Adapting computer vision-based marker-less pose estimation systems for rehabilitation applications through transfer learning

Abstract

The need for telerehabilitation services has become more evident in recent years, and new developments in computer vision have shown promise, providing evidence that these services may become a reality in the future. Even though the current solutions show a lot of promise, there are still improvements to be made before they can be used in the real world. This thesis is based on a markerless computer vision-based prototype for measuring the angle in the knee joint using, that was developed at Arcada University of Applied Science. The prototype showed some significant problems with keypoint accuracy in some positions, namely, the lying position with one leg bent. The goal of this thesis is to answer the question of whether the improvement of the accuracy of specific positions in a markerless computer vision-based prototype for rehabilitation through transfer learning is possible. The improvement of the prototype is focused on improving the underlying machine learning model that the prototype uses, in order to improve the accuracy for the positions that the current prototype struggles with. The thesis starts by first going through related research and existing solutions using computer vision for rehabilitation. The text explains the problems with the previous prototype and existing solutions, as well as the need for a new prototype. In conjunction with the explanation of the new prototype that was developed, the problems with the accuracies of some positions are explained. The need for a new dataset with the specific aim of improving the lying position through transfer learning is brought up as a solution. The results of the training are then analyzed and explained in detail, showing the difference in the original and improved accuracies, as well as the implications of a lack of data. The last part iterates the conclusions of this thesis including the problems that were found and the possible solutions to these problems as well as the future of the prototype.