改进YOLO V5的密集行人检测算法研究

Research on intensive pedestrian detection algorithm based on improved YOLO V5

针对在人员密集区或相互拥挤场景下进行的行人目标检测时,因行人遮挡或人像交叠所导致的跟踪目标丢失、检测识别率低的问题,提出了一种融合注意力机制的改进YOLO V5算法。通过引入注意力机制来深入挖掘特征通道间关系和特征图空间信息,进一步增强了对行人目标可视区域的特征提取。为提高模型的收敛能力,利用CIoU、DIoU_NMS代替YOLO V5的原有损失函数优化anchor的回归预测,降低了网络的训练难度,提升了遮挡情况下的检测率;同时,结合数据增强及标签平滑算法进一步提高了特征模型的泛化能力和分类器性能。相比于一般的YOLO V5算法,论文所提出的改进算法在人员密集区或相互拥挤场景下进行行人检测时,具有更高的准确率和更低的漏检率,同时保持了原有算法的实时性。

In order to solve the problem of tracking target loss and low detection and recognition rate caused by pedestrian occlusion or portrait overlap in pedestrian target detection in a crowded area or scene,an improved YOLO V5 algorithm integrating attention mechanism was proposed.Firstly,the relationship between feature channels and spatial information of feature map was deeply mined by introducing attention mechanism to further enhance feature extraction of pedestrian target visual area.Secondly,in order to improve the convergence ability of the model,CIoU and DIoU_NMS were used to replace the original loss function of YOLO V5 to optimize the regression prediction of anchor,which reduces the training difficulty of the network and improves the detection rate under occlusion.Meanwhile,the generalization ability and classifier performance of feature model were further improved by combining data enhancement and label smoothing algorithm.Compared with the general YOLO V5 algorithm,the improved algorithm proposed in this paper has higher accuracy and lower missing rate in pedestrian detection in crowded areas or scenes,while the real-time performance of the original algorithm is still maintained.