Decentralized Fusion of 3D Extended Object Tracking based on a B-Spline Shape Model

TL;DR

The paper addresses robust, scalable perception for multi-sensor autonomous systems by tackling decentralized fusion for 3D extended object tracking (EOT). It introduces a B-spline side-view profile extruded into 3D to model object shape, employing an EKF-based state estimator with a CTRV/CV motion model and a level-set inspired pseudo-measurement model. Covariance Intersection (CI) is adopted to fuse independent trackers under unknown cross-covariances, with design choices to mitigate orientation and shape parameter fusion issues. The approach is validated on CARLA simulations and the tumtraf real dataset, showing that CI-based decentralized fusion improves tracking when one sensor has an unfavorable perspective, while approaching centralized performance in favorable conditions. This yields a scalable, robust multi-sensor EOT framework for traffic scenarios.

Abstract

Extended Object Tracking (EOT) exploits the high resolution of modern sensors for detailed environmental perception. Combined with decentralized fusion, it contributes to a more scalable and robust perception system. This paper investigates the decentralized fusion of 3D EOT using a B-spline curve based model. The spline curve is used to represent the side-view profile, which is then extruded with a width to form a 3D shape. We use covariance intersection (CI) for the decentralized fusion and discuss the challenge of applying it to EOT. We further evaluate the tracking result of the decentralized fusion with simulated and real datasets of traffic scenarios. We show that the CI-based fusion can significantly improve the tracking performance for sensors with unfavorable perspective.

Figures (5)

• Figure 1: The concept of decentralized fusion of 3D EOT using the model of extruding a side-view profile represented by a B-spline curve. The estimation 1 and estimation 2 are computed locally. The fusion is performed using CI. The ground truth state of the object is shown together with its mesh. The points in the right image represent the control points of the B-spline curve.
• Figure 2: The test scenario in CARLA with a left turn drive and the results. The color blue, red and yellow indicates centralized fusion, sensor 2, and decentralized fusion, respectively.
• Figure 3: The test scenario in the real word dataset (sequence R4_S4) in tumtraf.
• Figure 4: Exampel result of real-world dataset evaluation. Red, yellow, blue results are from infrastructure sensor tracking, decentralized tracking and centralized tracking respectively. The ground truth bounding box is shown in black.
• Figure 5: The evaluation result using the TUMTraf dataset. The colors blue, red, and yellow indicate centralized tracking, tracking with infrastructure sensor, and decentralized tracking, respectively. Vehicle 2 enters the blind zone of the infrastructure sensor after $6 \ s$, so the tracking stops.