Master thesis presentation: Development of a motion tracking algorithm using a non-depth sensing camera to assist in rehabilitation of stroke and COPD patients

On 17 August 2016, Henrik Fransson and Tobias Petrén, MPBME, will be presenting their master thesis with the title “Development of a motion tracking algorithm using a non-depth sensing camera to assist in rehabilitation of stroke and COPD patients”.

When? Wednesday, August 17 at 14.00
Where? Blå Rummet (room 6414), Hörsalsvägen 11, 6th floor

Examiner: Fredrik Kahl

Abstract
Stroke and chronic obstructive pulmonary disease (COPD) are two common diseases and the patients suffering from them are in need of rehabilitation to improve their life quality. In order to facilitate the rehabilitation process, motion controlled computer games can be used, as studies suggest. This thesis involves the development of a tracking algorithm, used in such games, able to track the hands and head of a person in real time with a non-depth sensing camera, such as those found in laptops and tablets.

The resulting algorithm consists of four main parts: background subtraction, skin extraction using the RGB and HSV color spaces, classification of hands and head with a convolutional neural network (CNN) and tracking of the classified body parts with a Kanade-Lucas-Tomasi (KLT) tracker. Running the algorithm on video recordings obtained from a common motion pattern shows that the algorithm is able to correctly track the three body parts in approximately 88.9% of the 800 frames recorded at 30 FPS and the computation time is roughly 0.48 seconds per frame. The algorithm is able to recover from a situation in which not all body parts are tracked correctly and it can handle occlusions. Furthermore, the algorithm needs some preconditions, such as good lightning and no skin-colored background or clothes, to fully function.

It is concluded that the current implementation is too slow to function in real time. However, it is believed that the method used in the algorithm can be a viable approach for motion controlled games if implemented in a faster language than MATLAB and if further improvements are done to discard the currently required preconditions.