LaneMP: Robust Lane Attention Detection Based on Mutual Perception of Keypoints

Lane detection is a challenging task that requires predicting complex lane topology shapes in autonomous driving tasks. Some methods use instance segmentation to classify all pixels into lanes and backgrounds; There are also methods that predict each anchor into different lane categories based on the idea of anchor detection. However, these models are less robust and poorly detected. In order to solve these problems, we propose a robust lane attention detection network based on the mutual perception of keypoints (LaneMP), which uses the idea of keypoint detection to predict the keypoints on the lane and then clusters these keypoints into lane instances. Since the clustering process depends on the start points, a loss function is designed to guide the network to learn the correct start points. In addition, aiming at some special scenarios, we propose a horizontal stripe attention mechanism, which can adaptively capture the connection among keypoints through local symmetry of lanes, and improve the robustness of the network. Numerous experiments show that the network has an F1 value of 77.11% on CULane and 96.73% on the TuSimple dataset.