SODA-CitrON: Static Object Data Association by Clustering Multi-Modal Sensor Detections Online
New unsupervised ML method solves persistent robotics problem of tracking static objects from intermittent sensor data.
A team of researchers has introduced SODA-CitrON (Static Object Data Association by Clustering Multi-Modal Sensor Detections Online), a novel algorithm addressing a fundamental challenge in robotics and autonomous systems: accurately fusing and tracking static objects from intermittent, heterogeneous sensor data. Unlike classical approaches like Joint Probabilistic Data Association (JPDA) that excel with dynamic targets but struggle with static objects, this unsupervised machine learning method operates fully online, handling temporally uncorrelated measurements from multiple sensors while simultaneously estimating object positions and maintaining persistent tracks for an unknown number of objects. The work, submitted to the 2026 International Conference on Information Fusion, represents a significant step forward in environmental mapping for autonomous vehicles and robots.
The technical breakthrough lies in SODA-CitrON's combination of efficiency and performance. It maintains worst-case loglinear computational complexity relative to the number of sensor detections, making it scalable for real-time applications, while also providing full output explainability—a critical feature for safety-critical systems. In comprehensive Monte Carlo simulations, the algorithm consistently outperformed state-of-the-art methods including Bayesian filtering, DBSTREAM clustering, and JPDA across multiple metrics: F1 score (measuring accuracy), position Root Mean Square Error (RMSE), Multiple Object Tracking Precision (MOTP), and Multiple Object Tracking Accuracy (MOTA). This superior performance in static object mapping scenarios suggests immediate applications for improving the perception systems of self-driving cars, delivery robots, and automated mapping platforms that must reliably distinguish permanent environmental features from transient clutter.
- Outperforms Bayesian filtering, DBSTREAM, and JPDA in F1 score, RMSE, MOTP, and MOTA metrics for static object mapping
- Operates with worst-case loglinear complexity for scalable, real-time processing of sensor data
- Provides full output explainability while handling unknown numbers of objects with intermittent, multi-sensor detections
Why It Matters
Enables more reliable environment mapping for autonomous vehicles and robots, directly improving safety and operational capability in complex real-world settings.