软件导刊2025，Vol.24Issue(6)：175-184,10.DOI:10.11907/rjdk.241350
基于深度学习的复杂天气场景交通标志检测
Traffic Sign Detection in Complex Weather Conditions Based on Deep Learning
摘要
Abstract
In response to the problems of imbalanced distribution of target categories and single weather scenes in the TT100K traffic sign da-taset,as well as the decline in detection performance in complex weather scenes,three improvements are proposed based on the object detec-tion network CenteNet2.Firstly,a dataset extension method TT100K-sync is proposed based on sign replacement and synthesis images with depth map,which balances the number of traffic sign categories and adds two new weather scenarios:foggy and hazy images.Secondly,based on joint training,a U-shaped network structure U-subnet is constructed to enhance the performance of the feature extraction backbone when image quality decreases.Thirdly,a data preprocessing method based on image tiling called overlap-tiling is introduced to reduce the feature loss of the target during feature extraction and down sampling.The experimental results show that the expanded new dataset balances the maxi-mum difference in the number of traffic sign categories from 3 176 to 200 compared to the original dataset,and adds 11 330 fog and haze imag-es each.The network that adopted U-subnet and overlap-tiling improved the mAP50 on the conventional weather subset TT-normal,foggy weather subset TT-fog,and haze weather subset TT-haze in the new dataset by 1.13%,3.26%,and 6.38%,respectively,compared to the original network,significantly improving the network detection performance in complex weather conditions.关键词
交通标志检测/样本平衡/图像合成/联合训练/图像分块Key words
traffic sign detection/sample balance/image synthesis/joint training/image tiling分类
计算机与自动化引用本文复制引用
陆维,吴锡..基于深度学习的复杂天气场景交通标志检测[J].软件导刊,2025,24(6):175-184,10.基金项目
四川省科技计划项目(重点研发项目)(2023YFG0025) （重点研发项目）
