Accurate estimation of joint motion trajectories for rehabilitation using Kinect

Kinect as an effective tool for clinical assessment and rehabilitation, suffers from drawbacks of lower accuracy of measuring human body kinematic data when compared to clinical gold standard motion capture devices. The accuracy of time-varying 3D locations of a fixed number of body joints obtained from Kinect skeletal tracking utility is affected by the presence of noise and precision limits of the Kinect depth sensor. In this paper, a framework for improving accuracy of Kinect skeletal tracking is proposed, that uses a set of parametric models to represent and track the human body. Each of the models represents the 3D geometric properties of a body segment connecting two adjacent joints. The temporal trajectories of the joints are recovered via particle filter-based motion tracking of each model. The proposed method was evaluated on Active Range of Motion exercises by 7 healthy subjects. The joint motion trajectories obtained using the proposed framework exhibit a greater motion smoothness (by 36%) along with reduced coefficient of variation of radius (by 34%), and lower value of root-mean-squared-error (by 53%), when compared to Kinect joint trajectories. This indicates an improvement in accuracy of joint motion trajectories using Kinect device, rendering it more suitable for clinical assessment and rehabilitation.