面向弱光交通场景的YOLOv7道路标志检测算法优化  

Optimization of YOLOv7 Road Sign Detection Algorithm for Low-Light Traffic Scenes

作　　者：孙亭 杨洁[1] 李家璇 王耀宗 SUN Ting;YANG Jie;LI Jiaxuan;WANG Yaozong(School of Mechanics and Transportation,Southwest Forestry University,Kunming 650224,Yunnan,China)

机构地区：[1]西南林业大学机械与交通学院,云南昆明650224

出　　处：《计算机工程》2025年第3期342-351,共10页Computer Engineering

基　　金：云南省教育厅重点基金项目(2023J0711);农业推广理论与实践案例库的建设(503210305);农林研究生教育中产教融合和科教融合的探索(503210401)。

摘　　要：针对交通标志检测算法在黑夜及弱光条件下存在检测精度不高、漏检等问题,提出一种改进YOLOv7的交通标志检测算法。构建用于弱光增强的高斯图像滤波器,抑制其背景噪声,对图像实现像素增强。在YOLOv7网络中,构建新的AC-ResBlock残差模块来替代ELAN中的3×3卷积模块,以提高交通标志的特征提取能力和网络推理速度。引入SIoU损失函数提高模型的准确度,加速训练过程收敛。采用K-means++算法代替K-means重新标定锚框的尺寸,在扩展后的中国交通标志检测数据集CCTSDB上的实验结果表明,改进后的YOLOv7算法准确率达到95.7%,召回率达到94.8%,平均精度达到96.3%,优于YOLOv8、YOLOv5及其他主流检测算法,可以实现黑夜及弱光条件下的交通标志检测。对于复杂环境下的交通标志检测具有一定的研究意义。This paper proposes an improved YOLOv7 traffic sign detection algorithm to address the issues of low detection accuracy and missed detections under dark and low-light conditions by existing algorithms.A Gaussian image filter for low-light enhancement is constructed to suppress background noise and enhance the image pixels.In the YOLOv7 network,a new AC-ResBlock residual module has replaced the 3×3 convolution module in Efficient Layer Aggregation Network(ELAN),thereby enhancing the feature extraction capability and network inference speed for traffic signs.The Scylla-Intersection over Union(SIoU)loss function is introduced to improve model accuracy and accelerate training convergence.The K-means++algorithm is used instead of K-means to recalibrate the anchor box dimensions.Experiments on the expanded Chinese Traffic Sign Detection Benchmark(CCTSDB)have shown that the improved YOLOv7 algorithm achieves a accuracy of 95.7%,recall rate of 94.8%,and average accuracy of 96.3%.This performance surpasses those of YOLOv8,YOLOv5,and other mainstream detection algorithms,enabling the detection of traffic signs under night and low-light conditions,which is a significant advancement for traffic sign detection in complex environments.

关 键 词：交通标志检测 YOLOv7算法 黑夜图像增强 自注意力机制 损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]