Lane Line Detection Based on Improved U-Net Network

Abstract

The lane line detection and recognition are crucial research area for automatic driving. It aims at solving the problem of fuzzy feature expression and low time-sensitives of lane line detection based on semantic segmentation. This paper proposes to remove irrelevant background by dynamic programming region of interest while improving the lightweight neural network (U-Net). A group-by-group convolution and depth wise separable convolution in the backbone network are introduced, simplifies the branches of the backbone network, and atrous convolution is introduced into the enhanced path network with multi-level skip connection structure to retain the underlying coarse-grained semantic feature information. The full-scale skip connection fusion mechanism of the decoder is preserved, while capturing the fine-grained semantics and coarse-grained semantics of the feature map at full scale. The introduction of skip connections between the decoder and the encoder can enhance the lanes without increasing the size of the receptive field. The ability to extract line features and the ability to extract context improves the accuracy of lane lines. The experimental results show that the improved neural network can obtain good detection performance in complex lane lines, and effectively improve the accuracy and time-sensitives of lane lines.