基于优化DeepLab v3+的车道线缺损检测技术分析  

Analysis of Lane Line Defect Detection Technology Based on Optimized DeepLab v3+

作　　者：朱智键 刘宪国 宋炜桐 林思婷 ZHU Zhijian;LIU Xianguo;SONG Weitong;LIN Siting(Foshan University of Science and Technology,Guangdong 528000,China)

机构地区：[1]佛山科学技术学院,广东528000

出　　处：《电子技术（上海）》2024年第2期369-371,共3页Electronic Technology

基　　金：2022年度佛山科学技术学院学生学术基金立项项目(xsjj202203kjb12)。

摘　　要：阐述一种基于改进DeepLab v3+语义分割模型的车道线缺损检测方法。使用MobileNet v3网络替代原来的Xception网络来减少网络参数量。同时在主干特征提取后引入双注意力机制CBAM,通过最小矩形包围的方法检测出车道线缺损情况。使用自制数据集,将该数据集基于Pytorch对改进后的DeepLab v3+进行训练、验证和测试。结果表明,改进后的DeepLab v3+在MIoU和MPA上分别提高1.6%和1.3%,单幅图像分割时间7.4ms,与原模型对比减少16.9ms,可以满足车道线缺损检测的实时性和准确性。This paper describes a lane line defect detection method based on an improved DeepLab v3+semantic segmentation model.Replace the original Xception network with MobileNet v3 network to reduce the number of network parameters.At the same time,the dual attention mechanism CBAM is introduced after extracting the backbone features,and the lane defects are detected through the method of minimum rectangle bounding.Using a self-made dataset,train,validate,and test the improved DeepLab v3+based on Python.The results show that the improved DeepLab v3+improves MIoU and MPA by 1.6%and 1.3%respectively,with a single image segmentation time of 7.4ms,which is 16.9ms less than the original model.It can meet the real-time and accuracy requirements for lane line defect detection.

关 键 词：车道线缺损检测 语义分割 DeepLab v3+网络 双注意力机制CBAM 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] U491.31[自动化与计算机技术—控制科学与工程]