Reinstating Dlib Correlation Human Trackers Under Occlusions in Human Detection based Tracking

Human tracking in video feeds has been a popular research area, which consists of different state of the art approaches including tracking-by-detection, feature tracking and tracking by recognition/re-identification. Dlib correlation tracker is a state-of-the-art tracking-by-detection implementation, which can be used for human tracking except under occlusions. That is, it fails to continuously track a person if that person disappears due to occlusions and reappears immediately. In order to overcome this challenge, we propose a state estimation approach to extend the capabilities of Dlib correlation tracker under occlusions along with the use of human detection. The proposed approach has proven to give better results than directly using Dlib correlation tracker with human detection and was evaluated in real-world experiment scenarios. The approach can further be improved for wider use cases and a wider variety of trackers for human tracking.