Towards Improving Location Identification by Deep Learning on Images

Abstract

When we rely on GPS systems for navigating inside cities, localization errors might arise, especially when passing crossroads or in areas with bad signal due to high buildings. To address this, we investigated a new navigation method, based on identifying location through Deep Learning. We trained two Convolutional Neural Networks on street images, then used them for location recognition. The first neural network is responsible to identify the street, while the second one to identify the segment of the street we are on. We have obtained 99.70% accuracy for street recognition and 96.02% for segment recognition. The results show that, at a proof-of-concept level, the Convolutional Neural Networks are able to accurately identify the location using images, which could be used for complementing the GPS localization systems.