Improving lane detection with adaptive homography prediction

Lane marks regulate the routes and define the prior drivable areas for a vehicle. Robust detection of lanes plays a vital role in intelligent vehicle navigation. Lane detection algorithms are usually composed of two steps: lane candidate generation and lane curve fitting. The latter is not only used for fitting lane mark candidates with concise curve forms but also for removal of the outliers produced in the former step. Therefore, lane curve fitting is crucial for lane detection. In this step, a common way is carrying out the curve fitting on the bird’s-eye view (BEV), which can mitigate the distortion caused by the perspective projection and improve the fitting results. However, due to the sloping road surfaces in real scenarios, the relative pose between the camera and the ground can change frequently, where using a fixed pre-calibrated projection matrix could bring extra errors in curve fitting. In this paper, we propose a homography prediction network named HP-Net for robust lane mark fitting under various sloping roads. The network can adaptively predict the homographic projection matrix for each input image, producing a suitable BEV for lane fitting. Considering the parallel nature of multiple lanes, the HP-Net could skillfully be trained by reusing the lane labels originally for the task of lane mark segmentation, without introducing any extra manpower. Our method has been verified on a large dataset CULane and another dataset acquired by ourselves. Experiment results show that the proposed model can effectively improve the robustness and accuracy of lane detection.