关键词: 车道线检测；深度卷积神经网络；多项式拟合；辅助驾驶；MobileNetV3 网络

Abstract: Lane detection is an important part of vehicle assisted driving. In order to realize detecting lane lines accurately and quickly, a lightweight lane detection algorithm based on MobileNetV3 network is proposed. Firstly, the deep separable convolution module of MobileNetV3 network is improved, and the spatial attention mechanism module is added on this basis. Then the lane lines are represented as third-order polynomials, and the lane lines features in the image are extracted by using the optimized MobileNetV3 network to obtain the lane lines parameters used to fit the third-order polynomials. Finally, a regression model of lane-line is constructed to improve the detection accuracy of lane lines by constantly revising the lane lines parameters. The experimental results on Tusimple lane dataset show that the proposed algorithm has a frame processing speed of 210 fps and a detection accuracy of 83.35 %. It can run in real time and has high detection accuracy.

Key words: Lane detection; Deep convolutional neural network; Polynomial fitting; Assisted driving;MobileNetV3 network