Development of an unsupervised data-driven detection of GNSS outdoor-indoor transitions

Abstract

GNSS positioning systems are inside almost every device interested in getting their location in external environments. Recently, these devices have been performing tasks in environments inside infrastructures where they still require to know their location. Different methods exist and keep developing for indoor positioning, and outdoor positioning is practically solved using GNSS technologies. This work focuses on the scenario where devices benefit from a seamless transition between contexts. The seamless transition goal is to have uninterrupted access to the machine’s position, ensuring success for the overall objective of the tasks it is performing. A tool helpful in accomplishing a seamless transition is detecting when the device is transitioning from an outdoor environment to an indoor one. With this transition detection, the device can prepare accordingly to avoid problems in challenging surroundings that limit the positioning system’s capabilities. The method this thesis proposes involves machine learning to learn the distribution of outdoor data captured when the device interacts with this environment and later raises a flag when the conditions of the measurements change. The intent is to depend less on hard thresholds and adapt better to different locations while overcoming the challenges of data collection and labeling. The strategy relies on one-class support vector machines for their proven effectiveness with novelty and fault detection, along with delay embedding for its suitability to convert time series data to a set of vectors to accomplish the desired target. The resulting algorithm is evaluated in a series of trajectories covering an outdoor-to-indoor transition and portrays the functionality of the methodologies in fulfilling the objective. The evidence shows the potential and advantages of the process to make the detection. It also provides visibility on improvements and additions that can help integrate the algorithm into a general system to leverage the low need for manual configuration and high adaptability.