摘要
车道线检测任务包含道路磨损、阴影遮挡和弯道等困难样本,这些样本中线条信息均有不同程度的缺失,使检测结果出现漏检或误检现象。基于深度学习的检测方案通过卷积操作提取特征信息。卷积操作摒弃人工设计滤波器等一系列传统图像处理的繁琐操作,得益于权重共享和归纳偏置大大减少了特征提取的工作量。该操作在缩小图像分辨率的同时获取长距离的信息,导致小分辨率的特征图损失区域边缘等细节,影响检测结果的质量。深度学习中分割模型比检测模型处理的信息更细致,本文在分割模型的基础上引入Transformer改进采样方式,改善卷积操作在获取全局信息上的不足。模型改进后在Tusimple上测试准确率提高0.4%,像素精准度提高0.3,乘法累加运算量增加36.09 G。结果表明Transformer特有的采样方式可以改善卷积操作采样的不足,改善语义分割网络对车道线困难样本识别漏检的情况。
The task of lane line detection includes difficult samples such as road wear,shadow occlusion and curves.The line information in these samples can be missing with different levels,which results in missed or false detection of the detection results.The detection scheme based on deep learning extracts feature information through convolution operation.Convolution operation discards a series of tedious operations of traditional image processing,such as manually designing filters,and benefits from weight sharing and inductive bias,which greatly reduces the workload of feature extraction.This operation not only reduces the image resolution,but also obtains long-distance information,resulting in the loss of regional edge and other details of the small resolution feature map,which affects the quality of the detection results.In deep learning,the segmentation model processes more detailed information than the detection model.Based on the segmentation model,this paper introduces transformer to improve the sampling method and improve the lack of convolution operation in obtaining global information.After the model is improved,the test accuracy on Tusimple is improved by 0.4%,the pixel accuracy is improved by 0.3,and the amount of multiplication and accumulation operation is increased by 36.09 G.The results show that the transformer’s unique sampling method can improve the lack of convolution operation sampling,and improve the situation of missing detection of lane line difficult samples in semantic segmentation network.
作者
丁志江
李丹
马志程
张宝龙
Ding Zhijiang;Li Dan;Ma Zhicheng;Zhang Baolong(School of Electronics and Automation,Tianjin University of Science and Technology,Tianjin 300222,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2022年第10期227-234,共8页
Journal of Electronic Measurement and Instrumentation
作者简介
丁志江,2022年于天津科技大学获得硕士学位,现为苏州镁伽科技AI算法工程师。主要研究方向为测试计算技术与仪器与人工智能包括图像分割、关键点检测与目标检测。E-mai:dzj19960727@163.com;李丹,1999年于南开大学获得学士学位,2002年于南开大学获得硕士学位,2007年于香港大学获得博士学位,现为天津科技大学副教授,硕士生导师。主要研究方向为光电检测与计算机视觉检测。E-mail:lidan@tust.edu.cn;李丹,1999年于南开大学获得学士学位,2002年于南开大学获得硕士学位,2007年于香港大学获得博士学位,现为天津科技大学副教授,硕士生导师。主要研究方向为光电检测与计算机视觉检测。E-mail:lidan@tust.edu.cn;通信作者:张宝龙,1999年于南开大学获得学士学位,2001年于香港科技大学获得硕士学位,2006年于香港科技大学获得博士学位,现为天津科技大学教授,硕士生导师,主要研究方向为集成电路设计及半导体制备工艺的研究。E-mail:eezbl@tust.edu.cn
