Overhead View Person Detection Using YOLO

Abstract

In video surveillance system, one of the important task is to detect person. In recent years, different computer vision and deep learning algorithms have been developed, which provides robust person detection results. Majority of these developed techniques focused on frontal and asymmetric views. Therefore, in this paper, person detection has been performed from a significantly changed perspective i.e. overhead view. A deep learning model i.e. YOLO (You Look Only Once) has been explored in the context of person detection from overhead view. The model is trained on frontal view data set and tested on overhead view person data set. Furthermore, overhead view person counting has been performed using information of classified bounding box. The YOLO model produces significantly good results with TPR of 95% and FPR up to 0.2%.