【摘要】 遮挡是导致行人检测任务中漏检情况发生的主要原因之一，影响行人检测器的性能。为了增强检测器对遮挡行人的检测能力，论文提出一种改进的Faster R-CNN检测器，采用HRNet作为Faster R-CNN的特征提取网络，用于提取强语义的特征。在模型的训练和测试阶段，分别引入NMS-Loss和Soft-NMS，减少拥挤场景中由非极大值抑制算法（NMS）造成的漏检。此外，使用CrowdHuman行人数据集进行预训练，利用其中丰富的遮挡样本，增强Faster R-CNN检测器对遮挡行人目标的检测能力。在Caltech数据集上对本文提出的改进方法和其他对比方法进行了性能评估。实验结果表明，本文提出的改进方法在总体漏检率上具有优势，其中严重遮挡行人目标上的对数平均漏检率为29%，明显优于其他对比深度学习检测器。更多还原

【Abstract】 Occlusion is one of the main causes of missed detections in pedestrian detection tasks,affecting the performance of pedestrian detectors. To strengthen the detector’s ability to identify occluded pedestrians,we proposes an improved Faster R-CNN detector that employs HRNet as the feature extraction network for Faster R-CNN to extract strong semantic features. During the training and testing phases,NMS-Loss and Soft-NMS are introduced respectively to reduce the number of missed detections caused by the non-maximum suppression(NMS)algorithm in crowded scenes. Additionally,the CrowdHuman dataset is used for pre-training to leverage its rich sample of obstructed instances,thereby enhancing the occluded pedestrian detection capabilities of the Faster R-CNN detector. The proposed method and other comparative methods are evaluated on the Caltech dataset. Experimental results demonstrate that the proposed method has advantages in overall missed detection rates,with a logarithmic average missed detection rate of 29% for severely occluded pedestrian targets,significantly outperforming other comparative deep learning detectors.更多还原