(K-439) Calibration Algorithms for the Realignment of Randomly Oriented IMUs

Inertial measurement units (IMUs) are commonly used to determine joint angles as part of control schemes for assistive devices, including delivering neural stimulation to activate muscles after paralysis [1].

IMUs are often aligned to the body coordinate system to limit necessary calibration [2]. For implanted systems, the placement of the IMUs are subject to surgical constraints and cannot be guaranteed to align with the body’s axes. Misaligned IMUs could result in arbitrary joint angle calculations without a calibration process to align the coordinate systems to the body reference frame. People with paralysis may not be able to complete involved methods that are used to calibrate some IMU systems. As a result, methods that incorporate simple motions are needed to realign IMUs to anatomical axes for implanted systems regardless of original orientation. This project developed and evaluated two calibration approaches.

Two calibration methods, using either static postures or dynamic movements, were evaluated. Seven IMUs were placed in arbitrary locations and orientations on the trunk, thighs, shanks, and ankles, along with a set of VICON markers to capture lower limb angles. IMU orientation with respect to gravity was recorded when standing upright, and was used to realign the IMUs to the body’s inferior-superior axis. This first step was used in both calibration methods. 

The first method (static calibration) relies on accelerometer data recorded in a seated position with limbs rotated about the mediolateral axis. The measured acceleration due to gravity was used to realign the IMUs to the mediolateral and anteroposterior axes. 

The second method (dynamic calibration) utilized gyroscope recordings during movements about the mediolateral axis for all lower limb joints. The measured angular velocity was used to calculate the angle of rotation needed to realign the IMUs to the mediolateral and anteroposterior axes.

Joint angles were calculated for flexion/extension of the hip, knee, and ankle after offline calibration using the two methods. Calculated values were compared to VICON values during trials with overground walking, sit to stand transitions, and stand to sit transitions. Average RMSE was calculated across all timepoints.

The calibrated joint angles were compared to VICON motion capture for the two methods, and average RMSE for each joint angle across six trials was calculated. As shown in Table 1, both algorithms predicted joint angles within less than ten degrees of error; after accounting for an offset, average error was less than five degrees for each movement. These results suggest either calibration method could generate useful angles that could be incorporated into a real-time control system for an implanted neuroprosthesis.

This work has been supported by NIH grant NICHD R01 HD105008-01A1 and the APT Wen Ko and DEI Summer Internship Program.

1. Friederich, Aidan R., et al. “Trunk Posture from Randomly Oriented Accelerometers.” Sensors, vol. 22, no. 19, 2022, p. 7690.

2. Vitali, Rachel V., and Noel C. Perkins. “Determining Anatomical Frames via Inertial Motion Capture: A Survey of Methods.” Journal of Biomechanics, vol. 106, 2020, p. 109832.