Multi-Observer Vehicle Localization Case Study with Roadside Radar and Connected Vehicle Sensing

Accurately estimating vehicle positions is a fundamental requirement for intelligent transportation systems, especially in mixed-traffic environments where connected and conventional vehicles coexist. While roadside infrastructure and connected vehicles offer complementary perspectives, real-world evidence regarding decision-level fusion between heterogeneous sensors remains limited. This paper investigates how to effectively fuse compact object-level detections from a static roadside radar and a dynamic LiDAR-equipped connected vehicle to improve vehicle localization performance.

The authors propose a multi-observer vehicle tracking framework that operates in a radar-centric two-dimensional coordinate frame. Vehicle motion is propagated using a constant turn rate and velocity model within an extended Kalman filter framework. To combine the heterogeneous sensor observations, the study benchmarks two distinct fusion strategies: a sequential extended Kalman filter that updates measurements one by one, and an averaged extended Kalman filter that combines measurements into a single covariance-weighted position estimate prior to the update step. The framework is evaluated using real-world urban traffic data collected at an intersection in Helsinki, Finland, using an instrumented target vehicle with a precise GNSS/INS reference trajectory.

Under nominal conditions with full LiDAR availability, tracking performance is heavily dominated by the higher-accuracy LiDAR observations, while the noisier and less consistent radar data provide only marginal improvements. Nevertheless, the averaged extended Kalman filter achieves modest gains over a LiDAR-only baseline. Furthermore, the object-level connected vehicle observations remain useful for tracking even when shared at reduced update rates. These findings demonstrate that decision-level fusion is most valuable for handling asynchronous data, partial occlusions, and reduced-rate communication rather than uniformly outperforming a primary sensor in ideal conditions.