A visual SLAM loop closure detection method based on lightweight siamese capsule network.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Loop closure detection is crucial for accurate mapping in visual SLAM.
• Traditional methods (bag-of-words) suffer from complexity, slow speeds, and sensitivity to environmental changes.
• Existing deep learning approaches may not fully address these limitations.

Purpose of the Study:

• To develop a more accurate and robust loop closure detection algorithm for visual SLAM.
• To overcome the limitations of traditional bag-of-words models and existing deep learning methods.
• To enhance the performance of SLAM systems through improved loop closure detection.

Main Methods:

• Proposed a novel algorithm utilizing a Siamese capsule neural network for loop closure detection.
• Designed a new feature extractor tailored for capsule networks.
• Implemented parameter pruning techniques specific to capsule layers to reduce computational load.

Main Results:

• The proposed algorithm demonstrated higher accuracy and robustness compared to traditional and other deep learning methods on the CityCentre and New College datasets.
• The algorithm exhibited strong robustness against variations in illumination and viewing angles.
• The complete SLAM system incorporating the new algorithm was evaluated on the KITTI dataset.

Conclusions:

• The Siamese capsule neural network approach offers a significant advancement in loop closure detection for visual SLAM.
• The method provides a robust and accurate solution, particularly in challenging environmental conditions.
• This work contributes to building more reliable and precise maps in real-world robotic applications.