End-to-End Lane Detection: a Key Point Approach

Nowadays, autonomous driving is becoming more and more popular. Lane line detection is very important for trajectory planning and decision making in autonomous driving. Traditional lane detection methods rely on highly defined, manual feature extraction and heuristic methods, which usually require post-processing technology. More and more recently, the approach is modeling with deep learning. The lane line scheme based on segmentation usually requires large model and complex convolution structure design, and it cannot perceive the lane line geometric features. Similar to the heat map scheme, the detection of the key points of the lane line actually belongs to the same scheme as the segmentation in a certain angle, but it only reduces part of the amount of computation. The current methods all ignore the data imbalance between the lane line categories that the near lane line occupies most of the position of the picture, resulting in far lane line samples are far less than the near samples. In this paper, a novel detection scheme for key points of lane lines is proposed. The key points of lane lines are linearly sampled at different intervals on the longitudinal axis of images to solve the problem of data imbalance between lane lines. Then the sampled anchor points are fixed, and the model only needs to predict the abscissa of each lane line at the anchor points. At the same time, the geometric constraint loss function of the lane line is put forward to ensure the correct lane line shape. The method presented in this paper achieves 50 FPS on embedded devices, it achieved SOTA on the Culane and Tusimple datasets.