Enhancing 3D LiDAR Object Detection in Real-World Infrastructure With Trained Models Using Simulation Data

Abstract

The rapid development of autonomous driving technology has increased the need for accurate vehicle environmental perception, with infrastructure 3D LiDAR playing a crucial role in enhancing traffic safety and management. However, processing its large-scale, sparse, and noisy point clouds remains challenging, compounded by the high cost of annotation. This paper proposes a deep learning (DL)-based framework integrating foreground-background separation, noise modeling, and automatic point-cloud annotation. The DL-based separation method removes approximately 99% of background points, significantly improving object detection accuracy. The noise modeling approach transfers ray-wise geometric perturbations learned from real measurements to simulator-generated point clouds under a fixed LiDAR configuration, achieving up to 95% similarity based on normalized Chamfer Distance in the evaluated setting. This enables models trained solely on simulated data to reach 69% accuracy on real-world data at an IoU threshold of 0.5. Additionally, the proposed scenario framework integrates LiDAR point clouds with OpenSCENARIO 2.0 to support simulator-based automatic annotation and training-data generation. Experimental results demonstrate the effectiveness of the proposed framework for foreground reduction and sim-to-real training-data refinement in the evaluated infrastructure LiDAR setting. The present study does not claim validated robustness to adverse weather or broad vehicle-geometry variation, which remain important directions for future evaluation.