A Dynamic Network Approach to Inertial Motion Capture

Razavi, Hamid Reza | 2020

576 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 53739 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Alasty, Aria; Salarieh, Hassan
  7. Abstract:
  8. The current study introduces algorithms for inertial motion capture which use data from 9-DOF inertial-magnetic sensor modules to estimate the position and attitude of body links. First, an algorithm is proposed which is capable of IMU calibration without the use of external equipment with less than 0.5% error. Next, extended and unscented Kalman filter-based (EKF and UKF) inertial motion capturing algorithms are introduced that utilize biomechanical constraints in addition to kinematics. In addition to real-time sensor calibration, the algorithms are capable of real-time link geometry estimation, which allows for the imposition of biomechanical constraints without a priori knowledge regarding sensor placements. The stationary phases of the links, during which pseudo-measurements such as zero velocity or heading stabilization updates are applied, are detected using optically trained neural networks with buffered accelerometer and gyroscope data as inputs. Position and attitude error control, using only zero-velocity updates, results in estimation errors to grow during movement. The final part of the study enhances the error control capabilities of the Kalman filter-based inertial motion capture algorithms proposed earlier by multi-stage smoothing. The smoothing process, conducted over the stepping periods of each foot, comprises two stages; Kalman smoothing followed by error minimization by dynamic networks. The performance of the algorithms is experimentally evaluated during a fast-paced walking test using a custom-made inertial motion capture system. Comparison with an optical motion capture system shows that, compared to the EKF, the UKF performs up to 3 times and 40% better in terms of position and attitude estimation root mean square error, respectively. Furthermore, the computationally heavy smoothing and optimization method decreased pelvis position and attitude estimation errors by 19% and 29%, respectively
  9. Keywords:
  10. Sensor Calibration ; Unscented Kalman Smoother ; Inertial Motion Capture ; Dynamic Neural Network ; Unscented Kalman Filter ; Posture

 Digital Object List