Light Detection and Ranging (LiDAR) systems play a crucial role in automotive research and various sensing-driven tasks, including navigation, 3D mapping, and environmental monitoring. A key function of LiDAR is generating point cloud data, which serves as input for artificial intelligence (AI) and machine learning (ML) algorithms to perform 3D object detection. However, the performance of high-fidelity 3D object detection models primarily relies on LiDAR resolution in capturing precise geometric information, especially at extended ranges. Higher resolutions yield dense point clouds, which can improve detection accuracy but, in turn, result in significantly increased data volumes and computational costs. Therefore, a trade-off between LiDAR-generated point cloud density and ML-based detection accuracy is critical for designing efficient 3D perception systems.

In this study, we propose an AI-based approach to systematically investigate LiDAR configuration for an optimal design. Specifically, we examine the impact of varying LiDAR angular resolutions—ranging from fine (e.g., 0.1° × 0.1°) to coarse (e.g., 1.0° × 1.0°)—on the accurate detection and identification of obstacles. For this, we utilize advanced simulation tools to generate synthetic point cloud data at different angular resolutions based on urban conditions. Such synthetic data enable the analysis of LiDAR configuration at different point cloud densities virtually, reducing the high costs of physical prototyping and testing. In addition, we employ state-of-the-art ML algorithms across different resolutions to analyze how the generated point cloud densities influence 3D object detection performance. The investigated models include deep learning algorithms, categorized into voxel-based, point-based, and hybrid paradigms, each offering global contextual understanding alongside local geometric precision. Our findings provide practical insights into enhancing LiDAR configuration for various applications and enable engineers to select LiDAR settings that balance detection capability with computational efficiency.