Street direction classification using implicit vehicle crowdsensing and deep learning

Abstract

GPS navigation systems are nowadays used extensively by drivers to reach locations and minimize traveling time. It is crucial to have up-to-date map information for their correct functioning. Map updating approaches leveraging human intervention may lag behind frequent changes in urban road arrangements like allowed directions. In this work, we use a combination of crowdsensing and deep learning techniques to extract information about allowed directions of the streets in a completely automated way. We present a faster and more sustainable alternative to currently adopted methods to keep maps’ details updated, based on crowdsensing without user intervention. In particular, we use data from on-vehicle GPS receivers, onboard cameras, and OSM queries. We describe experiments with data from different sources, such as real datasets and trip simulators. To address the problem of classifying street allowed directions, we propose two approaches: the first is based on an MLP on GPS traces, and the second consists in adopting a ResNet50 model on onboard camera imagery. We discuss how the cheaper and more easily deployable first approach delivers performance comparable to the second.