary. In addition, the overall confidence value of the tracking result can be calculated accordingly, and the confidence level can be learned on-line for result classification. Experimental results have demonstrated that hMTT outperforms TLD and MTT in various tracking tasks.