基于改进YOLOv5的交通监控视频车辆检测方法研究

Research on Vehicle Detection Method in Traffic Surveillance Video Based on Improved YOLOv5

针对交通监控视频中车辆遮挡、车辆目标偏小导致的错检、漏检问题,提出一种改进的YOLOv5网络模型。将注意力机制SE模块分别引入YOLOv5网络的Backbone、Neck、Head,经过试验对比得出SE模块引入的最佳位置。针对YOLOv5模型存在的正负样本不平衡问题,将焦点损失函数Focal Loss引入模型的训练过程。试验表明,将SE模块与Backbone进行融合,模型的平均准确率mAP提高了0.011%,引入Focal Loss之后,mAP提高了0.02%,同时准确率Precision和召回率Recall均有提高。对自建的数据集进行测试,结果表明:改进的YOLOv5模型能有效地提高交通监控场景中的车辆检测性能。

Aiming at the wrong detection and missed detection caused by vehicle occlusion and small vehicle targets in traffic monitoring video, an improved YOLOv5 model is proposed. The SE module of the attention mechanism was introduced into Backbone,Neck, and Head of the YOLOv5 respectively, and the best position of the SE module was obtained through experimental comparison.Aiming at the imbalance of positive and negative samples in the YOLOv5 model, the Focal Loss is introduced into the training process of the model. The experiments show that the integration of the SE module and Backbone improves mAP by 0.011%. After the introduction of Focal Loss, the mAP is increased by 0.02%. At the same time, both the precision rate and the recall rate are improved. The test results based on a self-built data set show that the improved YOLOv5 model effectively improves the vehicle detection performance in traffic monitoring scenarios.