Abstract
Affordable home health care systems are extremely important for early diagnosis of disease and to track patient recovery. As part of these systems, the ADL score calculated from activities of daily livings (ADL) and instrumented activities of daily livings (IADL) provide valuable statistics about the functional and cognitive ability of patients and elderly citizens, which is required for deciding treatments and services. However, most of the existing available systems impose constraints on sensor values, the types of detected activities (no parallel/interleaved/joint activities), or the number of users, which reduces the robustness of the system in the real-world settings. Moreover, in order to provide a holistic solution for monitoring activities, it is important not only to provide information about complete activities but also detect attempted incomplete/partial activity instances and report the errors.
In this dissertation, we have designed, implemented and evaluated a novel activity
recognition and person identification system named SARRIMA, a rule-based general framework "QuActive" for modeling activity, a system for recognizing activity steps, and an activity quality monitoring system based on the "QuActive" framework for identifying partial/incorrectly performed activities and finding the errors within the activities (missing steps, wrong steps, wrong orders, and delays within and between activity steps). We show that by incorporating time difference while segmenting the occupancy episodes and feeding those segments in \textit{Apriori} rule association machine learning technique, the system is able to work in homes with multiple people and detect interleaved and parallel activities with higher performance than supervised machine learning systems. Moreover, we provide proof that by modeling activities in terms of activity steps and using grammar rules, it is possible to capture the different variations of the same activity. The grammar rules also enable to find the missing steps or errors due to performing wrong activity steps or doing them in the wrong order. Finally, we provide evidence that using user's activity history and occupancy correlation information, it is possible to infer which user performed an activity event for a significant number (the number varies in different datasets based on user lifestyle and sensor settings) of activity instances even when there is no direct personal identification information available. Our evaluation in different public datasets and collected data from the lab shows that our system performs better than the state-of-the-art activity recognition systems and is able to provide activity quality information of both complete and incomplete activities.
