OSS-OCL: Occlusion Scenario Simulation and Occluded-edge Concentrated Learning for pedestrian detection

Pedestrian detection plays an important role in realistic applications. However, heavily occluded pedestrians, with their incomplete and unusual appearances, are easily missed during detection. To address this issue, previous works use copy-paste method or generate dummies to assist the detectors in learning better detection capability. Nevertheless, these works focus on less frequent occlusion in natural scenes and lead to less performance gain. Therefore, we firstly propose a novel method named Occlusion Scenario Simulation (OSS), which simulates the most classical occlusion scenario by inserting objects adjacent to the non- or partial- occluded pedestrians. Secondly, in order to supervise the detector to better learn the occlusion information, we also propose anew method namely Occluded-edge Concentrated Learning (OCL) to predict the offset of occluded-edge between pedestrians and occlusions. Extensive experiments on popular pedestrian datasets demonstrate that our proposed OSS-OCL outperforms some state-of-the-art methods, particularly in the cases of heavy occlusion.