Feature diverse hierarchical classification of human gait with CW radar for assisted living

Activity recognition and estimation of gait parameter are medically essential components of remote health monitoring systems that can improve quality of life, enable personalized treatments, acquire continual medical data to better inform doctors of the patient's well-being, reduce health costs, and ensure rapid response to medical emergencies. Discriminating between a large number of oftentimes similar activities using the radar micro-Doppler effect, however, requires extraction of features that can capture differences in nuances within the signatures. This optimal feature set varies according to the number and type of classes involved. Thus, this work proposes a novel feature diverse hierarchical classification structure, which prevents significant sources of confusion between classes. Our results show a 19% reduction in confusion between creeping and crawling and an elimination of confusion between falling and walking, yielding an overall 7.3% performance improvement above a multi-class support vector machine classifier.