GPS-Free Navigation Using Vision-Based Convolutional Neural Networks

This paper presents a novel approach to address the challenge of self-localization of flying vehicles. It utilizes visual cues provided by the map imagery fed to a map-recognition convolution neural-network (CNN). This approach is invaluable during the navigation of flying vehicles in scenarios where the Global Positioning System (GPS) signal is unavailable. The proposed approach leverages the power of convolutional neural networks (CNNs) to imitate the visual perception and navigation abilities of homing pigeons, enabling the vehicle to navigate using solely real- time visual data with limited or no GPS information. Two pre-trained CNN’s (SqueezeNet and GoogLeNet) are selected and re-trained with Google Maps imagery, enabling them to efficiently learn and generalize from the diverse visual attributes present in the map. Extensive experimentation and evaluation have demonstrated the efficacy and resilience of the vision-based GPS-free navigation system. The resulting system predicts position accurately achieving an accuracy of 89.9% and 96.4% for SqueezeNet and GoogLeNet, respectively, for images with a resolution of (one km x one km) and reaching an accuracy of 94.7 for GoogLeNet for images with a resolution of (374 m x 374 m). Results underscore the potential of this approach for overcoming the challenge of GPS unavailability in aerial navigation.