In the context of Requirements Analysis, we intend to apply Human Computer Interaction Task Analysis techniques which will be used to gather the relevant information in very close cooperation with physiotherapist professionals and personal trainers. This task will identify the user scenarios, needs and activities in the rehabilitation field considering their workflow, as well as how they use kinematic data during diagnostic, and how can they take advantage of the proposed markerless motion capture system running a biomechanically valid segment orientation algorithm.
Motion capture acquisitions
In this task the movements identified in (Task 1) for exergaming and rehab will be acquired. Several trials will be simultaneously acquired with Kinect One sensors and highend markerbased motion capture system. Kinematic data from the latter system will be considered as ground truth. Several cameras will be used to track reflective markers placed on the subject’s body. The experimental protocol is based on the rigid body reference frames reported in the ISB guidelines for joint coordinate systems.
Anatomical Correction Algorithm for Estimating Segment Orientation
In this task, a skeletal tracking algorithm to estimate the relative orientations of body segments be formulated and implemented. The algorithm must accurately estimate the orientation of body segments in real-time. The only input must be joints and extremities positions of the stick figure captured by the Kinect sensor. By relying on Machine Learning techniques, it is possible to closely estimate the orientation of segments during human movement. Motion capture data from Task 2 will be used to train the machine learning algorithm. The algorithm will build a model that predicts the best orientations values for each anatomical segment.
Computational Model Validation
In order to assess the accuracy of the segment orientation predictions, validation and calibration will be performed with specific examples of the movements stored in the human motion database. Motion capture data from marker-based systems will be considered the ground truth for what is the anatomically correct orientations, whereas markerless mocap (i.e., data from the Kinect) is used as input for segment orientation estimation. Validation of proposed computational procedures will consider the root mean square error as the metric of choice to evaluate the difference between experimental and simulated data: for each time instant of the performed movement, the smaller the error then the better the estimation. Due to the large size of data, “number crunching” to obtain the machine learning model is intended to be performed at UT Austin, as we expect to make use of the supercomputers available at TACC.
Management and Dissemination of Results
The last task is to deal with all management related activities within the project, including the coordination of the four research teams, logistic and contractual tasks, and the production of (nontechnical) project reports resulting of the project’s activities. This project will, if successful, entail a string of follow up research actions to exploit the many research avenues opened by the new approach to skeletal tracking we aim at developing. All software components created within the project will be developed modularly to allow their reuse during the project, and also outside of the project as software libraries. We plan to freely disseminate project experimental data to foster their widest adoption, dissemination and take up by other research groups. Marker and markerless mocap data will be stored in an online human motion database to be hosted on a server at INESC-ID.