基于改进卷积神经网络模型的车道线检测方法

Lane Detection Method Based on Improved Convolutional Neural Network Model

文章基于卷积神经网络的车道线检测方法,提出了一种优化的车道检测方案,利用残差网络与简化的Transformer优化神经网络,首先将Transformer模型用作编码器或解码器,学习输入序列之间的关系,提高神经网络的性能,然后在编码器中增加残差层,用以更好地处理边缘与相似信息并提取低分辨率特征,最后设计一个主干架构用于整合优化内容并对优化后的模型进行训练。实验结果表明,优化后的模型在TuSimple验证集中的预测一致性超过了90%,并对多种干扰状况表现出良好的适应性。

This paper proposes an optimized lane detection scheme based on convolutional neural networks,which utilizes residual networks and simplified Transformers to optimize the neural network.Firstly,the Transformer model is used as an encoder or decoder to learn the relationships between input sequences and improve the performance of the neural network.Then,a residual layer is added to the encoder to better process edge and similar information and extract low resolution features.Finally,design a backbone architecture to integrate optimized content and train the optimized model.The experimental results show that the predicted consistency of the optimized model in the TuSimple validation set exceeds 90%and exhibits good adaptability to various interference situations.