Real-time robust human tracking based on Lucas-Kanade optical flow and deep detection for embedded surveillance
Quang Dao Vu, Sun-Tae Chung · 2017
Object tracking is one of the most important functions in surveillance systems, especially in the system with Pan/Tilt/Zoom Camera. In this paper, we propose a real-time robust human tracking method for embedded surveillance. The proposed human tracking method tracks human objects based on Lucas-Kanade (LK) optical flow algorithm [1], rectifies tracking error due to accumulation or object missing by readjusting tracked human object bounding boxes periodically. Human localization information is obtained from a reliable deep learning-based human. It also handles occlusion by combination of LK information and human detector information. In order to achieve fast and robust processing, computationally light but reliable human detector is developed based on YOLOv2 object detector [2] model. Through experiments in comparison with other state-of-the-art tracking methods, it is shown that the proposed human tracking method operates fast and reliably with occlusion handling, and that performs better than or comparable to others.