Identifying and predicting the locations that people wish to visit goes some way to understanding human behaviour, but it does not consider what the person wished to achieve, only where they were. For this, we turn to exploring identi- fyingactivitiesandcontexts, where activities are low-level actions performed by the user, and contexts represent times when a user had a specific goal or task to achieve. For example, a context could represent periods of time throughout which a person was exercising, but the activity being performed would specifi- cally bejogging or playing football.
This topic is revisited later in this thesis, in Chapters 6 and 7, where the identification and prediction of contexts respectively are considered.
2.5.1
Identifying Activities
Identifying the activities being performed by individuals has been considered as a hierarchical learning problem that can discover activities at multiple scales from video data [Kim et al., 2010]. There is little distinction between activity
2. Background and Related Work
extraction and activity labelling in existing work, where a group of sensor read- ings or part of a video are provided and the task is to classify which activity from a set is being performed. This includes identifying the current activity from video [Brand et al., 1997; Messing et al., 2009; Morris and Trivedi, 2011], accelerometers [Choudhury et al., 2008; Lee and Mase, 2002; Ravi et al., 2005], accelerometers and heart rate sensors [Lester et al., 2005; Pirttikangas et al., 2006], accelerometers and GPS devices [Subramanya et al., 2006], and custom sensor networks [Van Kasteren et al., 2008]. Relating specifically to trajecto- ries, the existing work is focused on 2D movement trajectories extracted from video data [Bashir et al., 2006, 2007; Brand et al., 1997]. Once extracted, these trajectories have been split into subtrajectories, typically based on changing velocity, where Principal Component Analysis (PCA) and Markov models have been used to detect the activity being performed [Bashir et al., 2007; Brand et al., 1997]. Geospatial trajectories have been considered as a source of activ- ity identification, but this typically entails identifying periods of time spent at specific locations (e.g. [Huang et al., 2015]), which we consider a di↵erent prob- lem and discuss in Section 2.3. Using time and features of the locations, Yu et al. [2015a] propose identifying activities from the types of locations visited, and Liao et al. [2007a] propose a hierarchical activity model for individuals that de- scribe the significant locations that a person visits and the activities performed at each of these locations, where activities are determined by assigning labels to grid cells on a map based on the speed of travel and proximity to transit routes in each cell.
Although more broad than activity identification, labelling of individuals has also been considered by using trajectories to classify students based on the course they study [Farrahi and Gatica-Perez, 2008a]. Expanding further on this, Farrahiet al. [2008b] label transitions in data in an attempt to summarise behaviour by identifying users with similar lifestyles. The labels they add take a form similar to ‘heading home at 10 p.m.’, however it is limited in that it only considers three class labels for locations, namelyHome,Work orOther.
2. Background and Related Work
2.5.2
Identifying Contexts
Context identification, in contrast, aims to discover periods of time in which a person is likely to have had similar goals or performed similar actions but the process is not necessarily concerned with the specific activity being performed. Identifying contexts has been considered from the locations visited by users, where properties of the interactions are used to determine whether a location is likely to have a single purpose (e.g. a restaurant), or multiple purposes (e.g. a shopping centre with restaurants and shops) [Assam and Seidl, 2014].
Identifying the contexts of the user, rather than the location, has been ex- plored using entropy-based clustering [Bao et al., 2011], and sequence-based ap- proaches that consider the transitions between contexts [Lemlouma and Layaida, 2004]. Utilising contexts, research has also focused on developing architectures and applications that adapt devices based on the current context [Anagnos- topoulos et al., 2006; Lemlouma and Layaida, 2004]. Situation and intention awareness are related areas that have a greater focus on developing tools and techniques to aid a person in conducting a particular task to achieve some goal [Howard and Cambria, 2013; Vinciarelli et al., 2015], with specific examples in defence [Howard, 2002] and aviation [Endsley, 1995, 2000].
2.5.3
Predicting Future Contexts
Similarly to location prediction, the task of context prediction has been consid- ered in the literature, where context and location prediction sometimes overlap. Using beacons placed around a smart home to identify di↵erent contexts, Seo and Lim [2016] predict the future context of occupants using classification tech- niques, aiming to identify what the user wishes to do in the house next. Addi- tionally, Assam et al. [2014] proposes using identified contexts of locations as a basis for location prediction. Separating out the context and location prediction stages, Yu et al. [2015a] and Bhyri et al. [2015] employ two-step approaches that first aims to predict the context a user will be in and then aims to identify the
2. Background and Related Work
specific location that the user will visit to fulfil the context, achieved through classification and statistical techniques. Contexts have also formed the basis for recommender systems, with Le et al. [2015] using the context history of users to predict a bundle of locations that the user may like to visit.