Context-Aware Sensor Modeling for Asynchronous Multi-Sensor Tracking in Stone Soup

Martin Vonheim Larsen; Kim Mathiassen

cs.CVMar 16, 2026

Context-Aware Sensor Modeling for Asynchronous Multi-Sensor Tracking in Stone Soup

Authors: Martin Vonheim Larsen, Kim Mathiassen

Organizations: Norwegian Defence Research Establishment (FFI) · University of Oslo, Department of Technology Systems

Abstract

Multi-sensor tracking in the real world involves asynchronous sensors with partial coverage and heterogeneous detection performance. Although probabilistic tracking methods permit detection probability and clutter intensity to depend on state and sensing context, many practical frameworks enforce globally uniform observability assumptions. Under multi-rate and partially overlapping sensing, this simplification causes repeated non-detections from high-rate sensors to erode tracks visible only to low-rate sensors, potentially degrading fusion performance. We introduce DetectorContext, an abstraction for the open-source multi-target tracking framework Stone Soup. DetectorContext exposes detection probability and clutter intensity as state-dependent functions evaluated during hypothesis formation. The abstraction integrates with existing probabilistic trackers without modifying their update equations. Experiments on asynchronous radar-lidar data demonstrate that context-aware modeling restores stable fusion and significantly improves HOTA and GOSPA performance without increasing false tracks.

Figures & tables

Panel of Figure 1 — Figure 1: Architectural comparison. Left: The tracker is configured with global detection probability $P_{D}$ and clutter intensity $\lambda$ . Right: The detection reader emits a per-timestep DetectorContext that exposes state and scene dependent functionals $P_{D}({\boldsymbol{\mathbf{x}}};\ \mathcal{S})$ and $\lambda({\boldsymbol{\mathbf{z}}};\ \mathcal{S})$ queried by the tracker. The key difference is that observability is no longer configured as static tracker parameters but evaluated dynamically via the emitted DetectorContext.

Panel of Figure 2 — Figure 2: Effective sensing regions and blind sectors for the radar-lidar configuration.

Figure 3: Scenario 1 ground-truth trajectories. Red markers show stationary objects (sea markers and islets). The ego-vessel remains fairly stationary at the position given by the black marker. The yellow target performs a close-range loop around the ego-vessel.

Figure 4: Scenario 2 ground-truth trajectories. Red markers show stationary objects (sea markers and islets). The ego-vessel follows the black trajectory in formation with a cooperating USV (green trajectory).

JPDA
Component	Setting
Motion model	Constant velocity, $\sigma_{\mathrm{acc}}=0.8$
State	${\boldsymbol{\mathbf{x}}}=[x,\dot{x},y,\dot{y}]^{\top}$
Track init cov.	$\mathrm{diag}([10,15,10,15])^{2}$
Meas./nav cov.	Varying; identical across trackers
Gating	$\chi^{2}$ gate, $P_{G}=0.95$

Table I: Tracker parameters used in all experiments.

	Combined		Scenario 1		Scenario 2
Method	HOTA (%) $\uparrow$	GOSPA $\downarrow$	HOTA (%) $\uparrow$	GOSPA $\downarrow$	HOTA (%) $\uparrow$	GOSPA $\downarrow$
JPDA	$67.5$	$35.2$	$71.2$	$29.6$	$64.3$	$39.5$
GM-PHD (uniform)	$42.0$	$49.2$	$10.1$	$61.4$	$57.3$	$35.2$
GM-PHD (radar-only)	$74.3$	$32.7$	$85.6$	$24.2$	$63.0$	$38.5$
GM-PHD ( $P_{D}$ -aware)	$80.1$	$29.4$	$89.5$	$22.1$	$71.4$	$34.6$
GM-PHD (context-aware)	81.4	27.1	89.7	21.8	73.5	31.0

Table II: Tracking performance across configurations. HOTA is computed using TrackEval multi-sequence aggregation. GOSPA is reported as RMS over all evaluation timesteps.

Panel of Figure 5 — Figure 5: Representative tracking output for Scenario 1 over a 480 s interval. Ground truth trajectories are shown as dashed gray lines. Inlier target tracks are rendered in shades of blue, while clutter tracks are shown in red. The ego-vessel position and heading are indicated by the black marker. (a) JPDA maintains medium-range tracks but exhibits limited clutter suppression in lidar-dominated regions and fragments a target track within lidar range. (b) GM-PHD with uniform observability loses radar-supported tracks due to accumulated missed detections. (c) Context-aware GM-PHD restores track continuity across sensing regimes while effectively suppressing clutter.

Radar-LiDAR Fusion

See all 13 papers

SODA-CitrON: Static Object Data Association by Clustering Multi-Modal Sensor Detections Online
Feb 24, 2026Jan Nausner, Kilian Wohlleben, Michael HubnerData Association Multimodal Sensor Fusion
Center for Digital Safety & Security, Austrian Institute of Technology GmbH, Vienna, Austria
ModTrack: Sensor-Agnostic Multi-View Tracking via Identity-Informed PHD Filtering with Covariance Propagation
Mar 16, 2026Aditya Iyer, Jack Roberts, Nora AyanianMulti-Object Tracking Uncertainty-Aware Multimodal Fusion
Department of Computer Science, Brown University, Providence, RI, USA · School of Engineering, Brown University, Providence, RI, USA
Context-aware Bayesian Fusion for Multimodal Object Detection
Oct 13, 2026Arunachalam Venkatachalam, Mhafuzul Islam, Marwan Yusuf +3Multimodal Object Detection Object Detection
imec Michigan, USA · imec and Ghent University Ghent, Belgium · imec Leuven, Belgium

Feb 24, 2026J/K move · Enter open · S save

SODA-CitrON: Static Object Data Association by Clustering Multi-Modal Sensor Detections Online

Jan Nausner, Kilian Wohlleben, Michael Hubner

Center for Digital Safety & Security, Austrian Institute of Technology GmbH, Vienna, Austria

Data Association Multimodal Sensor Fusion Clustering

Open paper AskPDF arXiv

The online fusion and tracking of static objects from heterogeneous sensor detections is a fundamental problem in robotics, autonomous systems, and environmental mapping. Although classical data association approaches such as JPDA are well suited for dynamic targets, they are less effective for static objects observed intermittently and with heterogeneous uncertainties, where motion models provide minimal discriminative power with respect to clutter. In this paper, we propose a novel method for static object data association by clustering multi-modal sensor detections online (SODA-CitrON), while simultaneously estimating positions and maintaining persistent tracks for an unknown number of objects. The proposed unsupervised machine learning approach operates in a fully online manner and handles temporally uncorrelated and multi-sensor measurements. Additionally, it has a worst-case loglinear complexity in the number of sensor detections while providing full output explainability. We evaluate the proposed approach in different Monte Carlo simulation scenarios and compare it against state-of-the-art methods, including POM-based filtering, DBSTREAM clustering, and JPDA. The results demonstrate that SODA-CitrON consistently outperforms the compared methods in terms of F1 score, position RMSE, MOTP, and MOTA in the static object mapping scenarios studied.

Abstract

Radar-LiDAR Fusion

Explore similar work

SODA-CitrON: Static Object Data Association by Clustering Multi-Modal Sensor Detections Online

ModTrack: Sensor-Agnostic Multi-View Tracking via Identity-Informed PHD Filtering with Covariance Propagation

Context-aware Bayesian Fusion for Multimodal Object Detection