Beyond the Vehicle: Cooperative Localization by Fusing Point Clouds for GPS-Challenged Urban Scenarios

Accurate vehicle localization is a critical challenge in urban environments where GPS signals are often unreliable. This paper presents a cooperative multi-sensor and multi-modal localization approach to address this issue by fusing data from vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) systems. Our approach integrates cooperative data with a point cloud registration-based simultaneous localization and mapping (SLAM) algorithm. The system processes point clouds generated from diverse sensor modalities, including vehicle-mounted LiDAR and stereo cameras, as well as sensors deployed at intersections. By leveraging shared data from infrastructure, our method significantly improves localization accuracy and robustness in complex, GPS-noisy urban scenarios.

💡 Research Summary

The paper tackles the well‑known problem of unreliable GNSS positioning in dense urban canyons, where multipath reflections, atmospheric delays, and satellite occlusion cause meter‑scale errors. To overcome these limitations, the authors propose a cooperative multi‑sensor, multi‑modal localization framework that fuses vehicle‑to‑vehicle (V2V) and vehicle‑to‑infrastructure (V2I) data through V2X communication. The core idea is to share 3‑D point clouds generated by LiDAR and stereo cameras mounted on the ego vehicle, nearby vehicles, and fixed roadside sensors (e.g., at intersections). These point clouds are integrated into a global reference map and used to correct the ego vehicle’s pose estimate in real time.

The system consists of four processing stages. First, the ego vehicle runs an online SLAM algorithm (graph‑based optimization or EKF) using its own LiDAR/camera data together with a noisy GNSS solution, producing a local trajectory and map but suffering from drift over time. Second, point clouds received from other agents and infrastructure are down‑sampled and described with Fast Point Feature Histograms (FPFH), a 33‑dimensional descriptor that captures surface normals, curvature, and angular relationships. Third, a hybrid registration pipeline aligns the ego‑vehicle’s local cloud to the global map: an initial coarse alignment is obtained from the FPFH correspondences, followed by Iterative Closest Point (ICP) refined with Random Sample Consensus (RANSAC) for robust outlier rejection. The resulting rigid transformation (T ∈ SE(3)) is applied to the ego cloud, effectively correcting the pose estimate. Finally, the corrected pose is fed back to the navigation stack, replacing the original GNSS‑based position.

The authors formalize GPS error sources (multipath attenuation, ionospheric/tropospheric delays, PDOP increase) and embed them in a mathematical model. They also detail the Bayesian formulation of online SLAM, the computation of FPFH, and the optimization problem solved by ICP (minimizing squared point‑to‑point distances). RANSAC is used to mitigate the risk of local minima and to handle large fractions of outliers typical in multi‑agent fusion.

Experimental validation is performed in the CARLA simulator across four traffic scenarios (parallel crossing, stationary agents, perpendicular crossing, left turn) and three sensor configurations: (i) four roadside sensors plus two vehicle agents (4_infra_2_agent), (ii) four roadside sensors only (4_infra), and (iii) two roadside sensors only (2_infra). Pose estimation errors are reported as root‑mean‑square error (RMSE) in meters, together with a “valid frame” filter that selects only those registration results that satisfy local fitness and inlier‑distance thresholds. Results show that raw GPS errors range from 7.9 m to 16.9 m, while standalone SLAM maintains low errors (0.09–0.16 m). Pure registration without fusion exhibits highly variable errors (0.02–59 m), depending on sensor density. The proposed fused approach, after discarding invalid frames, consistently reduces errors to the millimeter range (0.009–0.028 m). The best performance is achieved when both infrastructure and vehicle agents contribute data, confirming the benefit of multi‑agent cooperation.

The paper acknowledges several practical challenges: V2X communication latency, precise extrinsic calibration among heterogeneous sensors, and scalability of the registration pipeline when many agents are present. Future work is outlined to address these issues through deep‑learning‑based feature extraction and fusion, edge‑computing architectures for distributed processing, and real‑world field trials to validate robustness under varying traffic densities and environmental conditions.

In summary, the study demonstrates that integrating cooperative point‑cloud sharing with an ICP‑RANSAC‑enhanced SLAM pipeline can dramatically improve localization accuracy in GPS‑degraded urban settings, achieving millimeter‑level precision. This represents a significant step toward reliable, high‑precision positioning for autonomous vehicles operating in complex city environments.