Efficient Multi-Sensor Fusion for Cooperative Autonomous Vehicles Leveraging C-ITS Infrastructure and Machine Learning

The widespread deployment of Cooperative Intelligent Transport Systems (C-ITS) has elevated the need for robust, real-time sensor fusion strategies capable of handling noisy, asynchronous data from multiple infrastructure sensors. In this paper, we propose a two-stage data fusion framework that integrates a grid-based indexing method for efficient duplicate-object detection with a Light Gradient Boosting Machine (LGBM) augmented by an Extended Kalman Filter (EKF). In the first stage, the hybrid EKF-LGBM model mitigates noise, refines object trajectories, and synchronizes sensor streams under varying noise conditions. In the second stage, the grid-based indexing technique rapidly associates objects detected by multiple sensors, merging their measurements into unified state estimates. Extensive experiments-using both synthetic data, where noise scales ranged from 0.5 to 3, and a real-road dataset-confirm that our approach balances near-real-time performance with significantly improved trajectory accuracy. For instance, at a noise scale of 1, the hybrid method outperforms the Unscented Kalman Filter (UKF) while running up to 1.81 times faster, and real-world tests show a 1.54 times RMSE improvement over baseline measurements. By efficiently filtering out noise and minimizing the computational overhead of pairwise comparisons, the proposed system demonstrates practical feasibility with respect to C-ITS applications.