【摘要】 在行人目标检测中，小尺度、低像素的行人目标检测和行人遮挡等问题是模型训练的难点。针对深度学习Faster R-CNN网络对小尺度行人以及遮挡行人目标检测效果较差的情况，提出一种基于soft-NMS、GIoU和多尺度训练方法的改进型Faster R-CNN行人目标检测模型。在该改进模型中，Soft-NMS缓解行人密集检测中的因遮挡导致的漏检情况，GIoU对网络的损失计算进行改进，提升网络检测效果，多尺度训练方式能够提升小尺度与低像素的行人目标检测的准确率。仿真实验结果验证了该方法的有效性，改进的Faster R-CNN模型在Caltech行人数据集上的检测精度由64.2%提升到了70.4%，且在漏检率和误检率上都优于传统Faster R-CNN模型。更多还原

【Abstract】 In pedestrian target detection,small-scale and low-pixel pedestrian target detection and pedestrian occlusion are the difficulties of model training. To address the poor detection effect of the deep learning Faster R-CNN network on small-scale and occluded pedestrian targets,an improved Faster R-CNN pedestrian target detection model based on soft-NMS,GIoU and multi-scale training methods is proposed. In this improved model,Soft-NMS alleviates missed detection caused by occlusion in dense pedestrian detection.GIoU improves the loss calculation of the network and improves the network detection effect. The multi-scale training method can improve the accuracy of small-scale and low-pixel pedestrian target detection. Simulation results verify the effectiveness of the method.The detection accuracy of the improved Faster R-CNN model on the Caltech pedestrian data set increased from 64.2% to 70.4%,and both missed detection rate and false detection rate were better than those of the traditional Faster R-CNN model.更多还原