A Comparison of Deep Learning Architectures for WiFi-based Urban Localisation

Nowadays it is possible to find WiFi access points at almost any place in our cities. The growth of WiFi access points has made possible to consider, in densely populated areas, WiFi technology as a support to GPS. This paper presents different architectures of deep learning techniques for outdoors localisation tasks, to support the localisation of an autonomous vehicle using fingerprint based methods and WiFi RSS measurements. A new dataset covering a residential area of 30975 m2with 25 positions was collected on three different days, several weeks apart with three different devices. The area also showed different coverages of WiFi Access Points. Four different architectures were tested: firstly, a classic four layer DNN and a CNN based on ResNet bottleneck units. Then two sequence oriented NN's: an RNN and an LSTM. Results indicate that all the architectures are able to perform reasonably well on the localisation task being the LSTM one obtaining the best results with 93.26% accuracy and a mean distance error of 1.62m. All the architectures showed a decrease in performance in the areas where the number of Access Points was below 100. To reduce this effect, we have introduced class weighting achieving a reduction of the mean distance error around a 4–6% for all the tested architectures.