Identification of Noise Covariance Matrices to Improve Orientation Estimation by Kalman Filter

Abstract

Magneto-inertial measurement units (MIMUs) are a promising way to perform human motion analysis outside the laboratory. To do so, in the literature, orientation provided by an MIMU is used to deduce body segment orientation. This is generally achieved by means of a Kalman filter that fuses acceleration, angular velocity, and magnetic field measures. A critical point when implementing a Kalman filter is the initialization of the covariance matrices that characterize mismodelling and input error from noisy sensors. The present study proposes a methodology to identify the initial values of these covariance matrices that optimize orientation estimation in the context of human motion analysis. The approach used was to apply motion to the sensor manually, and to compare the orientation obtained via the Kalman filter to a measurement from an optoelectronic system acting as a reference. Testing different sets of values for each parameter of the covariance matrices, and comparing each MIMU measurement with the reference measurement, enabled identification of the most effective values. Moreover, with these optimized initial covariance matrices, the orientation estimation was greatly improved. The method, as presented here, provides a unique solution to the problem of identifying the optimal covariance matrices values for Kalman filtering. However, the methodology should be improved in order to reduce the duration of the whole process.

A Full-State Robust Extended Kalman Filter for Orientation Tracking During Long-Duration Dynamic Tasks Using Magnetic and Inertial Measurement Units

“…Adaptive gain tuning of filters Nez et al [32] showed that the filter gains must be selected carefully, or the filter performance could deteriorate. Also, Nez et al [32] showed that the values of the gains identified through an optimization were different from those obtained by the Allan Variance method [33]. The reason could be that the optimized gains implicitly account for error sources such as modelling or calibration [32].…”

“…Also, Nez et al [32] showed that the values of the gains identified through an optimization were different from those obtained by the Allan Variance method [33]. The reason could be that the optimized gains implicitly account for error sources such as modelling or calibration [32]. Therefore, to conduct a fair comparison and provide reliable results, we optimized the gains of all filters, according to [28].…”

Improved Attitude and Heading Accuracy with Double Quaternion Parameters Estimation and Magnetic Disturbance Rejection

“…Based on the error calculated in Equation ( 39 ), a rule was created to determine the error condition for switching to different values of the measurement error covariance. The covariance value of the measurement error indicates the degree to which the system relies on the gyroscope prediction or magnetometer data to estimate the heading [ 24 ]. If a severe magnetic disturbance is detected, a big measurement error covariance value is selected to prevent the effect of erroneous sensor data on the estimate.…”

Motion Capture Systems Using Optimal Signal Processing Algorithm: A State-of-the-art Literature

“…In this case, the extended KF can be used for estimating the three-dimensional orientation of a rigid body [18,19]. Furthermore, the multi view human pose estimation process is described in [20][21][22]. For determining heart rate, 2 Motion Capture Systems Using Optimal Signal Processing Algorithm: A State-of-the-art Literature a KF algorithm is used in [23,24].…”