Distance Measures in Behavioral Pattern Analysis

This paper presents the survey of multimodal data distance metrics applied to identify the subject's status in a personal and house-embedded surveillance system. The circadian repetitive status time series (behavioral patterns) are a background for learning of the subject's habits and for automatic detection of unusual behavior or emergency. We applied the active learning technique that allows for adaptation of the system to particular needs of the subject. In a prototype recording environment we captured the heart rate, motion factors, body posture, wrist acceleration and sounds during the scheduled performance of everyday activity. Our results show, that recognition of unusual pattern for possible alerting is performed sufficiently only with metrics compensating for temporal shift between patterns. Best performance (sensitivity 96.5% and specificity 97%) was achieved for suspicious status recognition with use of dynamic time-warping algorithm.