Automation enhancement and accuracy investigation of a portable single-camera gait analysis system

Yang, C. and Ugbolue, U. C. and McNicol, D. and Stankovic, V. and Stankovic, L. and Kerr, A. and Carse, B. and Kaliarntas, K. and Rowe, P. J. (2019) Automation enhancement and accuracy investigation of a portable single-camera gait analysis system. IET Science, Measurement and Technology, 13 (4). pp. 563-571. ISSN 1751-8822

[img]
Preview
Text (Yang-etal-SMT2019-Automation-enhancement-and-accuracy-investigation-of-a-portable)
Yang_etal_SMT2019_Automation_enhancement_and_accuracy_investigation_of_a_portable.pdf
Accepted Author Manuscript

Download (1MB)| Preview

    Abstract

    While optical motion analysis systems can provide high-fidelity gait parameters, they are usually impractical for local clinics and home use, due to high cost, requirement for large space, and lack of portability. In this study, the authors focus on a cost-effective and portable, single-camera gait analysis solution, based on video acquisition with calibration, autonomous detection of frames-of-interest, Kalman-filter + structural-similarity-based marker tracking, and autonomous knee angle calculation. The proposed system is tested using 15 participants, including 10 stroke patients and 5 healthy volunteers. The evaluation of autonomous frames-of-interest detection shows only 0.2% difference between the frame number of the detected frame compared to the frame number of the manually labelled ground truth frame, and thus can replace manual labelling. The system is validated against a gold standard optical motion analysis system, using knee angle accuracy as metric of assessment. The accuracy investigation between the RGB- and the greyscale-video marker tracking schemes shows that the greyscale system suffers from negligible accuracy loss with a significant processing speed advantage. Experimental results demonstrate that the proposed system can automatically estimate the knee angle, with R-squared value larger than 0.95 and Bland-Altman plot results smaller than 3.0127° mean error.