Towards a Virtual Coach for manual wheelchair users

Abstract

We introduce the concept of a Virtual Coach (VC) for providing advice to manual wheelchair users to help them avoid damaging forms of locomotion. The primary form of context for this system is the user's propulsion pattern. The contexts of self vs. external propulsion and the surface over which propulsion is occurring can be used to improve the accuracy of the system's propulsion pattern classifications. To obtain these forms of context, we explore the use of both wearable and wheelchair-mounted accelerometers. We show achievable accuracy rates of up to 80-90% for all desired contextual information using two common machine learning techniques: k-nearest neighbor (kNN) and support vector machines (SVM).