Real-time ROS Implementation of Conventional Feature-based and Deep-learning-based Monocular Visual Odometry for UAV

Abstract

Localization or state estimation is one of the most important tasks for UAVs based on different kinds of sensors such as GPS, IMU, Lidar or cameras. However, localization based on only a monocular camera or visual odometry is one of the most challenging research topics. Conventional methods are proposed based on the detection of key features in each image and matching them on consecutive images to estimate the camera motions. Deep-learning methods have also been studied to solve the problem. Although the current learning-based visual odometry methods score high results on public datasets, there is a lack of real-time implementation of the methods in common robot operating systems such as ROS to integrate them into a navigation system. In this paper, we introduce a ROS implementation of state-of-the-art conventional feature-based method, ORB-SLAM3, together with a deep-learning-based method, SC-SfMLearner for real-time UAV localization. A photo-realistic simulator, Flightmare, is used to test the implementation together with another navigation task such as control. The implementation can evaluate both algorithms in real-time operation to compare their performances. Based on evaluation results from the simulated environments, the limitation or failure cases of the algorithms could be found, then, the best parameters of the algorithms can be adjusted to improve the algorithms to avoid failures in practical experiments.