LiDAR-to-Graph Framework for Traffic Scene Understanding at Urban Intersections

Effective traffic monitoring is a cornerstone of intelligent transportation systems, enabling safer and more efficient urban mobility. Among emerging sensing modalities, Light Detection and Ranging (LiDAR) stands out as a robust technology that delivers high-resolution 3D spatial information, complementing traditional traffic monitoring systems. However, despite its potential, publicly available LiDAR datasets dedicated to traffic monitoring remain scarce, with most existing datasets primarily designed for autonomous driving from the ego-vehicle perspective. In this work, we develop a comprehensive LiDAR-driven pipeline for urban traffic assessment, with a specific focus on road intersections, one of the most complex and safety-critical areas of urban networks involving vehicles and pedestrians. The proposed simulation-based framework, combining SUMO for traffic scenario generation and Blender for high-fidelity 3D reconstruction with LiDAR emulation, integrates perception and relational reasoning through two main components: 1) a LiDAR-based perception module for 3D object detection and segmentation and 2) a graph-construction module that encodes spatial, geometric, and semantic relations between road users and infrastructure. The resulting datasets offer rich annotations, making them suitable for advanced AI-driven analysis of road user interactions and behaviors. We validate our approach using state-of-the-art models for LiDAR perception and graph neural networks, demonstrating high detection and relation-learning accuracy under both clean and distorted point cloud conditions. The results confirm the effectiveness of the proposed pipeline and highlight LiDAR’s ability to provide a holistic, reliable understanding of complex traffic scenes, paving the way for future smart mobility and intelligent traffic management applications.