Lane marking detection based on Convolution Neural Network from point clouds

Lane marking detection, as a key technique for the Highly Automated Driving (HAD) Map, has drawn much attention recently. Traditional methods mainly focused on processing color images, where the performance was affected by illumination variation, over-exposure and occlusion severely. Confronted with above problems, this paper proposes a lane detection algorithm based on Convolution Neural Network (CNN) from point clouds. Our contributions are twofold. On one hand, a CNN framework via gradual up-sampling is introduced, where robust and accurate detection results are achieved. Before applying the CNN model, we also design pre-processing steps, including point clouds registration, the road surface segmentation and orthogonal projection. On the other hand, we propose to analyze the layout of lanes by utilizing the global information and domain knowledge. Hence, false detections caused by ground arrows and texts could be eliminated. Visual and quantitative experiments demonstrate the effectiveness of our algorithm.