A bag-of-words-based hybrid approach for loop closure detection

Abstract

[eng] Loop closure detection is an important feature in mobile robotics, playing a key role in Simultaneous Localization and Mapping (SLAM) systems by allowing the robot to recognize previously visited locations. Effective loop closure detection ensures the consistency and accuracy of the generated maps, preventing drift in long-term navigation tasks. While the rise of deep learning techniques, particularly Convolutional Neural Networks (CNNs), has significantly advanced visual loop closure detection, challenges remain in scenarios where visual data may be insufficient or unreliable. Under this context, in this thesis we propose two novel approaches for enhancing loop closure detection in these situations. The first approach utilizes laser scan data along with binary descriptors generated using Fast and Adaptive Loop Closure Keypoint Detector (FALKO), a method specifically designed to extract robust keypoints and descriptors from 2D laser data. The second approach proposes a fused solution, combining both laser scans and image data, enabling the system to dynamically integrate both modalities for improved performance in diverse environments. To validate our methods, we benchmark them against a representative set of state-of-the-art approaches using publicly available datasets. The evaluation highlights the advantages of incorporating laser data, both independently and in combination with image data, for improving the performance of loop closure detection, particularly in challenging scenarios where visual-only solutions may struggle