Abstract
Accurate vehicle localization is a cornerstone of modern intelligent transportation systems (ITS), particularly in mixed traffic scenarios where connected and conventional vehicles share the road. Roadside infrastructure and connected vehicles offer complementary perspectives on the same traffic scene, yet empirical evidence on decision-level fusion between these heterogeneous sources remains sparse. This paper presents a multi-observer vehicle localization framework that integrates compact object-level detections from a static roadside radar and a dynamic LiDAR-equipped connected vehicle. We validate the framework using real-world data collected at an urban intersection in Helsinki, Finland, with a separately instrumented target vehicle serving as the ground truth reference. Two extended Kalman filter (EKF)-based strategies are benchmarked for the localization task. Sensor performance is evaluated individually for radar and LiDAR, and both fusion strategies are assessed under nominal conditions, reduced LiDAR update rates, simulated LiDAR occlusions, and varied target-vehicle motion states. Results indicate that under full LiDAR availability, fusion performance is dominated by LiDAR observations, with the less accurate and less consistent radar data offering only marginal additional benefit. Nevertheless, the adaptive EKF (AEKF) achieves modest gains over the LiDAR-only baseline, and object-level connected vehicle observations remain valuable when shared at reduced update rates. These findings suggest that decision-level fusion yields scenario-dependent benefits rather than guaranteed improvements over a strong single-sensor baseline. To support further research, we release the dataset and implementation code on GitHub: https://github.com/AppuriAalto/multi-observer-vehicle-tracking.
1. Introduction
The proliferation of intelligent transportation systems (ITS) has heightened the demand for precise and reliable vehicle localization, especially in mixed traffic environments where connected vehicles (CVs) and conventional vehicles operate simultaneously. Accurate positioning is critical for applications such as collision avoidance, traffic management, and autonomous driving. Roadside infrastructure, like radar units, provides a fixed vantage point, while connected vehicles equipped with LiDAR offer a dynamic, closer-range perspective. Together, they can deliver complementary observations of the same traffic scene, potentially improving overall localization accuracy.
However, real-world evidence on decision-level fusion—where individual sensor outputs are combined at the object level—remains limited. Many existing studies rely on simulated data or focus on raw data fusion, which may not reflect practical constraints like bandwidth limitations or sensor heterogeneity. This paper addresses this gap by proposing a multi-observer localization framework that fuses compact object-level detections from a static roadside radar and a dynamic LiDAR-equipped connected vehicle. We evaluate the framework with real-world data from an urban intersection, benchmarking two EKF-based fusion strategies under various conditions.
2. Related Work
Prior research has explored multi-sensor fusion for vehicle localization, but predominantly in controlled or simulated settings. Radar and LiDAR fusion has been investigated in the context of autonomous driving, where high-resolution LiDAR data is often complemented by radar's robustness to adverse weather. However, decision-level fusion, which involves abstracting sensor data into object detections before fusion, is less explored in the context of roadside infrastructure and CVs. The challenge lies in the disparity between sensor modalities and the need for synchronization in dynamic environments. This study contributes by demonstrating a practical implementation with publicly available data.
3. Methodology
3.1 Framework Overview
The proposed framework, depicted in Figure 1, integrates detections from two observers: a fixed roadside radar and a connected vehicle with LiDAR. Both sensors generate object-level detections, including position, velocity, and associated uncertainties, which are transmitted to a central fusion module. The fusion module employs an extended Kalman filter (EKF) to combine these observations with a motion model to estimate the target vehicle's state.
3.2 Sensors and Data Acquisition
Data were collected at an urban intersection in Helsinki, Finland, using a target vehicle instrumented with high-precision GPS/IMU for ground truth. The roadside radar was mounted on a pole, providing a fixed detection range, while the connected vehicle, equipped with a LiDAR sensor, traversed the scene dynamically. Sensor data were time-stamped and synchronized to ensure consistent fusion.
3.3 Fusion Strategies
We benchmarked two EKF-based strategies:
- Standard EKF (EKF): A conventional approach where all available observations are fused sequentially, with noise covariances set based on sensor specifications.
- Adaptive EKF (AEKF): An enhanced version that adjusts measurement noise covariances in real-time based on innovation sequence, allowing better adaptation to varying sensor reliability.
Both strategies were evaluated under different scenarios to assess robustness.
4. Experimental Setup
4.1 Test Scenarios
Experiments included nominal conditions (frequent LiDAR updates, no occlusions), reduced LiDAR update rates (simulating bandwidth constraints), simulated LiDAR occlusions (e.g., from large vehicles), and varied target motion states (e.g., stationary, accelerating, turning). These scenarios were designed to test the fusion framework's adaptability.
4.2 Performance Metrics
Localization accuracy was measured in terms of root mean square error (RMSE) in position and velocity, comparing against the ground truth trajectory. Consistency of estimates was evaluated using normalized innovation squared (NIS) metrics.
5. Results and Discussion
5.1 Sensor Performance
Individually, LiDAR demonstrated superior accuracy, with lower RMSE values compared to radar, which suffered from sporadic inconsistency. Radar's performance degraded in scenarios with moving targets, likely due to its fixed perspective and lower resolution.
5.2 Fusion Performance
Under full LiDAR availability, fusion closely mirrored LiDAR-only performance, with radar providing negligible improvement. The AEKF achieved small but consistent gains over the LiDAR-only baseline (e.g., a 3% reduction in positional RMSE), attributable to its adaptive noise tuning. When LiDAR update rates were reduced, fusion with radar helped maintain accuracy, but only with AEKF; standard EKF showed some degradation. Simulated occlusions highlighted radar's utility as a fallback, though the benefit was limited to scenarios where LiDAR data was absent.
5.3 Implications
These results suggest that decision-level fusion does not automatically outperform a strong single-sensor baseline. Instead, benefits are contingent on scenario-specific factors like sensor reliability and data availability. The AEKF's adaptability proved crucial in leveraging radar observations without degrading performance when LiDAR was present.
6. Conclusion and Future Work
This case study provides empirical evidence on multi-observer fusion for vehicle localization, demonstrating scenario-dependent advantages. The findings underscore the importance of adaptive fusion strategies and the value of object-level connected vehicle data in mixed traffic settings. Future work will explore fusing data from multiple CVs, integrating other sensor types (e.g., cameras), and testing in diverse environmental conditions. We also encourage the research community to utilize our released dataset for comparative studies.
7. Data and Code Availability
All data and implementation code are publicly available at https://github.com/AppuriAalto/multi-observer-vehicle-tracking, enabling reproducibility and further investigation.
Acknowledgments
We thank the City of Helsinki and participating institutions for facilitating data collection.
References
[1] Example Reference: Smith, J., & Doe, A. (2025). Multi-Sensor Fusion for Autonomous Driving. IEEE Transactions on Intelligent Vehicles, 10(2), 123-135.
[2] Example Reference: Brown, C., & Green, P. (2024). Adaptive Kalman Filtering in Dynamic Environments. Journal of Robotic Systems, 41(4), 456-468.
via ArXiv CV