Inferring Complex Activities for Context-aware Systems within Smart Environments
The rising ageing population worldwide and the prevalence of age-related conditions such as physical fragility, mental impairments and chronic diseases have significantly impacted the quality of life and caused a shortage of health and care services. Over-stretched healthcare providers are leading to a paradigm shift in public healthcare provisioning. Thus, Ambient Assisted Living (AAL) using Smart Homes (SH) technologies has been rigorously investigated to help address the aforementioned problems. Human Activity Recognition (HAR) is a critical component in AAL systems which enables applications such as just-in-time assistance, behaviour analysis, anomalies detection and emergency notifications. This thesis is aimed at investigating challenges faced in accurately recognising Activities of Daily Living (ADLs) performed by single or multiple inhabitants within smart environments. Specifically, this thesis explores five complementary research challenges in HAR. The first study contributes to knowledge by developing a semantic-enabled data segmentation approach with user-preferences. The second study takes the segmented set of sensor data to investigate and recognise human ADLs at multi-granular action level; coarse- and fine-grained action level. At the coarse-grained actions level, semantic relationships between the sensor, object and ADLs are deduced, whereas, at fine-grained action level, object usage at the satisfactory threshold with the evidence fused from multimodal sensor data is leveraged to verify the intended actions. Moreover, due to imprecise/vague interpretations of multimodal sensors and data fusion challenges, fuzzy set theory and fuzzy web ontology language (fuzzy-OWL) are leveraged. The third study focuses on incorporating uncertainties caused in HAR due to factors such as technological failure, object malfunction, and human errors. Hence, existing studies uncertainty theories and approaches are analysed and based on the findings, probabilistic ontology (PR-OWL) based HAR approach is proposed. The fourth study extends the first three studies to distinguish activities conducted by more than one inhabitant in a shared smart environment with the use of discriminative sensor-based techniques and time-series pattern analysis. The final study investigates in a suitable system architecture with a real-time smart environment tailored to AAL system and proposes microservices architecture with sensor-based off-the-shelf and bespoke sensing methods. The initial semantic-enabled data segmentation study was evaluated with 100% and 97.8% accuracy to segment sensor events under single and mixed activities scenarios. However, the average classification time taken to segment each sensor events have suffered from 3971ms and 62183ms for single and mixed activities scenarios, respectively. The second study to detect fine-grained-level user actions was evaluated with 30 and 153 fuzzy rules to detect two fine-grained movements with a pre-collected dataset from the real-time smart environment. The result of the second study indicate good average accuracy of 83.33% and 100% but with the high average duration of 24648ms and 105318ms, and posing further challenges for the scalability of fusion rule creations. The third study was evaluated by incorporating PR-OWL ontology with ADL ontologies and Semantic-Sensor-Network (SSN) ontology to define four types of uncertainties presented in the kitchen-based activity. The fourth study illustrated a case study to extended single-user AR to multi-user AR by combining RFID tags and fingerprint sensors discriminative sensors to identify and associate user actions with the aid of time-series analysis. The last study responds to the computations and performance requirements for the four studies by analysing and proposing microservices-based system architecture for AAL system. A future research investigation towards adopting fog/edge computing paradigms from cloud computing is discussed for higher availability, reduced network traffic/energy, cost, and creating a decentralised system. As a result of the five studies, this thesis develops a knowledge-driven framework to estimate and recognise multi-user activities at fine-grained level user actions. This framework integrates three complementary ontologies to conceptualise factual, fuzzy and uncertainties in the environment/ADLs, time-series analysis and discriminative sensing environment. Moreover, a distributed software architecture, multimodal sensor-based hardware prototypes, and other supportive utility tools such as simulator and synthetic ADL data generator for the experimentation were developed to support the evaluation of the proposed approaches. The distributed system is platform-independent and currently supported by an Android mobile application and web-browser based client interfaces for retrieving information such as live sensor events and HAR results.
- PhD