Vehicle-mounted vision-lidar intermediate fusion perception algorithm based on dynamic DBSCAN and adaptive weights

Abstract

To address issues in autonomous vehicles (AVs) scenarios such as insufficient robustness of single-sensor perception, unstable long-distance point cloud clustering results, and the difficulty of fixed-weight fusion strategies in adapting to the physical characteristics of sensors, this paper proposes a decision-level mid-level fusion perception algorithm based on YOLOv8 visual detection and dynamic DBSCAN point cloud clustering. The algorithm uses the KITTI dataset as a verification platform, complements multi-modal information from vehicle-mounted cameras and LiDAR, and improves the reliability of target detection and distance estimation under different distances and densities through dynamic clustering parameters, distance-adaptive weights, and clustering quality-driven confidence calculation. While ensuring lightweight and real-time performance, it effectively reduces missed detections and false detections, making it suitable for vehicle-mounted embedded perception systems.