A Pedestrian Detection Method Based on Decoupling Prediction with Counting and Occlusion Locating in Crowds
In crowded scenes,occlusion causes partial semantic loss of pedestrians,and traditional Non-Maximum Suppression(NMS)algorithm struggles with highly overlapped detection boxes,limiting the ef-fectiveness of proposal-based pedestrian detectors.To address it,we propose a prediction decoupling module.It is trained with separate branches for predicting full-body and visible-body boxes,enhancing the network's understanding of visible-body features.Additionally,we introduce a strategy for assigning positive and negative samples based on comprehensive annotations of visible-body boxes and full-body boxes,guiding the network to effectively regress full-body boxes from visible-body features.Furthermore,we propose a Counting and Locating based NMS strategy.By utilizing the local counting branch and the occlusion-aware locating branch,we can obtain local counting and occlusion localization,thereby adjus-ting the confidence of full-body boxes.Experiments on the CrowdHuman validation set within the Cascade R-CNN framework demonstrates that we achieves 3.8%AP gains,0.9%MR-2 gains,2.5%JI gains,il-lustrating the advancement of our method.
prediction decoupling modulecounting and occlusion Locatingpedestrian detectionvisible regions
万方数据
维普
