Complex activity recognition and context validation within social interaction tools

Abstract: Human activity recognition using sensing technology is crucial in achieving pervasive and ubiquitous computing paradigms. It can be applied in many domains such as health-care, aged-care, personal-informatics, industry, sports and military. Activities of daily living (ADL) are those activities which users perform in their everyday life and are crucial for day-to-day living. Human users perform a large number of these complex activities at home, office and outdoors. This thesis proposes, develops and validates a mechanism which combines knowledge and data-driven approaches for the recognition of complex ADLs. This thesis focuses on the challenging task of capturing data from sensors and recognizing user’s complex activities which are concurrent, interleaved or varied-order in nature. We propose, develop and validate mechanisms and algorithms to infer complex activities which are concurrent and interleaved and build a test-bed which enables the sharing of this information on social interaction tools (SITs). In particular, we propose, develop and validate a novel context-driven activity theory (CDAT) to build atomic and complex activity definitions using domain knowledge and activity data collected from real-life experimentation. We develop a novel situation- and context-aware activity recognition system (SACAAR) which recognizes complex activities which are concurrent and interleaved and validate it by performing extensive real-life experimentation. We build a novel context-driven complex activity recognition algorithm to infer complex concurrent and interleaved activities. We build and validate an extended-SACAAR system which utilizes probabilistic and Markov chain analysis to discover complex activity signatures and associations between atomic activities, context attributes and complex activities. Our proposed algorithms achieve an average accuracy of 95.73% for complex activity recognition while maintaining Complex Activity Recognition and Context Validation within Social Interaction Tools computational efficiency. We build an ontological extension to SACAAR called semantic activity recognition system (SEMACT) which is based on CDAT and evaluate it using ontological reasoning for activity recognition. It achieves a high accuracy of 94.35% for the recognition of complex activities which are both concurrent and interleaved.We propose and develop novel mechanisms for context validation within SITs. Users might be keen on sharing their activity related information with family and friends. The wide scale use of SITs has made anytime, anywhere communication of user’s activities with family, friends and other interested parties possible. It is important that any activity-based updates based on user’s current context shared on these SITs are up-to-date, correct and not misleading. This thesis further focuses on the validation of context such as activity-based updates within social interaction tools. We validate the correctness and freshness of activity-based updates on SITs for a user by matching them against the inferred activity information by our SACAAR system. This provides the user with a mechanism to always have a correct and timely update which does not mislead their friends, family and other followers. We perform extensive experimentation and our validation algorithm is able to detect 81% of the incorrect updates made to social interaction tools. We propose, develop and implement a novel context-aware Twitter validator which checks and validates user’s tweets on Twitter. During the course of the thesis work, 9 peer-refereed international conference papers and 1 journal paper have been produced.