Indoor Location Prediction through Modeling of Human Spatiotemporal Behavior

The ability to capture people’s location within large indoor spaces (such as office buildings, university campus buildings, or hospitals) and to use this information to
predict when and where they will go next is a potentially powerful computational tool. Indoor location prediction algorithms could enable smarter automated receptionists, support ad-hoc meetings whilst occupants are transitioning between locations, or improve room temperature control. However, predicting occupant indoor locations and when occupants will transition between these locations is
challenging. Existing systems that monolithically approach this task perform poorly, which is likely because occupants’ spatiotemporal routines within the interior of a
building, such as a workspace, can be very complex.
To address this problem, this work explicitly models the behavior of occupants using a range of spatiotemporal features such as time of day, previous locations,
or day of the month. We developed a pattern extraction algorithm, ABC-Pattern-Extract, based on Conditional Frequent-Pattern Trees (FP-Trees) that describes a
person’s behavior as a collection of spatiotemporal FP-Trees varying depth. Using 53 occupancy traces from a four-month data collection our algorithm is able to extract on average 1011.6 FP-Trees with an average depth of 5.3. Due to
computational restrictions put in place on the algorithm, ABC-Pattern-Extract was able to create these trees in on average less than 25 minutes. To evaluate our approach we show how an extracted set of patterns can be used to accurately predict the future location of an occupant through a secondary predictive algorithm, ABC-Pattern-Predict. To show the impact of the algorithm in a real world scenario
we chose efficient room temperature control and simulated the energy impact of ABC-Pattern-Predict. Our evaluation shows that ABC-Pattern-Predict achieves a
very high accuracy of 92.56% and an increase in Recall by 10.93% over the current state of the art, PreHeat. The high accuracy and recall improvement over PreHeat leads to an estimated energy consumption reduction of 16.1%, while
maintaining the same level of comfort.