Continual Learning via Multiple Support Vector Models for Localization with a Single Aerial Image

Aerial localization using images captured by Unmanned Aerial Vehicles (UAVs) opens the door to different applications where the GPS may become inaccessible or lost. Deep learning approaches using Convolutional Neural Networks (CNNs) have shown promising localization results in known and static environments. However, these approaches can produce inaccurate pose estimations in scenarios with dynamic changes occurring over time. Therefore, we propose a methodology to generate multiple models that are trained online, and that can be immediately available for localization inference during the same operation flight. To this end, we study the use of Support Vector Regression (SVR) based on a Continual Learning Rehearsal Strategy. Using the SVR, we seek to generate compact models that can be integrated into a search strategy based on sub-maps created along the operation flight. Our method receives a single aerial image as input, thus leveraging features extracted by backbone networks, which will be used to regress the localization variables. For evaluation purposes, we compare our approach with PoseNet, ORB-SLAM2 and single-model training using the entire dataset, comprising no more than 300 images. The experiments were carried out on four trajectories, with our approach achieving a percentage error of around 7% of the total trajectory, with a processing time of 67 ms.