Grunnleggende konsepter
A versatile hybrid visual SLAM system that combines deep feature extraction and deep matching methods to enhance adaptability in challenging environments.
Sammendrag
The paper introduces SL-SLAM, a robust visual-inertial SLAM system that integrates deep learning-based feature extraction and matching algorithms to achieve superior performance in challenging environments.
Key highlights:
- SL-SLAM supports multiple sensor configurations including monocular, stereo, monocular-inertial, and stereo-inertial.
- It applies deep feature extraction and matching techniques throughout the entire SLAM pipeline, including tracking, local mapping, and loop closure.
- An adaptive feature screening strategy and deep feature bag-of-words adaptation are designed to enhance the system's robustness.
- Extensive experiments on public datasets and self-collected data demonstrate that SL-SLAM outperforms state-of-the-art SLAM algorithms in terms of localization accuracy and tracking robustness, especially in low-light, dynamic lighting, weak-texture, and severe jitter conditions.
- The system is implemented in C++ and ONNX, enabling real-time performance.
Statistikk
SL-SLAM achieves an average ATE (RMSE) of 0.034m on the challenging Euroc V203 sequence, outperforming ORB-SLAM3 by over 0.08m.
On the TUM-VI dataset, SL-SLAM exhibits the lowest accumulated drift in most sequences compared to other VINS methods.
In the authors' self-collected dataset with challenging conditions, SL-SLAM demonstrates superior robustness and stability compared to ORB-SLAM3.
Sitater
"To better adapt to challenging environment, we apply deep feature extraction and matching to the whole process of SLAM system, including tracking, local mapping, and loop closure."
"Adaptive feature screening as well as deep feature bag of words adaptation to SLAM system are designed."
"We conduct extensive experiments to demonstrate the effectiveness and robustness, and the results on public datasets and self-collected datasets show that our system is superior to other start-of-the-art SLAM systems."