融合Transformer和改进PANet的YOLOv5s交通标志检测  

Fusion Transformer and Improved PANet for YOLOv5s Traffic Sign Detection

作　　者：张倩 刘紫燕 陈运雷[1] 吴应雨 郑旭晖 ZHANG Qian;LIU Ziyan;CHEN Yunlei;WU Yingyu;ZHENG Xuhui(College of Big Data and Information Engineering,Guizhou University,Guiyang Guizhou 550025,China;State Key Laboratory of Public Big Data,Guizhou University,Guiyang Guizhou 550025,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵州贵阳550025 [2]贵州大学公共大数据国家重点实验室,贵州贵阳550025 [3]中国科学院计算机技术研究所,北京100190

出　　处：《传感技术学报》2023年第2期232-241,共10页Chinese Journal of Sensors and Actuators

基　　金：贵州省科学技术基金资助项目(黔科合基础[2016]1054);贵州省联合资金资助项目(黔科合LH字[2017]7226号);贵州大学2017年度学术新苗培养及创新探索专项项目(黔科合平台人才[2017]5788)。

摘　　要：针对交通标志检测速度慢和目标大小与类别极度不平衡等问题,提出一种融合Transformer和改进PANet网络的YOLOv5s交通标志检测算法。首先在不增加模型复杂度的前提下,将主干网络末端与Transformer融合以提高网络特征提取能力;其次由于所采用交通标志数据集的目标尺度太小,导致网络32倍大尺度检测层检测效果不佳,故不采用相关网络层,同时采用K-means算法得出适合的预测候选框;然后改进损失函数以解决正负样本极度不平衡问题。最后将所提出的改进算法在Jetson AGX Xavier平台上部署验证。实验结果表明,所提算法检测性能更佳,其准确率和召回率在原网络的基础上分别提高了2.2%和0.7%,模型参数量和计算复杂度分别减少了25.8%和10.1%。在Xavier上的检测速度达到76FPS,满足实时交通标志检测的要求且易于在实际场景部署。Aiming at issues of low speed of traffic sign detection and extreme imbalance of target size and category,a YOLOv5s traffic sign detection algorithm of fused Transformer and improved PANet network is proposed.Firstly,without increasing model complexity,the end of backbone network is fused with Transformer to improve network feature extraction capability.Secondly,the small size of target scale in traffic sign datasets causes poor detection effects of the scale detection layers with a scale of 32 times larger,thus correlation network layers is not used,and K-means algorithm is adopted to yield appropriate prediction candidate frames.Then the loss functions is improved to address extreme imbalance problems between positive and negative samples.Finally,the proposed algorithm is applied on Jetson AGX Xavier platform for validation.Experimental results demonstrate that the proposed algorithm achieves better performance,the detection precision and recall rate improve 2.2%and 0.7%,respectively,and the number of model parameters and computational complexity reduce 25.8%and 10.1%respectively in comparison with the original network.The detection speed of 76 FPS on Xavier meets the requirements for real-time traffic sign detection,which is easily deployed in real scenarios.

关 键 词：交通标志检测 Jetson AGX Xavier TRANSFORMER PANet YOLOv5s 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]