Human-Cascaded network for Robust Detection of Occluded Pedestrian

Abstract

Occlusion is a key challenge in real-world on-road pedestrian detection task. Due to constrained viewpoint geometry, a pedestrian is very likely to be obstructed by other pedestrians and/or other objects such as cars and bicycles. For Advanced Driver Assistance System (ADAS), heavily occluded pedestrians are as important as reasonable pedestrians because they may burst out from crowds or roadside obstacles. In this paper, a human-cascaded network is proposed for robust detection of heavily occluded pedestrians. Specifically, a sharp-response proposal network (SRPN) is designed to refine the feature responses in a narrow area to handle crowded pedestrian detection, followed by outputting the head and full body proposals for diverse occlusion situations. After RoI-pooling, a visible-guided attention (VGA) module is developed to leverage the head and visible area information. The VGA module also suppresses the feature noise of occluded area to enhance the feature representation learning of the backbone network. Finally, a head-cascade RCNN (HRCNN) network is proposed to predict the pedestrian bounding box from the head proposal. The proposed approach is validated through a widely used pedestrian detection dataset: CityPersons. Experimental results show that our approach achieves promising detection performance (log-average miss rate, MR) improvement of 11.4% on heavy occlusion subset, compared to the baseline detector.