基于自注意力机制的车道线检测算法

Lane Detection Algorithm based on self-attention mechanism

环境感知是自动驾驶中获取外界信息的关键核心技术,而车道线检测则是环境感知中重要的基础任务之一。但由于车道线过于细长,其较大的空间跨度使得车道线的检测易于受到遮挡、光照变化和自然侵蚀褪色等复杂环境的干扰。目前的车道线检测一般都包括车道线空间位置的定位和后处理算法两部分组成,如SCNN、LaneNet、H-Net等。但是CNN的感受野大小远远小于理论值,较小的感受野无法将重要信息充分融入到网络之中。文章针对车道线的特征提出带有自注意力的语义分割网络模型和相应的后处理算法,通过引入的自注意力机制模块可以更好的获取丰富的上下文信息,从而有效提取更具代表性的全局特征。(同时考虑到检测的实时性,简化后处理算法的时间复杂度)。通过实验结果表明,使用结合自注意力机制的车道线检测算法,可以在几乎不增加额外计算量的情况下取得更高的准确率。

Environment perception is the key technology of obtaining the outside information in automatic driving,and lane detection is one of the important basic tasks in environment perception.But because the lane is too long and thin,the detection of lane is easily disturbed by the complex environment such as occlusion,illumination change and natural erosion fading.At present,lane detection generally includes two parts:Lane location and post-processing Algorithm,such as Scnn,LaneNet,H-Net,etc.However,the CNN receptive field is much smaller than the theoretical value,and the smaller receptive field can not fully integrate the important information into the network.We propose a semantic segmentation network model with self-attention and a corresponding post-processing algorithm based on lane features.By introducing self-attention mechanism module,rich contextual information can be obtained better,thus,more representative global features can be extracted effectively.At the same time,we consider the real-time performance of the detection and simplify the time complexity of the post-processing Algorithm.The experimental results show that the lane detection algorithm combined with self-attention mechanism can achieve higher accuracy with almost no additional computation.