Anomaly Detection on GPS Data

Abstract

Security companies around the world use GPS trackers to track the valuable goods of their customers. Using the trackers you know where they are but you cannot assess whether that is okay or not, a car may be moving in the middle of the night that should be parked or a parked car may be in a location that it has never been before. The goal of this research was to find a way to recognize novel – possibly troubling – situations from data of GPS trackers, to ultimately help recognize the theft of the vehicles they are mounted in. We compare multiple classification methods on their ability to achieve this feat, scoring them on things like the possibility of providing an Outlier Score, Computational Cost and whether the model has the means to add a time dimension. The One-class Support Vector Machine did rather well to get a feeling how a model like this would work in practice. We made a proof of concept implementation. Keywords: Novelty Detection; Anomaly Detection; Outlier Detection; Support Vector Machine (SVM); Unsupervised