Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (2): 264-271.DOI: 10.3778/j.issn.1002-8331.2210-0166

• Graphics and Image Processing • Previous Articles     Next Articles

Lane Detection Method Based on Improved Multi-Head Self-Attention

GE Zekun, TAO Fazhan, FU Zhumu, SONG Shuzhong   

  1. 1.College of Information Engineering, Henan University of Science and Technology, Luoyang, Henan 471023, China
    2.Henan Key Laboratory of Robot and Intelligent Systems, Henan University of Science and Technology, Luoyang, Henan 471023, China
  • Online:2024-01-15 Published:2024-01-15

改进多头注意力机制的车道检测方法

葛泽坤,陶发展,付主木,宋书中   

  1. 1.河南科技大学 信息工程学院,河南 洛阳 471023
    2.河南科技大学 河南省机器人与智能系统重点实验室,河南 洛阳 471023

Abstract: In order to solve the problems of low efficiency of network and weak modeling ability for slender structures in lane detection methods based on convolution neural network (CNN), a lightweight lane detection method based on improved multi-head self-attention (MHSA) is proposed. Firstly, the MHSA is introduced to integrate Fuse MBConv, MBConv and feature compression module to reduce the parameters of the model. At the same time, the context information embedding module is used to establish a global attention network that gives consideration to both detection accuracy and reasoning speed. Then, the lane structures are parameterized by using the encoder-decoder of Transformer and the feedforward feedback network, and the Hungarian fitting loss function is combined to improve the modeling ability of the proposed method for the slender structure of lane lines. Finally, the proposed method is verified on the TuSimple. The results show that accuracy of the method reaches 96.3% and the inference speed reaches 95 FPS. At the same time, it runs at 60 FPS in Apollo autonomous driving platform, which can meet the requirements of real-time detection.

Key words: multi-head self-attention, global context, lightweight lane detection method, autonomous driving platform

摘要: 针对基于卷积神经网络(convolution neural network,CNN)的车道线检测方法存在的网络处理效率低和对车道线细长结构的建模能力不佳的问题,提出一种基于改进多头注意力机制(multi-head self-attention,MHSA)的轻量级车道检测方法。引入MHSA,融合Fuse MBConv、MBConv模块与特征压缩模块,降低模型的参数,同时利用上下文信息嵌入模块,建立兼顾检测精度和推理速度的全局注意力网络;利用Transformer的编码和解码器以及前向反馈网络将车道线参数化,结合匈牙利拟合损失函数提高所提出方法对车道线细长结构的建模能力。在TuSimple数据集对所提出的方法进行验证,结果表明,所提出的方法识别精度达到96.3%,推理速度达到95帧/s,同时在Apollo无人驾驶平台上的运行速度达到60帧/s,能够满足实时检测的要求。

关键词: 多头注意力机制, 上下文信息, 轻量级车道检测方法, 无人驾驶平台