Aiming at the problems of complex algorithm structure and large number of parameters existing in road lane visual detection technology under all-weather conditions, a lane detection method based on depth-separable convolution and residual attention module is proposed, and the LPINet network model is established. We use depth-separable convolution to reduce the size of the input images, design three bottleneck residual units with different structures to reduce the number of network parameters, and introduce the ECANet attention mechanism, which can increase the weight of important feature channels, to improve the lane detection accuracy. The experimental results on Tusimple dataset and GZUCDS self-built dataset show that the LPINet network lane detection accuracy can reach 96.62% in sunny scenarios, and the number of model parameters is reduced to 1.64 MB, which realizes the lightweight design. We carried out exploratory researches in complex scenes such as foggy, rainy, night and tunnel, and the accuracy of lane detection reaches 93.86%, which proves the effectiveness of our method.