Author(s)

Qianghuang Huang, Fuxin Xu and Su Wang

Corresponding Author

Qianghuang Huang

ABSTRACT

Common lane segmentation methods have disadvantages such as complicated pretreatment process, rough detail segmentation, low robustness and large computation. In view of these shortcomings, we propose a new lane segmentation method that combine full convolution neural network and conditional random field (CRF). The front end is feature represented by U-Net network, and the image is segmented into two parts, lane and background, while the back end adopts CRF for smooth segmentation of the lane edge. The main improvements made in this paper include: in order to improve the accuracy and robustness of lane segmentation method, we give up migration learning and adopt the latest, largest and most complete database -BDD100K for model training; LeakyReLU activation function is used to replace the original relu function to avoid parameter necrosis. In order to make the segmentation method able to deal with the complex lane environment, we add the CRF to the back end for edge segmentation. We achieve 98.86% ACU and the processing speed of 0.09s per image, which is improved compared with other methods. The test results prove that our method is effective.

KEYWORDS

Full convolution neural network-conditional random field-pattern recognition-lane segmentation