Using sensor orientation information for computational head stabilisation in 3d electromagnetic articulography (EMA)

Christian KroosInterspeech 2009

TL;DR

This paper introduces a novel algorithm for computationally stabilizing head movements in 3D Electromagnetic Articulography (EMA) recordings. By effectively leveraging sensor orientation data, the method offers improved accuracy, especially when head tracking sensors are placed close together, and provides a robust solution for scenarios with only two available sensors.

What Problem Does This Paper Address?

Accurately measuring human speech articulation, particularly tongue and lip movements, is crucial for understanding speech production. Electromagnetic Articulography (EMA) is a powerful tool for this, but its measurements are only meaningful within a head-centered coordinate system. Head movements, which are natural and convey prosodic information, can contaminate the articulatory data. Therefore, computationally stabilizing the head post-hoc is essential to isolate true articulatory movements. Traditional methods often rely on position coordinates of multiple sensors, requiring them to be spread out to ensure accurate rotational estimation. This can be problematic if sensors need to be placed in specific, potentially close, locations on the head or if only a limited number of sensors are available. This paper addresses the challenge of achieving accurate head stabilization when sensor placement is constrained or limited.

Methods

The authors propose a three-step algorithm for head stabilization that incorporates sensor orientation information.

  1. Spherical Coordinate Augmentation: Orientation angles (azimuth and elevation) are converted into full spherical coordinates. A unit radius is assigned, effectively treating these as points on a unit sphere.
  2. Cartesian Conversion: These spherical coordinates are then transformed into Cartesian coordinates.
  3. Pose Estimation (PCM): A standard Orthogonal Procrustes algorithm (a least-squares method) is applied to these Cartesian coordinates to estimate the head's rotational parameters. Crucially, since the points lie on a unit sphere, only rotation around the origin is needed, and no translation parameters are determined in this step.

The algorithm also offers a hybrid method (HM) that combines position coordinates (PCM) and orientation angles (OAM). In this hybrid approach, the centroid is calculated from raw position data, and then orientation angles are added to the measured sensor orientation angles before applying the OAM. For the special case of only two sensors, a third virtual sensor's orientation is derived to enable the OAM.

Datasets

The evaluation used a previously recorded dataset for assessing the measurement accuracy of the Carstens AG500 3D EMA system. This dataset involved a custom-built container with three EMA calibration cartridges (each holding 12 sensors) and 8 markers from an optical motion tracking system (Vicon). Various translational, rotational, and free movements of the container were recorded. The paper does not introduce a new benchmark dataset but rather uses existing EMA data for evaluation.

Key Insights and Ideas

  • Orientation is Key for Constrained Placement: The study highlights that sensor orientation information is particularly valuable when head tracking sensors cannot be spread out spatially. This opens up possibilities for more flexible sensor placement.
  • Two Sensors Can Be Enough: The proposed algorithm provides a robust method for head stabilization even when only two sensors are available, a scenario that is challenging for purely position-based methods.
  • Smoothness as a Quality Metric: The authors introduce the integral of the squared second derivative of motion parameters as a measure of smoothness, arguing that it's a better indicator of stabilization quality than residual error alone, as it's less susceptible to mistracking of individual sensors.
  • OAM's Independence from Sensor Distance: Unlike PCM, the accuracy of OAM for rotational estimation is not dependent on the spatial distance between sensors, only on the spread of their orientation angles.

Key Results

  • OAM Outperforms PCM in Smoothness: For sensor distances within the studied range (33.9mm to 79.4mm), the Orientation Angle Method (OAM) consistently produced smoother motion parameters compared to the Position Coordinate Method (PCM).
  • Hybrid Method (HM) Falls Between: The Hybrid Method (HM), combining both position and orientation, generally performed between PCM and OAM in terms of smoothness.
  • No Difference Between 3 and 2 Sensors for HM: Remarkably, the HM showed no significant difference in performance whether using three or two sensors.
  • RMS Error Corroborates Smoothness: The Root Mean Square (RMS) error analysis generally supported the findings from the smoothness analysis, indicating that OAM provided more reliable results.

Limitations and Future Work

  • Limitations: The paper acknowledges that while OAM performs well when sensors are close, it relies on the spread of orientation angles. If all sensors have very similar orientations, OAM would break down. The study focused on a specific EMA system (Carstens AG500), and generalizability to other systems might require further validation.
  • Future Work: The authors suggest that this orientation-based approach has significant potential for accurately determining jaw movements. Since jaw sensors are often limited in placement and cannot be widely separated, OAM could be crucial for obtaining reliable jaw motion parameters, enabling better analysis of jaw-tongue coordination. They also note that the position of the condyle could be factored in later as the center of rotation for jaw stabilization.

Summary updated

BibTeX
@inproceedings{kroos2009using,
  author = {Christian Kroos},
  title = {{Using sensor orientation information for computational head stabilisation in 3d electromagnetic articulography (EMA)}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure