A robust method to improve estimation accuracy of walking gait kinematics by considering geometrical constraints on kinect data

In this paper, we present a robust post processing method to improve the accuracy of kinematics information of human walking gait obtained from the Kinect sensor to be used for home-based gait analysis purposes. The accuracy of raw skeleton tracking data provided by Kinect suffers from a considerable level of uncertainty that compromises any reliable motion analysis. To address this issue, we have developed a comprehensive framework that reconstructs the joint trajectories from the Kinect's uncertain measurements. The proposed algorithm detects valid motion periods as well as valid segments that represent starting and ending points of fully observed walking gait cycles. It then estimate the skeleton parameters based on the data within these valid periods. The variations of the estimated parameters is significantly reduced when only the data within valid periods are used. Moreover, by considering human motor control principles, the orientation of each limb is filtered through a 5^th order polynomial fitting algorithm (Savitzky-Golay). This process removes sudden jumps/deviations which are inconsistent with human motor control. This fitting process along with the estimated skeleton parameters as geometrical constraints are used to reconstruct the joint trajectories. The experimental results demonstrate higher repeatability and less dispersion of the reconstructed joint trajectories compared to the raw skeleton information.