基于深度学习的智轨信号灯检测和识别方法
Traffic light detection and recognition based on deep learning for autonomous-rail rapid tram
熊群芳 1林军 1袁希文 1徐阳翰 1岳伟 1李源征宇1
作者信息
- 1. 中车株洲电力机车研究所有限公司,湖南 株洲 412001
- 折叠
摘要
智轨电车作为中车株洲电力机车研究所有限公司自主研发的智能轨道快运系统(简称"智轨"),其交通信号灯的检测与识别是提升智轨自动驾驶系统安全性的关键技术.智轨交通信号灯除了少部分通用信号灯,绝大多数为定制信号灯,而目前已有的信号灯检测与识别方法无法满足智轨自动驾驶环境下的检测要求.因此,文章利用深度学习算法对智轨信号灯检测与识别开展了相关研究工作,首先,通过高精度地图信息确定信号灯的RoI(Region of Interest)区域,缩小对智轨信号灯检测的范围,提升检测速度;其次,采用改进的YOLOV5s网络对RoI区域进行特征提取,检测出智轨交通信号灯;最后,对提取的交通信号灯图片采用MobileNetV2轻量级网络进行识别分类,确定信号灯的具体类别.为进一步增强模型的泛化性能,在信号灯检测之前增加了图像诊断算法,针对曝光、逆光等复杂环境及时提示预警,同时保存这些非正常数据,用于信号灯的检测与分类模型训练,进一步优化模型.试验结果表明,文章提出的方法对于智轨信号灯的检测与识别取得较好效果,白天在指定道路上平均精度均值达到84.76%,且实时性能良好.
Abstract
The autonomous rail rapid transit(ART)system is an intelligent express transport solution developed independently by CRRC Zhuzhou institute co.,ltd.The detection and recognition of traffic lights are critical for improving the safety of the ART automatic operation system.However,existing means for traffic light detection and recognition often fall short of meeting the detection require-ments in the automatic operation environments characterized by customized traffic lights,with only a few regular signals.This paper presents study efforts in this field through the application of a deep learning algorithm.Firstly,regions of interest(RoIs)for traffic lights were determined using high-precision map information to narrow the detection range and improve the detection speed.Secondly,an im-proved YOLOV5s network was employed to extract features from the RoIs,facilitating the recognition of ART traffic lights.Finally,the extracted traffic signal images were classified using a MobileNetV2 lightweight network to identify specific signal categories.In order to further enhance the model's generalization performance,an image diagnosis algorithm was introduced before signal recognition,which generates warnings for complex conditions,such as overexposure and backlighting.These abnormal image data were saved and utilized for further training and optimizing the model.Experimental results reveals that the proposed approach effectively detected and recog-nized ART traffic lights,achieving an average detection precision of 84.76%on designated roads during the daytime while exhibiting good real-time performance.
关键词
深度学习/智轨交通信号灯/高精度地图/YOLOV5s/MobileNetV2/图像诊断Key words
deep learning/ART traffic light/high-precision map/YOLOV5s/MobileNetV2/image diagnosis引用本文复制引用
出版年
2024
NETL
NSTL
万方数据