Convolutional Neural Networks for Autonomous UAV Navigation in GPS-Denied Environments

This work addresses the challenge of autonomous Unmanned Aerial Vehicle (UAV) navigation in Global Positioning System (GPS)-denied environments by proposing a new approach that is an amalgamation of data-driven and model-based philosophies. The proposed method exploits datasets acquired from existing frameworks like the Generalized Trust Region Sub-problem (GTRS) and the Weighted Least Squares (WLS). These datasets are then used to feed the proposed Convolutional Neural Network (CNN) specially tailored to create models for UAV navigation. Afterwards, these models are used to make predictions of an optimal trajectory. The obtained numerical results reveal that the proposed CNN reveals improvements in accuracy and robustness to noise when compared to other Machine Learning approaches, while reducing the required training time.