基于深度学习的车道线检测算法的研究
Research on lane detection algorithm based on deep learning
柯红梅 1徐远2
作者信息
- 1. 电子科技大学成都学院计算机学院,四川 成都 611731
- 2. 成都产品质量检验研究院有限责任公司国家移动互联网软件产品质量检验检测中心,四川 成都 610100
- 折叠
摘要
近年来,随着软硬件设备的快速升级,智能设备得到了快速发展,智能辅助驾驶技术逐渐落地.而智能辅助驾驶要解决的关键问题之一是车道线、车辆以及行人三者之间的矛盾,因此对智能辅助驾驶系统中的车道线检测展开研究.针对车道线检测算法,采用了不规则编码器-解码器的网络结构以及多任务学习的思想,对车道线检测算法LaneNet进行改进以及对卷积模块进行优化,解决了复杂场景难以检测以及模型精度低的问题.经过大量数据训练和实测数据集的验证,车道线检测在Tusimple数据集上的准确率为96.42%,参数量为5.14 M.测试结果表明:车道线检测效果较好,满足车道线检测的研究目标.
Abstract
In recent years,with the rapid upgrading of software and hardware devices,intelligent devices have developed rapidly.Intelligent assisted driving technology is gradually being implemented.One of the key issues that intelligent assisted driving needs to solve is the contradiction between lane lines,vehicles,and pedestrians.The following research has been conducted on lane line detection in intelligent assisted driving systems.For the lane detection algorithm,an irregular encoder-decoder network structure and the idea of multi-task learning were adopted to improve the lane detection algorithm LaneNet and optimize the convolution module,solving the problems of difficult detection in complex scenes and low model accuracy.The method used in the article has been trained on a large amount of data and validated on actual datasets.The accuracy of lane detection on the Tusimple dataset is 96.42%,with a parameter size of 5.14 M.The test results indicate that lane detection has a good effect and meets the research objectives of this article.
关键词
车道线检测/辅助智能驾驶/深度学习Key words
lane detection/assisted intelligent driving/deep learning引用本文复制引用
出版年
2024
NETL
NSTL
万方数据