Lane Detection and Pixel-Level Tracking for Autonomous Vehicles

Lane detection and tracking play a key role in autonomous driving, not only in the LKA System but help estimate the pose of the vehicle. While there has been significant development in recent years, traditional outdoor SLAM algorithms still struggle to provide reliable information in challenging dynamic environments such as lack of roadside landscape or surrounding vehicles at almost the same speed or on the road in the woods. On the structured road, lane markings as static semantic features may provide a stable landmark assist in robust localization. As most of the current lane detection work mainly on separated images ignoring the relationship between adjacent frames, we propose a pixel-level lane tracking method for autonomous vehicles. In this paper, we introduce a deep network to detect and track lane features. The network has two parallel branches. One branch detects the lane position, while the other extracts the point description on a pixel level. In our approach, the performance of lane detection improves by using the features extracted from past frames, and the description branch has been pre-trained on a synthetic dataset with known ground truth. Then we calculate the Euclidean norm between the description vectors of the same lane to find lane point matches and achieve a better performance against the occlusion by surrounding vehicles by using a modified NW algorithm to calculate the matching scores. To validate the system, we experiment on a video-instance lane detection dataset VIL-100. Experiments show that the proposed method can get a precise matching result.