A Robust and Efficient Loop Closure Detection Approach for Hybrid Ground/Aerial Vehicles

Abstract

Frequent and dramatic viewpoint changes make loop closure detection of hybrid ground/aerial vehicles extremely challenging. To address this issue, we present a robust and efficient loop closure detection approach based on the state-of-the-art simultaneous localization and mapping (SLAM) framework and pre-trained deep learning models. First, the outputs of the SuperPoint network are processed to extract both tracking features and additional features used in loop closure. Next, binary-encoded SuperPoint descriptors are applied with a method based on Bag of VisualWords (BoVW) to detect loop candidates efficiently. Finally, the combination of SuperGlue and SuperPoint descriptors provides correspondences of keypoints to verify loop candidates and calculate relative poses. The system is evaluated on the public datasets and a real-world hybrid ground/aerial vehicles dataset. The proposed approach enables reliable loop detection, even when the relative translation between two viewpoints exceeds 7 m or one of the Euler angles is above 50°.