摘要
针对夜间光照条件不足等条件下交通环境的多目标检测问题,提出一种改进YOLOv5s的目标检测算法。该算法首先在原始的YOLOv5s网络中嵌入三分支结构并行卷积注意力模块,通过计算跨维度注意力权值矩阵,实现了一种轻量级的有效注意力机制。其次,为了解决小目标和遮挡目标的检测问题,嵌入残差遮挡感知注意力机制,通过不同卷积核大小的卷积块对图像进行类分块操作,更准确地突显小目标和被遮挡目标。通过在FLIR数据集上的对比实验表明,改进算法在夜间交通环境下的多目标检测任务中能够提高检测精度,相较于传统YOLOv5s,其检测准确率mAP@0.5提高2.9%。
A modified YOLOv5s object detection algorithm is proposed to address the problem of multi-target detection in traffic environments under conditions such as insufficient nighttime lighting.This algorithm first embeds a three branch parallel convolutional attention module into the original YOLOv5s network,and achieves a lightweight and effective attention mechanism by calculating the cross dimensional attention weight matrix.Secondly,in order to solve the detection problem of small and occluded targets,a residual occlusion perception attention mechanism is embedded.The image is segmented into different convolution blocks with different kernel sizes to more accurately highlight small and occluded targets.Comparative experiments on the FLIR dataset show that this improved algorithm can improve detection accuracy in multi-object detection tasks in nighttime traffic environments,and its detection accuracy is higher than that of traditional YOLOv5s map@.5 Increase by 2.9%.
作者
刘丽伟
王玲
戚星烁
LIU Liwei;WANG Ling;QI Xingshuo(School of Computer Science&.Engineering,Changchun University of Technology,Changchun 130102,China)
出处
《长春工业大学学报》
CAS
2024年第5期428-436,共9页
Journal of Changchun University of Technology
基金
吉林省科技发展计划项目(20220201099GX)
吉林省预算内创新动力建设资金项目(2022C047-7)。
