基于改进的DeeplabV3+模型的车道线检测方法

Lane Line Detection Method Based on Improved DeeplabV3+ Model

为兼顾车道线检测的准确性与实时性,本文提出一种基于改进的DeeplabV3+模型的车道线检测方法。首先通过水平翻转、改变图像的亮度、饱和度等方法对车道线图像进行数据增广,加强模型对于眩光、车道线破损情况下长直型、大曲率车道线的泛化能力。其次将模型的主干网络更换为轻量级的MobilenetV2网络,提高模型训练速度,并依据车道线图像特点改进ASPP结构,合理设计组合多采样率空洞卷积,提高模型对边缘车道线及远处车道线的预测效果,利用深度可分离卷积,减少模型参数量。最后本文依据车道线图像特点提出了双注意力机制DAMM结构,通过合理分配注意力资源,提高模型分割能力。实验表明,改进的DeeplabV3+模型像素精度为99.35%,平均交并比为86.08%,单图预测时间为22.62ms,说明改进DeeplabV3+模型兼顾准确性和实时性。

In order to balance the recognition accuracy and real-time performance of lane detection, this paper proposes a lane detection method based on the improved Deeplab V3+ model. Firstly, the lane line images are expanded by horizontal flip and the brightness and saturation of the images are changed to enhance the generalization ability of the model for both long straight and large curvature lane lines under glare and lane line breakage. Secondly, the backbone network of the model is replaced with a lightweight Mobilenet V2 network to improve the training speed of the model, and the ASPP structure is improved according to the characteristics of the lane line images, and the multi-sampling rate null convolution is reasonably designed to improve the prediction effect of the model on the edge lane lines and the distant lane lines. Finally, the dual-attention mechanism DAMM is fused according to the characteristics of lane line images, and the attention resources are reasonably allocated to improve the model segmentation capability. Experiments show that the accuracy of the improved Deeplab V3+ model is 99.35%, the m Io U is 86.08%, and the single-image prediction time is 22.62ms, which indicates that the improved DeeplabV3+ model takes into account both accuracy and real-time performance.