Towards accurate dense pedestrian detection via occlusion-prediction aware label assignment and hierarchical-NMS

Abstract

With the development of self-driving technology, pedestrian detection models are becoming more useful and important. Although object detection models based on Convolutional Neural Networks have achieved favorable detection results, due to the occlusion in crowded scenarios, dense pedestrian detection is still a challenging task compared with generic object detection. For a CNN-based object detector, an appropriate label assignment method is significant. In this paper, we propose an effective training sample allocation method called Occlusion-Prediction aware Label Assignment (OPLA) to improve the performance of the detector in crowded scenarios. For those Ground Truths (GTs) without severe occlusion, we directly assign prediction boxes with the best matching quality as the corresponding positive samples. In terms of GTs in heavily crowded areas, we fine-tune the matching quality based on visible ratio to enhance the training sample selection priority of severe occluded GTs. In addition, we reveal the deficiency of currently popular post-processing algorithms, Non-Maximum Suppression (NMS) and Soft-NMS for dense pedestrian detection, and propose Hierarchical-NMS (H-NMS) whose strategy diverges with different Intersection over Union (IoU) levels. Experimental results on Crowdhuman, the largest dense pedestrian detection dataset at present, show our approach optimizes the Average Precision (AP) and Miss Rate (MR) of Fully Convolutional One Stage Object Detection (FCOS) by 4.05% and 5.35% and outperforms many state-of-the-art detectors, respectively. Besides, our method does not bring any increase in calculation during inference.