Generalization capability of a wearable early morning activity detection system

In this paper, we study the generalization capability of a classifier system which can detect, classify and monitor the activities of daily living for assisting patients with cognitive impairments due to traumatic brain injuries. Generalization implies that the system does not need subject specific training or minimal training, if needed, when the system is deployed in a home setting. We briefly describe the infrastructure of a cost-effective system and show initial applications in detecting activities executed in the early morning. A set of in-home fixed wireless sensors and wearable wireless sensors were used to detect the activity of the user. Both time and frequency-domain features were extracted and used to classify activities using Gaussian Mixture Models post processed with a Majority Voter. We show promising experimental results from 7 subjects while completing washing face, shaving face and brushing teeth activities. We compare results from intra subject classification study with inter subject classification study and show the generalization capability of our wearable system to detect several early morning activities.