YOLOv5 Based Visual Localization For Autonomous Vehicles

In this paper, we use the advances brought by neural networks for the implementation of a vision based localization framework for autonomous vehicles namely UAVs. We base our work on monocular visual odometry. It is used for incremental localization of autonomous vehicles. This method suffers from drift. Loop closure detection is a way to improve its accuracy. Thus, we introduce a Siamese network able to perform binary classification in order to detect the visited places and the loop closures. This gives us an accurate, light and fast vision based localization framework.