改进YOLOv7的电动自行车违章行为检测算法
Electric Bicycle Illegal Behavior Detection Based on Improved YOLOv7
林雪勤 1潘立琼2
作者信息
- 1. 安徽三联学院 工学部,安徽 合肥 230601
- 2. 安徽三联学院 智慧交通现代产业学院,安徽 合肥 230601
- 折叠
摘要
针对电动自行车密度大和容易互相遮挡而导致违章行为难判断及难定位的问题,构建了融合协调置换注意力机制,并优化位置损失函数的YOLOv7-CSA-DSIoU算法.通过将通道维度分成多个子特征,并将每个子特征的通道注意力模块分解为两个并行的一维特征编码模块,用于捕捉空间方向的长程依赖和位置信息来加强网络特征提取能力.另外,改进基于斯库拉交并比损失函数的位置回归损失函数,加入预测框和目标框的顶点距离和中心点间距离的度量,实现点边双重快速收敛并提高目标检测的定位能力.实验结果表明:相对于基准模型,YOLOv7-CSA-DSIoU算法平均精度均值mAP@0.5 提升了 6.4%,mAP@0.75 提升了 4.2%,mAP@0.5:0.95 提升了 4.3%.
Abstract
To address the issues of high electric vehicle density and difficulty in violation detection and location due to overlapping,the YOLOv7-CSA-DSIoU algorithm,which integrates an attention mechanism and optimizes position loss function,is constructed.The network's ability for feature extraction is strengthened by dividing channel dimension into multiple sub-features,each sub-feature is decomposed into two parallel one-dimensional feature encoding modules for capturing the long-range dependence of spatial direction and location information.In addition,the location regression loss function based on the Scylla IoU is improved by adding the measures of vertex distance and inter-center distance between the prediction frame and the target frame to achieve the point-edge double fast convergence and improve the localization ability of target detection.The experimental results indicate that YOLOv7-CSA-DSIoU algorithm outperforms the baseline model,mAP@0.5 improved by 6.4%,mAP@0.75 improved by 4.2%,and mAP@0.5:0.95 improved by 4.3%.
关键词
电动自行车/违章行为/YOLO/IoU/注意力机制Key words
electric bicycle/illegal beharior/YOLO/IoU/attention mechanism引用本文复制引用
基金项目
安徽省普通高校交通信息与安全重点实验室开放课题(JTX202204)
安徽省省级重点自然科学基金项目(KJ2021A1191)
安徽省优青人才支持项目(gxyqzd2021140)
出版年
2024
NETL
NSTL
万方数据