Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)

Legacy Department

Computer Engineering


Walker, Ian D

Committee Member

Green , Keith E

Committee Member

Groff , Richard E

Committee Member

Brooks , Richard R


For many individuals, aging is frequently associated with diminished mobility and dexterity. Such decreases may be accompanied by a loss of independence, increased burden to caregivers, or institutionalization. It is foreseen that the ability to retain independence and quality of life as one ages will increasingly depend on environmental sensing and robotics which facilitate aging in place. The development of ubiquitous sensing strategies in the home underpins the promise of adaptive services, assistive robotics, and architectural design which would support a person's ability to live independently as they age. Instrumentation (sensors and processing) which is capable of recognizing the actions and behavioral patterns of an individual is key to the effective component design in these areas.
Recognition of user activity and the inference of user intention may be used to inform the action plans of support systems and service robotics within the environment. Automated activity recognition involves detection of events in a sensor data stream, conversion to a compact format, and classification as one of a known set of actions. Once classified, an action may be used to elicit a specific response from those systems designed to provide support to the user. It is this response that is the ultimate use of recognized activity. Hence, the activity may be considered as a command to the system. Extending this concept, a set of distinct activities in the form of hand and arm gestures may form the basis of a command interface for human-robot interaction. A gesture-based interface of this type promises an intuitive method for accessing computing and other assistive resources so as to promote rapid adoption by elderly, impaired, or otherwise unskilled users.
This thesis includes a thorough survey of relevant work in the area of machine learning for activity and gesture recognition. Previous approaches are compared for their relative benefits and limitations. A novel approach is presented which utilizes user-generated feedback to rate the desirability of a robotic response to gesture. Poorly rated responses are altered so as to elicit improved ratings on subsequent observations. In this way, responses are honed toward increasing effectiveness. A clustering method based on the Growing Neural Gas (GNG) algorithm is used to create a topological map of reference nodes representing input gesture types. It is shown that learning of desired responses to gesture may be accelerated by exploiting well-rewarded actions associated with reference nodes in a local neighborhood of the growing neural gas topology. Significant variation in the user's performance of gestures is interpreted as a new gesture for which the system must learn a desired response. A method for allowing the system to learn new gestures while retaining past training is also proposed and shown to be effective.