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

Aleksi Pippuri; Nilusha Jayawickrama; Risto Ojala

cs.CVAug 17, 2026

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

Authors: Aleksi Pippuri, Nilusha Jayawickrama, Risto Ojala

Organizations: Department of Mechanical Engineering, Aalto University, Espoo, Finland

Abstract

In modern intelligent transportation systems, it is essential to accurately estimate vehicle positions, especially in mixed traffic conditions where both connected and conventional vehicles coexist. Roadside infrastructure and connected vehicles can provide complementary observations of the same traffic scene, but real-world evidence on decision-level fusion between these sources remains limited. This paper proposes a multi-observer vehicle 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 collected at an urban intersection in Helsinki, Finland, with a separately instrumented target vehicle used as the reference trajectory. Two extended Kalman filter based strategies for the localization task were benchmarked. The performance of the radar and LiDAR sensors were evaluated separately, and the two fusion strategies were explored under nominal sensing conditions, reduced LiDAR update rates, simulated LiDAR occlusions, and different target-vehicle motion states. The results show that, under full LiDAR availability, fusion performance is dominated by the LiDAR observations, while the less accurate and less consistent radar observations provide only limited additional improvement. Nevertheless, AEKF achieves small gains over the LiDAR-only baseline, and object-level connected vehicle observations remain useful when shared at reduced update rates. These findings indicate that decision-level fusion provides scenario-dependent benefits rather than automatic improvement over a strong single-sensor baseline. We release the dataset and implementation on Github to support further research: https://github.com/AppuriAalto/multi-observer-vehicle-tracking

Figures & tables

Fig. 1: The proposed framework takes as input vehicle detections from a roadside radar and a CV-equipped LiDAR. The detections are then fused and evaluated against a known reference track from a target vehicle.

Fig. 2: Proposed tracking and fusion framework. Existing tracks are propagated using the CTRV model and associated with radar and LiDAR detections. Depending on observation availability, the framework applies sequential updating, averaged updating, single-sensor updating, or prediction-only propagation. Track management handles initialization, confirmation, deletion, and trajectory output.

Fig. 3: Overview of the experimental workflow. Data collected in the experiment are processed, calibrated, and then used in the proposed fusion framework. Finally the resulting tracks are evaluated against the target vehicle’s reference trajectory.

Fig. 4: Target vehicle trajectory as well as the radar location shown. The segments of the route inside the radar FoV is highlighted, the right-side track being the depart segment and the left one approach. Map provided by OpenStreetMap [ 36 ] and CARTO [ 37 ]

Fig. 5: Connected vehicle preprocessing pipeline. LiDAR-based vehicle detections are georeferenced using the CV GNSS trajectory and transformed into the common radar-centric ENU frame for fusion.

Fig. 6: Roadside radar preprocessing pipeline. Radar-based vehicle detections are extracted, aligned to the CV time base, and transformed into the radar-centric ENU frame for fusion.

Parameter	LiDAR	Radar	SEKF	AEKF
Association gate $\gamma_{\mathrm{assoc}}$	24	44	32	36
Process noise $\sigma_{a}$ [m/s 2 ]	3.5	8.0	4.5	5.0
Yaw-rate noise $\sigma_{\dot{\psi}}$ [rad/s]	0.65	0.85	0.95	0.95
Radar position noise $\sigma_{r}$ [m]	–	1.0	9.0	9.0
LiDAR position noise $\sigma_{l}$ [m]	0.5	–	0.4	0.5

TABLE I: Final parameter values selected after tuning. Parameters not used by a method are marked with –.

Parameter	LiDAR	SEKF	AEKF
Association gate $\gamma_{\mathrm{assoc}}$	24	24	46
Process noise $\sigma_{a}$ [m/s 2 ]	8.0	10.5	7.5
Yaw-rate noise $\sigma_{\dot{\psi}}$ [rad/s]	0.65	0.6	0.9
Radar position noise $\sigma_{r}$ [m]	–	10.0	8.0
LiDAR position noise $\sigma_{l}$ [m]	0.8	0.6	0.5

TABLE II: Final parameter values selected after tuning with LiDAR at

1~\mathrm{Hz}

. Parameters not used by a method are marked with –.

Motion bin	Speed range	Samples
Stationary	$v_{\mathrm{ref}}<0.3$ m/s	27845
Slow	$0.3\leq v_{\mathrm{ref}}<2.0$ m/s	5163
Moving	$v_{\mathrm{ref}}\geq 2.0$ m/s	27815
Total		60823

TABLE III: Distribution of reference samples across motion bins.

Method	RMSE [m] (std)	MAE [m]	Match rate	Drop rate
LiDAR	0.84 (0.30)	0.74	0.983	0.017
Radar	3.27 (0.65)	3.12	0.215	0.785
SEKF	0.84 (0.29)	0.74	0.982	0.018
AEKF	0.81 (0.29)	0.73	0.984	0.016

TABLE IV: Overall performance comparison of evaluated methods across all validation segments.

Method	Segment	RMSE [m] (std)	MAE [m]	Match rate
LiDAR	Approach	0.90 (0.36)	0.82	0.969
LiDAR	Depart	0.78 (0.23)	0.67	0.995
Radar	Approach	2.93 (0.57)	2.76	0.379
Radar	Depart	3.61 (0.56)	3.48	0.067
SEKF	Approach	0.92 (0.37)	0.82	0.969
SEKF	Depart	0.78 (0.19)	0.66	0.994

TABLE V: Performance comparison of evaluated methods across all validation segments.

Fig. 7: Pooled localization error clouds relative to the target vehicle’s reference trajectory for the LiDAR-only, radar-only, SEKF, and AEKF methods. The dashed circles indicate radial error magnitudes of $1~\mathrm{m}$ , $2~\mathrm{m}$ , and $5~\mathrm{m}$ .

Rate	Method	Tune	RMSE [m] ( $\Delta$ )		MAE [m]	Match
$5~\mathrm{Hz}$	LiDAR	Full	0.84	(+1.5%)	0.74	0.981
	LiDAR	$1~\mathrm{Hz}$	0.83	(+0.9%)	0.74	0.981
	SEKF	Full	0.86	(+3.8%)	0.75	0.979
	SEKF	$1~\mathrm{Hz}$	0.85	(+3.0%)	0.75	0.980
	AEKF	Full	0.82	(+0.0%)	0.74	0.981
	AEKF	$1~\mathrm{Hz}$	0.84	(+2.4%)	0.75	0.980

TABLE VI: Performance of the compared methods under varying LiDAR refresh rates. Values in parentheses next to RMSE denote the relative RMSE change compared with the best-performing method at the same refresh rate.

Occlusion	Method	RMSE [m]	MAE [m]	Match rate
$3~\mathrm{s}$	LiDAR	1.34 (+0.64)	1.17 (+0.48)	0.904 (-0.087)
	SEKF	1.46 (+0.74)	1.26 (+0.56)	0.899 (-0.091)
	AEKF	1.43 (+0.72)	1.23 (+0.53)	0.852 (-0.140)
$5~\mathrm{s}$	LiDAR	1.91 (+1.21)	1.59 (+0.91)	0.774 (-0.219)
	SEKF	2.03 (+1.30)	1.70 (+1.00)	0.767 (-0.225)
	AEKF	2.07 (+1.36)	1.71 (+1.02)	0.710 (-0.283)

TABLE VII: Effect of simulated LiDAR occlusions. Metrics are computed only during the occlusion interval and averaged over occlusion start positions. Values in parentheses denote the change relative to the corresponding non-occluded same-window reference. Match-rate changes are reported as absolute proportions.

Rate	Motion bin	LiDAR	SEKF	AEKF
10 Hz	Stationary	0.60	0.60	0.60
	Slow	0.72	0.72	0.72
	Moving	0.92	0.94	0.88
5 Hz	Stationary	0.61	0.61	0.60
	Slow	0.73	0.72	0.73
	Moving	0.92	0.97	0.90

TABLE VIII: Motion-conditioned localization RMSE in meters under varying LiDAR refresh rates using the full-rate tuned configurations.

Radar-LiDAR Fusion

See all 13 papers

ESAFusion: LiDAR--4-D Radar Fusion via Local Geometric Complementation and Multiscale Adaptive Interaction for 3-D Object Detection
Sep 13, 2026Gang Ma, Senjie Hu, Junjie Liu +2Multi-Scale Feature Fusion 3D Object Detection
School of Future Technology, Shanghai University, Shanghai 200444, China
Ultra-Fusion: A Resilient Tightly-Coupled Multi-Sensor Fusion SLAM Framework under Sensor Degradation and Spatiotemporal Perturbation for Intelligent Transportation Systems
Jun 19, 2026Yihong Tian, Junjie Zhang, Liuyang Li +3Simultaneous Localization and Mapping Robot Localization
Beijing Institute of Technology · Chongqing University · Sichuan University +2
CLIFE: Camera-LiDAR Fusion Framework for Edge-Deployable Roadside VRU Perception
Jul 17, 2026Tam Bang, Hoang H. Nguyen, Lei Cheng +6Edge Inference Camera-LiDAR Fusion
Center for Urban Informatics and Progress, University of Tennessee at Chattanooga, USA · University of Maryland, College Park, USA · University of Arizona, USA

Sep 13, 2026J/K move · Enter open · S save

ESAFusion: LiDAR--4-D Radar Fusion via Local Geometric Complementation and Multiscale Adaptive Interaction for 3-D Object Detection

Gang Ma, Senjie Hu, Junjie Liu, Chao Wang+1

School of Future Technology, Shanghai University, Shanghai 200444, China

Multi-Scale Feature Fusion 3D Object Detection Radar-LiDAR Fusion Multimodal Sensor Fusion

Open paper AskPDF arXiv

LiDAR--4-D radar fusion combines accurate spatial geometry with motion and reflectivity cues from radar, offering a promising solution for 3-D object detection in complex driving environments. However, sparse radar observations and differences in spatial sampling between the two modalities complicate reliable cross-modal complementation. Moreover, the relative importance of modalities and feature scales varies across spatial regions, making adaptive fusion challenging. To address these challenges, we propose ESAFusion, an evidence-aware and scale-adaptive framework that combines local geometric complementation with multiscale adaptive interaction. Specifically, we introduce an Evidence-Aware Radar Selection (ERS) module to suppress radar clutter using motion and observation-quality evidence while retaining foreground confidence for subsequent fusion. Then, the Pillar-Level Complementary Encoder (PCE) improves cross-modal complementation under mismatched spatial sampling using local geometric support from neighboring LiDAR pillars. We further design an Intra- and Inter-Scale Adaptive Fusion (ISAF) module to adaptively adjust the contributions of different modalities and feature scales in bird's-eye-view (BEV) space. Extensive experiments on the View-of-Delft (VoD) dataset show that ESAFusion achieves the highest mean average precision (mAP) among the compared methods, reaching 74.60% in the Entire Annotated Area and 88.89% in the Driving Corridor. It also attains the highest average precision (AP) for Cyclist among these methods in both regions while running at 19.23 FPS. Evaluations on VoD-Fog further demonstrate robustness under progressively degraded LiDAR observations. The source code will be made publicly available at https://github.com/SenJieHu549/ESAFusion.

Abstract

Radar-LiDAR Fusion

Explore similar work

ESAFusion: LiDAR--4-D Radar Fusion via Local Geometric Complementation and Multiscale Adaptive Interaction for 3-D Object Detection

Ultra-Fusion: A Resilient Tightly-Coupled Multi-Sensor Fusion SLAM Framework under Sensor Degradation and Spatiotemporal Perturbation for Intelligent Transportation Systems

CLIFE: Camera-LiDAR Fusion Framework for Edge-Deployable Roadside VRU Perception