基于注意力网络的车道线检测算法研究

Lane Detection Algorithm Research based on Attention Network

随着智能网联汽车的发展,越来越多的学者投身于稳定的L4级以上的自动驾驶算法研究中。车道线检测在智能网联汽车领域具有极其重要的地位,它是实现车辆自动驾驶、保障行车安全的关键技术。由于车道线的变化可以反映道路条件和驾驶路径的信息,因此车道线的准确检测与识别对于提高自动驾驶汽车的导航精度、预防车辆偏离车道或发生交通事故具有重大的研究意义。在众多传统的车道线检测算法中,常常面临诸如阴影、反光、复杂背景和实时性低等问题,对这些问题的处理往往依赖于预处理和手动特征提取,对各种不同情况的适应性较差。本文提出了基于注意力机制的深度车道检测模型,并引入锚点进行特征池化,从而整合全局特征,使得该模型检测精度和实时性得到有效提升。所提出的算法部署在某品牌的线控改装车上,在真实道路上得到了优异效果。

With the progression of intelligent connected vehicles,an increasing number of scholars have dedicated their efforts to the research of stable autonomous driving algorithms,particularly those above level 4.Lane detection plays a pivotal role in this domain,serving as a key technology to enable autonomous vehicle operation and ensure road safety.The variation in lane lines reflects information on road conditions and driving paths;hence,accurate detection and identification of these lines are crucial for improving the navigation accuracy of autonomous vehicles,preventing lane departure,and averting traffic accidents.Traditional lane detection algorithms often face issues such as shadows,glare,complex backgrounds,and low real-time performance.The handling of these issues generally relies on pre-processing and manual feature extraction,which have limited adaptability to diverse scenarios.This paper proposes a deep lane detection model based on an attention mechanism and introduces anchors for feature pooling,thereby integrating global features,effectively enhancing detection accuracy and real-time performance.The proposed algorithm,when deployed on a steer-by-wire retrofit car based on passenger car,yielded excellent results on real roads.