Ontology engineering and reasoning to support real world human behavior recognition

Resumen

----- ABSTRACT ----- Human behavior recognition [1] has attracted much attention during the recent years due to its multiple applications in the health and wellness domain [2]. Despite their popularity, most existing behavior recognition systems suffer two main constrains which make them practically unsuitable to work in real-world conditions: they are bounded to a specific sensor deployment setup and they are defined to operate in a single application domain. Human behavior recognition systems may certainly undergo changes unforeseen at system design time, with sensors subject to diverse types of anomalies such as failures [3] or deployment changes [4]; thus, a pre-defined, well-known and steady sensor setup cannot be guaranteed. Moreover, the categories of behavior recognized by these systems tend to be quite primitive and with limited applicability; however, their appropriate combination could lead to more meaningful and richer expressions of context for human behavior analysis [1]. In the light of these limitations, there is a clear necessity of comprehensively describing the set of heterogeneous resources involved in the human behavior recognition system, dynamically selecting replacement sensors to ensure continuity of recognition, exhaustively describing human context information, and automatically inferring meaningful and rich expressions of context for human behavior analysis. This thesis investigates novel mechanisms to solve the above limitations of human behavior recognition systems in order to facilitate their seamless, robust and accurate use in realistic conditions. Ontologies [5] are considered here to be the cornerstone technology to realize this idea. The extraordinary characteristics of ontologies, which provide implicit semantics, support interoperability and enable automatic reasoning, fit particularly well with the necessities posed by the problem considered here. Besides, ontologies largely exceed other similar and non-semantic models in terms of flexibility, extensibility, generality, expressiveness, and decoupling of the knowledge from the implementation, thus making it a perfect option to create more advanced behavior-aware systems. This work proposes MIMU-Wear, an OWL 2 ontology [6] which comprehensively describes mainstream wearable sensor platforms consisting of magnetic and inertial measurement units (MIMUs), including the MIMUs capabilities and the characteristics of the wearable sensor platform. This ontology provides implicit semantics enabling the automatic interpretation of the resource descriptions, their abstraction from the underlying technology, and the abstraction of the sensor selection method from the actual sensing infrastructure. The dynamic selection of sensors is enabled through ontology reasoning and querying. The proposed sensor selection method builds on the MIMU-Wear Ontology, applies ontological reasoning to infer candidate replacement sensors from a set of heuristic SWRL rules [8], and iteratively poses SPARQL queries [9] on the ontological sensor descriptions to select the most appropriate MIMU for the replacement of an anomalous one. The proposed ontology-based sensor selection method proves to ensure continuity of recognition as it helps recovering the system capabilities after the replacement takes place. MIMU-Wear could also serve at system startup to identify which sensors should be activated based on the necessities of the behavior recognition system or for the self-calibration of some parameters of the sensing network according to energy constraints or efficiency goals, and based on processing power or memory resources. This thesis further proposes the Mining Minds Context Ontology, an OWL 2 ontology for exhaustively modeling rich and meaningful expressions of context. This ontology enables any combination of cross-domain behavior primitives, also referred to as low-level contexts, in order to infer more abstract human context representations, also called high-level contexts. The context ontology extends beyond the state-of-the-art while uniting emotion information as a novel behavioral component together with activity and location data to model new contextual information. An ontological method based on descriptive logic is developed for deriving high-level context information out of the combination of cross-domain low-level context primitives, namely activities, locations and emotions. The proposed method not only proves efficient while deriving new contextual information but also robust to potential errors introduced by low-level contexts misrecognitions. This method can be used for determining any type of high-level context information from diverse sources of low-level context data. Thus, it can be easily applied to any new domain, only requiring the extension of the ontology itself. The proposed models and methods enable comprehensive descriptions and dynamic selection mechanisms for heterogeneous sensing resources to support the continuous operation of behavior recognition systems; likewise, exhaustively descriptions and automatic inference of abstract human context information is supported to enhance the operation of behavior-aware systems. Hence, these ontologies and ontology reasoning-based methods pave the path to a new generation of behavior recognition systems readily available for their use in the real-world.