A hybrid Bayesian approach is proposed for fusing ultrasonic range-based and IMU-based sourceless position estimates under conditions of varying observability. The method selectively switches between Single-Hypothesis-Tracking (an EKF) and Multi-Hypothesis-Tracking (a particle filter with Gaussian mixture likelihoods) based on how many range measurements are available at any given moment. Results show up to 10% improvement over SCAAT and 24% improvement over a standalone EKF for intermittent 3-second occlusions.
When the Ultrasonic Signal Goes Dark
In a hybrid motion capture system for rehabilitation, ultrasonic transmitters are fixed at known positions and wearable receivers measure time-of-flight ranges to compute 3D position. When a patient moves, body limbs can occlude the direct path between transmitter and receiver. For short periods, only one or two ranges are available instead of the three needed for full triangulation.
This is called partial observability and it is the key challenge this paper addresses. With sufficient ranges, a standard EKF correction step handles things well. But with one or two ranges, a single Gaussian is a poor model for the true uncertainty. When you only have two ranges, the most likely position lies somewhere on the circumference of a circle. That is a ring-shaped probability mass, not a blob. A Gaussian centred on the ring midpoint and spread wide enough to cover the ring wastes probability mass in places that are geometrically impossible.
The core insight: partial observability produces multimodal, non-Gaussian uncertainty. Using a Gaussian mixture to model this via a particle filter captures the true ring-shaped distribution, while switching back to an EKF when full observability returns keeps computational cost down.
What Happens During Real Exercises
The paper identifies five distinct occlusion patterns that arise when a sensor unit on a limb undergoes rehabilitation exercise motions.
Switching Between Single and Multi Hypothesis
The correction step of the filter has three components: a Single-Hypothesis-Tracking (SHT) unit based on an EKF, a Multi-Hypothesis-Tracking (MHT) unit based on a Sequential Importance Sampling particle filter, and an information exchange unit that passes mean and covariance between the two when switching.
The switching decision is based on the rank of the observation matrix. When the rank is sufficient (three or more ranges), the state is fully observable and the SHT unit runs. When it falls below the threshold, the MHT unit takes over. The information exchange unit ensures that the covariance information built up in the SHT phase seeds the particle cloud for MHT, which means fewer particles are needed to achieve the same accuracy.
Performance Under Occlusion
Simulated on linear motion, random roto-translation and upper-arm flexion-extension over 10 seconds at 50 Hz IMU rate and 30 Hz range rate. Two ranging error levels tested: 1 cm and 5 cm.
| Scenario | Proposed (cm) | SCAAT (cm) | EKF (cm) | DPF (cm) | vs SCAAT |
|---|---|---|---|---|---|
| 1 (full LOS) | 3.02 | 3.02 | 3.06 | 5.07 | 0% |
| 2 | 3.05 | 3.05 | 3.12 | 5.05 | 0% |
| 3 | 3.20 | 3.33 | 3.41 | 5.30 | 4% |
| 4 | 3.24 | 3.40 | 4.18 | 5.40 | 4.8% |
| 5 (worst occlusion) | 3.70 | 4.10 | 4.83 | 5.90 | 10% |
The computational advantage is equally notable. The hybrid approach needs far fewer particles than a conventional SIS filter to achieve the same accuracy, because the EKF phase provides a tight prior for particle seeding during MHT. At 1000 particles, the hybrid achieves 2.57 cm RMS against 3.74 cm for SIS.
Probabilistic Reasoning Under Physical Constraints
The problem addressed here is more general than it sounds. Whenever you are fusing a primary ranging system with a secondary dead reckoning system, you face the question of what to do when the primary system partially fails. Simply falling back to dead reckoning ignores whatever partial information remains from the one or two surviving ranges. This work shows that Bayesian filtering with the right likelihood model can extract value from incomplete observations.
The Gaussian mixture approach to ring-shaped uncertainty also points toward a broader principle: uncertainty models should respect the geometry of the measurement. One range measurement places you on a sphere. Two ranges place you on a circle. The appropriate prior is shaped like that constraint, not like a blob.
Place in the PhD research arc
This paper addresses the sensor fusion layer of the wearable rehabilitation system. The dual IMU dead reckoning work from 2011 improved the sourceless subsystem. The rehabilitation system architecture paper from 2010 defined what the system needed to do for stroke patients. This 2012 paper connects them: it answers how to combine range and IMU data in a principled way when range availability fluctuates during natural movement. The 2014 work on magnetic distortion then addressed orientation accuracy, the remaining major error source that dead reckoning does not fix.
DOI: 10.1109/IS.2012.6335119
UCD Research Repository: hdl.handle.net/10197/3965