VRGNet: A Robust Visible Region-Guided Network for Occluded Pedestrian Detection

Abstract

Pedestrian detection has made significant progress in both academic and industrial fields. However, there are still some challenging questions with regard to occlusion scene. In this paper, we propose a novel and robust visible region-guided network (VRGNet) to specially improve the occluded pedestrian detection performance. Specifically, we leverage the adapted FPN-based framework to extract multi-scale features, and fuse them together to encode more precision localization and semantic information. In addition, we construct a pedestrian part pool that covers almost all the scale of different occluded body regions. Meanwhile, we propose a new occlusion handling strategy by elaborately integrating the prior knowledge of different visible body regions with visibility prediction into the detection framework to deal with pedestrians with different degree of occlusion. The extensive experiments demonstrate that our VRGNet achieves a leading performance under different evaluation settings on Caltech-USA dataset, especially for occluded pedestrians. In addition, it also achieves a competitive of 48.4%, 9.3%, 6.7% under the Heavy, Partial and Bare settings respectively on CityPersons dataset compared with other state-of-the-art pedestrian detection algorithms, while keeping a better speed-accuracy trade-off.