TL;DR
This paper investigates how different inertial sensor calibration methods impact the performance of various Kalman filter (KF) algorithms for estimating body segment angles. They found that more complex calibrations consistently lead to better results, and cooperative KFs (which fuse data from multiple sensors) generally outperform local KFs, especially when using higher-quality calibration data.
What Problem Does This Paper Address?
Estimating human body segment angles accurately is crucial for applications like gait analysis and exoskeleton control. Inertial Measurement Units (IMUs), composed of accelerometers and gyroscopes, are widely used for this purpose. However, low-cost IMUs suffer from inherent errors like bias and scale factors due to manufacturing imperfections and self-heating. While sensor fusion techniques, particularly Kalman Filters (KFs), are standard for mitigating these errors, and calibration methods exist to correct sensor readings, the interaction between the quality of calibration and the type of KF algorithm has not been thoroughly studied. This paper aims to bridge this gap by systematically comparing different calibration routines with local and cooperative KF approaches to understand their combined effect on orientation estimation accuracy.
Methods
The study compares three KF algorithms:
- Local KF: A standard KF that uses data from a single IMU.
- Matricial KF: A cooperative KF that fuses data from all IMUs attached to the body.
- Markovian KF: Another cooperative KF that fuses IMU data with information from exoskeleton potentiometers.
Two calibration methods were employed:
- Simplified Calibration: A basic method replicable in most labs, involving static accelerometer calibration against gravity and 90-degree gyroscope rotations.
- Complex Calibration: A more sophisticated approach, similar to commercial vendor methods, using a high-precision turntable for gyroscope calibration across various angular velocities and accounting for temperature effects.
The performance of these combinations was evaluated using walking data from a healthy subject wearing a 6-DoF exoskeleton. The Root Mean Square Error (RMSE) of the estimated absolute segment angles (trunk, thigh, shank, foot) against a reference optoelectronic system was the primary metric. A three-way mixed ANOVA was used for statistical analysis.
Datasets
The study used walking data collected from a single healthy subject (33 years old, 1.75m, 82kg) walking on a treadmill at 1.5 km/h for 60 seconds while wearing an exoskeleton. The data was synchronized from an optoelectronic system (BTS Bioengineering), exoskeleton potentiometers (Technaid S.L.), and inertial sensors (Technaid S.L.). The optoelectronic system served as the ground truth.
Key Insights and Ideas
- Calibration Quality is Paramount: The study strongly emphasizes that the quality of sensor calibration has a significant impact on the accuracy of orientation estimation, regardless of the KF algorithm used. More complex calibration routines consistently yield better results.
- Cooperative Fusion Offers Advantages: Cooperative KFs (Matricial and Markovian) generally outperform the Local KF, especially when using higher-quality calibration data. This highlights the benefit of leveraging information from multiple sensors.
- Markovian KF's Robustness: The Markovian KF, which integrates data from both IMUs and exoskeleton potentiometers, showed robust performance across different calibration qualities. It was particularly effective when calibration was less ideal or unknown, suggesting its value in real-world scenarios where calibration might be compromised.
- Matricial KF's Sensitivity to Calibration: The Matricial KF, which fuses data from all IMUs, demonstrated the largest performance gain when moving from simplified to complex calibration. This suggests it is highly sensitive to the quality of the input sensor data.
- Segment-Specific Performance: The trunk segment was identified as the most challenging to estimate accurately, likely due to fewer quasi-static moments for accelerometer updates.
Key Results
- RMSE Reduction: The complex calibration consistently resulted in lower RMSE values compared to the simplified calibration across all filters and segments.
- Overall Performance: The Markovian KF generally performed best, with average RMSEs below 1.22 degrees when using complex calibration. The Matricial KF also performed similarly well under complex calibration.
- Local KF Limitations: The Local KF showed the poorest performance, especially for the trunk segment, with RMSEs as high as 4 degrees.
- Statistical Significance: A three-way mixed ANOVA confirmed significant main effects for calibration, filter, and segment. A significant interaction between filter and calibration was also observed.
- Trunk Segment: The trunk segment exhibited significantly worse performance than other segments, with the Markovian KF showing a significant advantage over the Matricial KF when considering only the filter effect (ignoring calibration).
Limitations and Future Work
- Single Subject Data: The study was conducted on data from a single subject, which limits the generalizability of the findings.
- Controlled Environment: The data was collected during a controlled walking trial on a treadmill. Performance in more dynamic or varied activities (e.g., reaching, grasping, object manipulation) remains to be explored.
- Future Work: The authors suggest incorporating in-field calibrations, evaluating performance on more complex tasks, and exploring cooperative KFs with more explicit physical models that describe the relationships between IMUs. They also propose evaluating cooperative KFs with a more explicit physical model that declares the relation between the IMUs.
Summary updated