Camera-LiDAR Traffic Light Relevance Estimation as Graph-Matching for Autonomous Vehicles

This paper addresses the challenge of identifying which traffic light in an intersection regulates the motion of the ego vehicle. We propose a modular pipeline that combines standard traffic light recognition with a novel relevance estimation module based on human-like logic. The core contribution is a graph-based reasoning method that encodes semantic cues from camera images and geometric information from LiDAR into a structured scene graph. This graph is matched against an offline topological map using inexact graph matching to determine the most relevant traffic light. We evaluate our approach on a proprietary dataset collected on Italian roads, demonstrating promising performance in identifying relevant signals and ensuring appropriate behavior at intersections. Our results highlight the potential of the proposed method as a reliable and efficient component for traffic light understanding in autonomous driving systems.