基于改进YOLOv5s模型的车辆及行人检测方法

Vehicle and Pedestrian Detection Method Based on Improved YOLOv5s Model

针对道路交通环境中车辆及行人目标较小或被遮挡造成的检测精度低以及误检、漏检问题,提出一种基于改进YOLOv5s模型的车辆及行人目标检测方法。针对小目标和遮挡目标,引入SIoU边界框损失函数,增加小目标检测层,增强对小尺度特征的获取;改进特征金字塔结构,增加横向特征图传递,并使用CSP stage替换C3_F特征提取网络,使其获得更多的语义信息和图形信息;改进后处理NMS算法,优化冗余边界框剔除方法,筛选出高质量检测结果。试验结果表明:改进YOLOv5s模型算法的Precision、Recall、mAP@0.5和mAP@0.5:0.95指标均优于Faster-RCNN、YOLOv3-tiny和YOLOv8s算法,与原YOLOv5s模型算法相比Precision下降了0.4%,但Recall、mAP@0.5和mAP@0.5:0.95提高了3.4%、2.1%和6.0%,分别达到了86.1%、92.9%和70.0%,对小目标和遮挡目标的检测效果明显提高,证明此改进方法有效解决了对小目标和遮挡目标检测精度低以及误检、漏检问题。

A vehicle and pedestrian detection method based on improved YOLOv5s model is proposed to address the low detection accuracy,false alarms,and missed detection caused by small or occluded targets in the road traffic.For small and occluded targets,the SIoU bounding box loss function is introduced,a small target detection layer is added to enhance the acquisition of small-scale features.The feature pyramid structure is improved by adding lateral feature map connections and using CSP stage instead of C3_F feature extraction network,enabling the model to capture more semantic and spatial information.The post-processing NMS algorithm is modified to optimize the removal of redundant bounding boxes and select high-quality detection results.Experimental results show that the improved YOLOv5s model algorithm outperforms Faster-RCNN,YOLOv3-tiny and YOLOv8s algorithms in terms of Precision,Recall,mAP@0.5,and mAP@0.5:0.95 metrics.Compared to the original YOLOv5s model algorithm,Precision metric decreases by 0.4%,but Recall,mAP@0.5 and mAP@0.5:0.95 metrics improve by 3.4%,2.1%and 6.0%,reaching 86.1%,92.9%and 70.0%respectively.The detection performance for small and occluded targets significantly improved,demonstrating that this improvement method effectively addresses the issues of low detection accuracy,false positives and missed detection for small and occluded targets.