Abstract: In view of the problem that existing research rarely uses the same network to achieve simultaneous detection of vehicles and lane lines, and the real-time performance is poor, the traditional DeepLabv3+ network is improved for lightweighting in this paper. The original Xception backbone network is replaced by MobileNet v2 network. The MobileNet v2 network is further changed from five downsampling to four and the learning rate and other parameters are adjusted. An image data set containing both vehicles and lane lines is constructed. The Xception, MobileNet v2 and the improved method proposed in this paper are trained and tested on the data set. The experimental results show that the mean pixel accuracy (mPA) and mean intersection over union (mIoU) of the proposed method is 86.78% and 77.66% respectively, which is basically the same as that of the original Xception backbone network. The model volume is 22.44 m 3 , which is reduced by 89.3%, and the frames per second (FPS) is increased by 72.43%. This proposed method not only guarantees the detection accuracy, but also greatly improves the detection speed. Therefore, this study provides suggestions for vehicles and lane lines detection in the process of autonomous driving of intelligent networked vehicles, and provides strong support for vehicle driving safety

Key words: vehicles detection; lane lines detection; DeepLabv3+; model lightweighting

摘要： 针对现有研究较少采用同一网络实现车辆和车道线的同时检测，且实时性较差的问题， 文章对传统的 DeepLabv3+网络进行轻量化改进，用 MobileNet v2 网络替换原有的 Xception 主干网络，进一步将 MobileNet v2 网络的五次下采样改成四次，并调整学习率等参数；构建 同时包含车辆和车道线的图像数据集，将 Xception、MobileNet v2 和文章提出的改进方法在 数据集上进行训练和测试。实验结果表明，该改进方法的平均像素准确率（mPA）和平均交 并比（mIoU）分别为 86.78%和 77.66%，与原 Xception 主干网络相比基本相同；模型体积为 22.44 m 3，减小了 89.3%，帧率（FPS）提高了 72.43%，该方法在保证检测精度的同时大幅提 升了检测速度。因此，该研究为智能网联汽车自动驾驶过程中的车辆和车道线检测提供了建 议，为汽车行驶安全提供了有力支持。

关键词: 车辆检测；车道线检测；DeepLabv3+；模型轻量化