自动驾驶场景中基于改进YOLOX-NANO模型的人行横道实时检测  

Real-time Detection of Crosswalks in Autonomous Driving Scenarios based on YOLOX and Hard Sample Mining Strategy

作　　者：崔蕾[1] 陈易平 CUI Lei;CHEN Yi-ping(Department of Information Engineering,Yantai Vocational College,Shandong,Yantai 264670;School of Computer Science and Engineering,Guangzhou Institute of Technology,Guangdong,Guangzhou 510540)

机构地区：[1]烟台职业学院信息工程系,山东烟台264670 [2]广州理工学院计算机科学与工程学院,广东广州510540

出　　处：《贵阳学院学报(自然科学版)》2024年第4期74-82,共9页Journal of Guiyang University:Natural Sciences

基　　金：2022年山东省职业教育教学改革研究项目“新职教背景下数字化赋能高职专业建设转型升级路径研究与实践”(项目编号:2022092)。

摘　　要：提出了基于改进YOLOX-NANO的人行横道检测模型,对训练数据依次应用不同的数据增广策略,并通过基于性能梯度分析选择合适的增广策略。使用改进的YOLOX-NANO模型执行目标检测任务,利用注意力特征增强模块改善有用特征的提取性能,并使用EIoU损失函数提高模型鲁棒性。在模型训练过程中,提取出难分正负样本,利用迁移学习技术进行微调。公开数据集的实验结果表明,所提方法在嵌入式平台Jetson nano上的处理速度达到35.7 FPS,人行横道检测的平均精度(AP)值达到了96.7%,性能优于其他先进方法,能够显著提升各种自动驾驶场景中人行横道的检测性能,有助于提高自动驾驶应用的安全性。nighttime and partial occlusions in realistic complex scenes must be considered.To this end,a crosswalk detection scheme based on improved YOLOX-NANO is proposed.Different data augmentation strategies are sequentially applied to the training data,and appropriate augmentation strategies are selected based on performance gradient analysis.The object detection task is performed using the modified YOLOX-NANO model,in which the attention feature enhancement module is utilized to improve the extraction performance of useful features,and the EIoU loss function is used to improve the model robustness.During the model training process,hard positive and negative samples are extracted,and the model is fine-tuned through transfer learning technology.The experimental results on the public dataset show that the processing speed of the proposed method on the embedded platform Jetson Nano reaches 35.7 FPS,and the average accuracy(AP)value of crosswalk detection reaches 96.7%,which is superior to other advance methods being compared.The proposed method can significantly improve the detection performance of crosswalks in various autonomous driving scenarios,which helps to improve the safety of autonomous driving applications.

关 键 词：自动驾驶 人行横道检测 智能交通管理系统 YOLOX-NANO 难分样本挖掘 

分 类 号：U461[机械工程—车辆工程] TP391[交通运输工程—载运工具运用工程]