In this thesis, an activity classification algorithm is developed to support a human ambulatory monitoring system. This algorithm, to be deployed on an e-textile garment, represents the enabling step in creating a wide range of garments that can use the same classifier without having to re-train for different sensor types. This flexible operation is made possible by basing the classifier on an abstract model of the human body that is the same across all sensor types and subject bodies. In order to support low power devices inherent for wearable systems, the algorithm utilizes regular expressions along with a tree search during classification.

To validate the approach, a user study was conducted using video motion capture to record subjects performing a variety of activities. The subjects were randomly placed into two groups, one used to generate the activities known by the classifier and another to be used as observation to the classifier. These two sets were used to gain insight on the performance of the algorithm. The results of the study demonstrate that the algorithm can successfully classify observations, so as long as precautions are taken to prevent the activities known by the classifier to become too large. It is also shown that the tree search performed by the classification can be utilized to partially classify observations that would otherwise be rejected by the classifier. The user study additionally included subjects that performed activities purely used for observations to the classifier. With this set of recordings, it was demonstrated that the classifier does not over-fit and is capable of halting the classification of an observation.