Decision-Level Fusion of YOLOv8 and PointPillars: Initial Findings

This paper presents a multi-modal perception system tailored for autonomous driving and safety monitoring in public parking environments, utilizing a dual-stage decision-level fusion of YOLOv8-seg and PointPillars. The architecture ensures precise occupancy monitoring and safety by integrating 2D instance segmentation with 3D LiDAR point clouds. A specialized decision fusion rngine features a distance-based matching phase and a rescue phase to maintain detections during sensor occlusions. By implementing a 2.5m Euclidean threshold and a high-confidence YOLO override mechanism, the system effectively compensates for LiDAR sparsity. Experimental results on the KITTI dataset demonstrate significant reliability gains, notably increasing the F1-score for pedestrians by 10.11% and cars by 6.99%. These findings prove that the synergy of visual masks and geometric data provides a robust solution for real-time monitoring of vehicles and vulnerable road users (pedestrians and cyclists) in automated parking environments.