Closed-Loop Edge Following for High-Definition Map-Free Autonomous Driving With Real-Time Road Edge Detection

Edge following-where an autonomous vehicle tracks a road edge-is a critical component of intelligent transportation systems, underpinning applications such as curbside sweeping, line marking, and container docking at port terminals. Traditional edge-following planners are often open-loop and rely on manually crafted High Definition maps, limiting adaptability. However, substituting static map inputs with real-time sensor references exposes the limitations of single-modal sensing, which struggles to capture irregular curb geometries or subtle height variations, degrading detection and trajectory quality. In this paper, we propose a novel map-free, closed-loop framework that integrates real-time LiDAR-vision fusion with an any-time constrained quadratic-program planner. Our perception pipeline extracts coarse curb candidates directly in the 3-D point cloud-preserving centimetre-scale vertical gradients-and then refines them via a lightweight image segmentation head on a cropped region of interest. The planner formulates each along-edge trajectory as a sparse quadratic programming in the Frenet frame, enforcing hard collision and comfort constraints while guaranteeing feasibility at every iteration and supporting safe interruption under real-time deadlines. This continuous feedback loop adapts instantly to road geometry and dynamic obstacles without High Definition map dependence. Extensive simulations and real UGV trials in a campus sweeping scenario demonstrate the framework’s robustness, responsiveness, and deployability in real-world conditions.