A Modular Lane Detection Method Based on Scene Understanding

Hui Lu, Xiaoqun Tan, Guanqi Ding

Abstract: Lane detection is an important part for autonomous driving vehicles to locate properly. Traditional lane detection methods rely on a series of complicated algorithms to detect the lane features, followed by some post-processing techniques to reduce the effect of noise. However, these methods require high quality images, and very likely fail when the driving environment has significant variation. Based on the development of deep learning, researchers have proposed pixel-wise lane segmentation with many learning models. However, most methods need a large amount of database to reduce the error and the accuracy will have high fluctuation when the driving environment changes. In this paper, we proposed a lane detection modular to extract the lane area to reduce the environment effect. Also we proposed a learning model which utilizes the lane features history information to predict the lane position when no features in the next images. The proposed method demonstrated improved accuracy and robustness compared with recent methods based on deep learning.