Evaluating Vision Based Human Pose Estimation for use in Gait Assessment of Older Adults with Dementia

Becky Ng (1, 2), Babak Taati (1, 2, 3, 4)

1) Toronto Rehabilitation Institute, 2) Institute of Biomaterials and Biomedical Engineering, University of Toronto, 3) Department of Computer Science, University of Toronto, 4) Vector Institute

1. Background
Falls are detrimental to older adults, leading to injury and, in many cases, mortality [1]. Fall risk is increased in older adults with dementia due to the individual’s inability to self-monitor their decline in mobility [2]. One way to reduce this risk is to perform gait assessments and implement mobility aids as required [3]. In today’s healthcare setting its infeasible for clinicians to perform gait assessments often enough to effectively reduce fall risk. This study aims to build on the many recent advances in computer vision, deep learning and, specifically, human pose estimation, by applying these methods for gait assessment. This study aims to correlate parameters extracted from the pose estimation data with clinical gait assessments.

2. Methods
Patient data has been collected by recording video of elderly dementia patients on a daily basis in their long term care facility. Pose estimation algorithms has been run over the videos in post-processing, and gait parameters are being extracted from the resulting sequence of joint and limb locations. A regression model will be developed to correlate the extracted gait parameters with a clinical gait assessment protocol which is performed on each study participant by a trained clinician.

3. Conclusions
By developing a method which uses pose estimation to continuously, unobtrusively and automatically estimate clinical gait assessments, it may be possible to create a warning system which, if implemented, could reduce falls. Furthermore, if successful, this technology could be adapted for use in any multitude of applications which require not only mobility, but any movement assessment.