Kalman filter based on marker and force plate measurements and a full multibody model, for the real–time capture, reconstruction and analysis of human movement

Abstract Currently, optical motion capture remains the gold standard for human motion analysis. This technique estimates the movement of a subject by tracking a set of markers attached to their skin, then using the captured trajectories to reconstruct the movement of an underlying rigid–body model. However, since the markers are not rigidly fixed to the skeleton, their motion relative to the bones can induce significant estimation errors, especially when using the computed accelerations to calculate the joint torques by solving the inverse dynamics. This paper presents an extended Kalman filter aimed at reducing such errors, through the incorporation of a complete multibody model of the human body. The method builds upon an existing Kalman filter based on a purely kinematic plant model, which uses only optical markers as sensors. The proposed observer adds the motor efforts and external reactions to the system states, and incorporates force plates into the set of sensors. This allows the measured ground reactions to directly affect the dynamics, rather than remaining an input to a subsequent inverse dynamics. The dynamic model can be added to the Kalman filter with multiple options and simplifications, which are here examined to find the combination with the best balance between efficiency and accuracy. The performance of the selected implementation is then compared with the existing kinematics–based observer. The results show that the proposed method improves the estimation of joint torques and reduces the influence of marker bounces on them, in exchange for a higher computational cost and an increased estimation delay.