双注意力机制下自动驾驶汽车车道线深度感知研究

Research on lane depth perception of autonomous vehicles with dual attention mechanism

为改善车道线分割存在的计算量大、融合效果不明显以及遮挡、丢失、误识别等问题,设计了一种轻量化的基于语义分割的编解码卷积神经网络结构,在网络中引入通道注意力机制与行、列注意力机制。采用轻量化的训练网络ResNet-18对输入图片进行快速下采样,用来产生多阶段特征图;将通道注意力机制用于高阶特征图以提取高阶语义信息;将行、列注意力机制用于低阶特征图以提取车道线的空间信息,采用特征融合机制FFM将高阶特征图上采样后与低阶特征图融合,以提高车道线分割精度。取代传统聚类方法,构建了3层全连接网络,对分割出的像素进行类别预测,实现了背景及车道线的分类,使整个网络得到了端到端的训练与输出。将轻量化的编解码网络模型在Tusimple数据集上完成了车道线检测的训练与测试,并与以往研究模型进行了对比。结果表明,在车道线存在遮挡、模糊、阴影干扰及曝光等场景下,所设计的深度卷积网络仍可以准确且快速地识别出车道线,与现有车道线检测模型相比,在分割精度和检测速度上均有所提高,能够满足自动驾驶实时性检测的需求。

In order to solve the problems caused by lane segmentation,such as heavy computation,weak fusion effect,occlusion,loss and misrecognition,this paper designs a lightweight convolutional neural network structure based on semantic segmentation to introduce the channel attention mechanism,as well as row and column attention mechanism into the network.A lightweight training network ResNet-18 is frstly used to rapidly downsample the input images to generate multi-stage feature maps.Then,the channel attention mechanism is applied to higher-order feature maps to extract higher-order semantic information.The row and column attention mechanism is applied to the low-order feature map to extract the spatial information of the lane lines.Furthermore,the feature fusion mechanism FFM is used to sample the high-order feature map and get fused with the low-order feature map to improve the segmentation accuracy of lane lines.A three-layer fully connected network is constructed to predict the categories of the segmented pixels,which replaces the traditional clustering method,classifies the background and the lane lines,and enables the whole network to get end-to-end training and output.The lightweight codec network model is trained and tested on Tusimple data set for lane detection,and later compared with previous research models.The results show that the designed deep convolutional network can still accurately and quickly identify lane lines in the case of lane lines with occlusion,blurring,shadow interference and exposure.Compared with the existing lane detection model,the segmentation accuracy and detection speed are improved,which can meet the requirements of real-time detection of automatic driving.