A novel bio-kinematic encoder for human exercise representation and decomposition - Part 2: Robustness and optimisation

Bio-kinematic characterisations of human exercises constitute dealing with parameters such as velocity, acceleration, joint angles, etc. A majority of these are measured directly from various sensors ranging from RGB cameras to inertial sensors. However, due to certain limitations associated with these sensors, such as inherent noise, filters are required to be implemented to subjugate the effect from the noise. When the two-component (trajectory shape and dynamics) bio-kinematic encoding model is being established to represent an exercise, reducing the effect from noise embedded in raw data will be important since the underlying model can be quite sensitive to noise. In this paper, we examine and compare some commonly used filters, namely least-square Gaussian filter, Savitzky-Golay filter and optimal Kalman filter, with four groups of real data collected from Microsoft Kinect©, and assert that Savitzky-Golay filter is the best one when establishing an underlying model for human exercise representation.