一种基于深度学习的车道线识别方法
A Lane Line Recognition Method Based on Deep Learning
史炎锦 1金文智 1李勇 1赵子豪 1高琪 1樊星男1
作者信息
- 1. 太原学院机电与车辆工程系,山西 太原 030032
- 折叠
摘要
基于深度学习和语义分割的车道线识别方法能够对车道线图片进行端到端的识别,能够适应复杂多变的车道环境.本文设计了一种基于深度学习和语义分割的车道线识别模型,该模型以 Segnet 为基础,由编码器和解码器两部分组成.编码器采用 4 级下采样结构,主要由卷积层和最大池化层组成,并将 PRelu 函数作为卷积层的激活函数,该函数能有效提高网络的拟合能力,并降低过拟合分险;解码器采用 4 级上采样结构,主要由上采样层、卷积层和批标准化层组成.为解决车道线图片中车道线和背景像素点数量严重不平衡的问题,使用加权交叉熵函数计算网络的损失值,并用MFB算法确定权值.最后,在tuSimple数据集上进行了验证,在大量实验的基础上,通过对交叉熵函数权值进行修正,获得了良好的识别效果和较高的鲁棒性.
Abstract
The lane line recognition method based on deep learning and semantic segmentation can recognize lane line images end-to-end and adapt to complex and ever-changing lane environments.This lane recognition model is based on deep learning and semantic segmentation,which is based on Segnet and consists of two parts:an encoder and a decoder.The encoder adopts a 4-level down sampling structure,which is mainly composed of the convolution layer and the maximum pooling layer,and uses the PRelu function as the Activation function of the convolution layer,which can effectively improve the fitting ability of the network and reduce the risk of over fitting;The decoder adopts a 4-level upsampling structure,mainly composed of upsampling layer,convolution layer,and batch standardization layer.In order to solve the problem that the number of lane lines and back-ground pixels in the lane line image is seriously unbalanced,the weighted Cross entropy function is used to calcu-late the loss value of the network,and the MFB algorithm is used to determine the weight value.Finally,valida-tion was conducted on the tuSimple dataset,and based on extensive experiments,good recognition performance and high robustness were achieved by modifying the weights.
关键词
车道线识别/语义分割/深度学习/卷积网络Key words
Lane line recognition/Semantic segmentation/Deep learning/CNN引用本文复制引用
基金项目
山西省教育厅2023年省级大学生创新创业训练资助项目(20231516)
出版年
2024
NETL
NSTL
万方数据