Loop closure detection of visual SLAM based on variational autoencoder.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Loop closure detection is crucial for Simultaneous Localization and Mapping (SLAM) to minimize positioning drift.
• Traditional methods using handcrafted features struggle with environmental changes, leading to false positives and inaccurate maps.

Purpose of the Study:

• To propose a novel loop closure detection method using a Variational Autoencoder (VAE) for robust feature extraction.
• To replace handcrafted features with learned, low-dimensional image representations for improved SLAM accuracy.

Main Methods:

• A Variational Autoencoder (VAE) is employed as a neural network-based feature extractor.
• An attention mechanism and improved loss function with added constraints enhance image representation.
• Geometric checking is integrated into the backend for filtering incorrect matches and mitigating false positives.

Main Results:

• The proposed VAE-based method outperforms traditional bag-of-words models and other deep learning methods in precision-recall.
• The method demonstrates high robustness against environmental changes.
• Experiments across diverse datasets confirm its applicability in real-world SLAM scenarios.

Conclusions:

• The VAE-based loop closure detection method offers superior performance and robustness compared to existing techniques.
• This approach effectively addresses the limitations of handcrafted features in dynamic environments.
• The method shows significant potential for real-world SLAM applications.