eWatch: Virtual Coach for Wheelchair Users

This project addresses the creation of a virtual coach for manual wheelchair users and deals with basic research in areas such as sensing, logging and analyzing contextual data, user engagement techniques, and context aware reminders.

We developed a manual wheelchair classification system which recognizes different patterns using our wrist mounted eWatch accelerometer (see top image at right). There are four distinct propulsion patterns that wheelchair users tend to follow – semicircular (SEMI), single loop over (SLOP), double loop over (DLOP) and arcing (ARC) which have been identified in a limited user study. We explored classification of propulsion patterns using two common machine learning algorithms, k-Nearest Neighbor (kNN) and Support Vector Machines (SVM) with a Radial Basis Function (RBF) kernel.

It can be seen that the classification accuracy tends to be higher on surfaces with higher resistance (dynamometer, low carpet, asphalt), when compared to surfaces with low resistance (tile). Classification accuracy for arcing was considerably lower than the other propulsion patterns. Namely, the arcing is a subset of each of the other patterns, and hence, is most susceptible to misclassification (see bottom image at right).

In the next study we used both wrist mounted and wheelchair-mounted accelerometers to identify and evaluate the following three contexts: propulsion pattern, self vs. external propulsion, and surface type.