Abstract
Indoor localization in GPS-denied environments often relies on Ultra-wideband (UWB) technology, which typically requires at least three anchors for reliable 2D positioning. This paper introduces CT-UIO, a continuous-time localization framework that fuses UWB, IMU, and odometer data using non-uniform B-splines. By implementing adaptive knot-span adjustment and virtual anchor generation, CT-UIO maintains high accuracy in sparse networks with only one or two physical anchors. Experimental results demonstrate significant improvements in both accuracy and computational efficiency compared to state-of-the-art discrete-time and uniform continuous-time methods.
Introduction
Reliable indoor localization remains a significant challenge for mobile robotics. While UWB is a leading solution due to its high precision, its dependence on anchor density is a major drawback. In many practical scenarios—such as long corridors or energy-constrained deployments—maintaining line-of-sight to three or more anchors is often impossible.
Traditional discrete-time (DT) methods struggle with sensor synchronization and lack observability in these "few-anchor" environments. While continuous-time (CT) methods using B-splines address synchronization, they often suffer from computational inefficiency due to uniform knot spacing, which fails to account for varying motion complexity.
Methods
The CT-UIO framework addresses these challenges through three primary innovations: adaptive non-uniform B-splines, virtual anchor (VA) generation, and an adaptive sliding window backend.
1. Adaptive Non-Uniform B-splines
Unlike standard CT methods that use fixed time intervals, CT-UIO dynamically adjusts the density of trajectory control points (knot-spans) based on the robot's velocity and rotational changes. Higher density is allocated during rapid maneuvers or sharp turns to capture complex motion, while density is reduced during steady motion to conserve computational resources.
2. Virtual Anchor (VA) Generation
To overcome the lack of observability in sparse UWB networks, the system generates virtual anchors. By utilizing short-term motion priors from an EKF-fused IMU/odometer model, the system projects the robot's position across time. These projections, combined with UWB ranging measurements, effectively create a multi-anchor geometry from a single physical source, making the system fully observable.
3. Adaptive Sliding Window Backend
The backend performs factor graph optimization within a sliding window. Crucially, the length of this window adapts to the control point density, ensuring real-time performance regardless of the trajectory's complexity.
Key Results
The system was evaluated using a Turtlebot3 in three distinct environments: a corridor, an exhibition hall, and a complex office space. Ground truth was established using a high-accuracy LiDAR-Inertial SLAM system.
- Accuracy: CT-UIO achieved localization accuracies of 0.403m (Corridor), 0.150m (Exhibition Hall), and 0.189m (Office), representing a 15.2% to 26.1% improvement over existing UIO systems.
- Efficiency: The adaptive knot-span strategy allowed the system to operate nearly twice as fast as uniform B-spline methods during slow-motion sequences by optimizing significantly fewer control points.
- Robustness: 90% of the Absolute Positioning Error (APE) remained below ~0.5m even in challenging hybrid motion sequences.
Conclusion
CT-UIO provides a robust solution for localization in sparse UWB environments by leveraging continuous-time trajectory modeling and motion-aware optimization. Future work will focus on the auto-calibration of anchor positions and integrating visual sensors to further enhance robustness against wheel slippage and extreme non-line-of-sight (NLOS) conditions.
Summary updated