Comparison, validation and calibration of map-matching algorithms for GPS data

点击翻译按钮获取中文摘要

Since the end of 2013, Intact Insurance offers usage-based car insurance programs designed to leveragedata-driven technologies and improve insurance products. Under these programs, Intact collects driving datafrom customers using either a GPS dongle connected to the car on-board computer or a smartphoneapplication. This data is used to characterize driving behavior and to detect risky events. In order to pre-processthis massive amount of data collected, a map-matching algorithm needs to be implemented. Map matching isthe process of estimating a user's position on a road segment. Concretely, GPS positions are provided only interms of latitude and longitude and are not linked spatially to the road network. This is particularly of concernin urban centers, where tall buildings can completely block GPS signals or create spurious signals, leading topositional noise of several meters in magnitude. If the goal is to determine travelling patterns, then it isnecessary to explicitly match each trip to the travelled network links. Although several map-matchingprocedures have been developed in the literature, no formal comparison exists in terms of accuracy andcomputational speed. These issues are extremely important for Intact Insurance, who needs to find the optimaltrade-off between computational cost and precision. Furthermore, the performance of existing methods mayvary greatly depending on the type of environment (urban, suburban, highways, etc.) and the use of a singlemethod might not be the optimal solution. Moreover, the combination of GPS and sensor data has been rarelyinvestigated in the context of big data; this could improve the map-matching accuracy but increases thecomputing complexity. Therefore, the objective of this research project is to: 1) evaluate the performance ofexisting map-matching methods both in terms of efficiency and accuracy, 2) investigate the benefits ofcombining GPS and sensor data in the map-matching accuracy, and 3) calibrate the optimal method(s) intoIntact Insurance's analytical pipeline.