Online lane mapping based on multi-sensor SLAM and Catmull–Rom splines

Abstract Traditional offline high-definition maps have significant limitations owing to their high cost and difficulty in updating. Online vector maps represent a new trend in the development of autonomous driving. In this paper, we propose an online lane mapping system, which leverages both multi-sensor simultaneous localization and mapping (SLAM) and Catmull–Rom splines. The map is constructed in real-time through the localization provided by SLAM and the fitting of splines. Existing SLAM algorithms perform poorly in complex urban environments, a deep learning network is incorporated into the visual front-end of our system to improve stability. Deep learning can extract and study deeper and less observable features, so that the system enhances the accuracy of visual tracking in scenarios characterized by low texture, lighting variations, and rapid motion. The system state is optimized by a factor graph framework to achieve precise localization. To construct continuous and consistent online lane maps, our system combines SLAM localization results with spline curves to achieve lane connection. Additionally, we develop a convenient method for initializing, extending, and optimizing the control points of splines to obtain more precise results. Our method combines the geometric characteristics with historical observations of lanes, enhancing the association and consistency of maps. Experiments conducted on the mainstream datasets demonstrate that our system exhibits greater robustness and improves the accuracy of localization and mapping across diverse environments.