【摘要】 基于深度卷积神经网络的人物检测方法是目前检测效果最好的方法。在同等环境下，YOLOv3运行速度最快，但其采用的非极大值抑制算法(NMS)导致很多正确的检测框被错误移除。通过加入取回算法来恢复被NMS错误移除掉的人物检测框，而且将NMS替换为Soft-NMS进一步提高了准确率。在PASCAL VOC数据集上的实验表明，使用Soft-NMS和取回算法改进的YOLOv3相比于原算法提升了大约3.1百分点的准确率，同时运行速度没有发生太多的变化。更多还原

【Abstract】 Human detection algorithm based on deep convolutional neural network is the most effective method at present. Under the same circumstances, YOLOv3 ran the fastest, but its None Max Suppression(NMS) algorithm caused many correct detection frames to be removed by mistake. The incorrectly removed human bounding boxes could be recovered by using the get back algorithm. NMS was replaced by Soft-NMS to further improve the accuracy. Experiments on the PASCAL VOC dataset show that compared with the former model, the accuracy of the improved YOLOv3 increases by about 3.1 percentage points and its speed does not change too much.更多还原