Enhancing motion tracking accuracy of a low-cost 3D video sensor using a biomechanical model, sensor fusion, and deep learning

• Published in: Frontiers in rehabilitation sciences 2022-08, Vol.3, p.956381
• Main Authors: Agami, Shahar, Riemer, Raziel, Berman, Sigal
• Format: Article
• Language: English
• Subjects: kinematics; recurrent neural network (RNN); rehabilitation; Rehabilitation Sciences; upper limb; virtual reality
• ISSN: 2673-6861; 2673-6861
• DOI: 10.3389/fresc.2022.956381

Low-cost 3D video sensors equipped with routines for extracting skeleton data facilitate the widespread use of virtual reality (VR) for rehabilitation. However, the accuracy of the extracted skeleton data is often limited. Accuracy can be improved using a motion tracker, e.g., using a recurrent neural network (RNN). Yet, training an RNN requires a considerable amount of relevant and accurate training data. Training databases can be obtained using gold-standard motion tracking sensors. This limits the use of the RNN trackers in environments and tasks that lack accessibility to gold-standard sensors. Digital goniometers are typically cheaper, more portable, and simpler to use than gold-standard motion tracking sensors. The current work suggests a method for generating accurate skeleton data suitable for training an RNN motion tracker based on the offline fusion of a Kinect 3D video sensor and an electronic goniometer. The fusion applies nonlinear constraint optimization, where the constraints are based on an advanced shoulder-centered kinematic model of the arm. The model builds on the representation of the arm as a triangle (the arm triangle). The shoulder-centered representation of the arm triangle motion simplifies constraint representation and consequently the optimization problem. To test the performance of the offline fusion and the RNN trained using the optimized data, arm motion of eight participants was recorded using a Kinect sensor, an electronic goniometer, and, for comparison, a passive-marker-based motion tracker. The data generated by fusing the Kinect and goniometer recordings were used for training two long short-term memory (LSTM) RNNs. The input to one RNN included both the Kinect and the goniometer data, and the input to the second RNN included only Kinect data. The performance of the networks was compared to the performance of a tracker based on a Kalman filter and to the raw Kinect measurements. The accuracy of the fused data was high, and it considerably improved data accuracy. The accuracy for both trackers was high, and both were more accurate than the Kalman filter tracker and the raw Kinect measurements. The developed methods are suitable for integration with immersive VR rehabilitation systems in the clinic and the home environments.