Automatic range-of-motion calculation using machine learning and pose estimation, for online physical rehabilitation

Abstract

This thesis examines the feasibility of automatically calculating the range-of-motion of a person's legs from user-recorded videos. Automating this process would significantly benefit online rehabilitation platforms, such as the one in development by MoveUP. To achieve this, we have first performed a feasibility study, examining the impact of different datasets, state-of-the-art networks and specific training techniques. By doing this, we were able to generate pose estimations with an accuracy of up to 96.66\% on our own dataset; out-performing other state-of-the-art networks, using a Stacked Hourglass network, 4-stacks deep, trained on our synthetic and custom dataset through transfer learning. This achievement has been made possible by the fact that our network is limited to recognizing only a small pose space, namely only one specific range-of-motion exercise. Afterwards, we examined possible errors still returned by the network, and defined ways in which these can be filtered or corrected. From this, we are able to generate range-of-motion estimations that are correct within 5 to 10 degrees, which is sufficient for our application. Although, some shortcomings still remain. For example, because of the limited size of our dataset the network can have trouble in videos with bad composition, and the final network is fairly heavy on system resources, making it unsuitable for running on mobile devices.