Lane Detection Algorithm Based on Dilated Convolution Network

Abstract

Aiming at the complex and diverse problems of road scenes and lane lines in the actual driving environment, a lane detection algorithm based on improved lanenet network is proposed. The expansion layer is constructed by dilated convolution to extract local and context information in a simple and effective way. Then the parallel dilated convolution module is used to extract lane line features of different scales. It is used to classify each lane line pixel, and the lane line marking point set is output to optimize the segmentation effect. Finally, the lane line is fitted by hnet network. Experiment results shows that the proposed algorithm achieves 65.7% mean Intersection-over-Union(mIOU) on TuSimple test set at the speed of 73 Frames Per Second(FPS). The accuracy of the proposed algorithm is 1.1 % higher than lanenet network. The proposed algorithm is helpful to complete tasks such as efficient and accurate street scene image segmentation in automatic driving.