LiDAR Ground Detection-Based Dynamic Inverse Perspective Mapping of BEV Lanes

This article proposes a novel method for projecting lane markings detected from camera images onto the road surface to form a bird's eye view (BEV) lane detection result using inverse perspective mapping (IPM). Existing methods that use a fixed homography matrix (fixed IPM) have been observed to lead to distortions due to vehicle pose variations relative to the road surface. Such distortions can cause inaccurate BEV representation of lane detection results, especially in larger vehicles like buses. This article proposes a "dynamic" IPM method to address these challenges to enhance the quality of the BEV perspective transformation of the lane detection result. This approach utilizes light detection and ranging (LiDAR) point clouds to estimate road plane parameters in the vehicle's frame of reference. Lane markings are then projected onto the road plane using dynamic IPM, which considers the vehicle's pose relative to the measured road surface model. The proposed dynamic IPM method is compatible with existing camera-based lane detection methods that detect lane marking points in the image space, given an accurate LiDAR and camera pose in the vehicle's reference frame. The proposed method has been evaluated in simulated and real-world scenarios.