Loop Closure Detection by Using Global and Local Features With Photometric and Viewpoint Invariance

Area of Science:

• Robotics
• Computer Vision

Background:

• Loop closure detection is crucial for Simultaneous Localization and Mapping (SLAM) systems.
• Photometric and viewpoint variations pose significant challenges to traditional loop closure methods.

Purpose of the Study:

• To develop a novel loop closure detection algorithm robust to photometric and viewpoint variance.
• To improve the reliability of SLAM systems in dynamic environments.

Main Methods:

• A Siamese Network learns global features invariant to photometric and viewpoint changes using intensity, depth, gradient, and normal vector distributions.
• Local features are extracted based on relative pixel intensity histograms and geometric properties (curvature, coplanarity) for rotation invariance.
• Jointly leveraging global and local features for enhanced loop closure detection.

Main Results:

• The proposed algorithm demonstrates superior robustness in challenging scenarios with significant photometric and viewpoint variations.
• Experiments on TUM and KITTI datasets show performance exceeding current state-of-the-art methods.
• Effective loop closure detection is achieved even under adverse environmental conditions.

Conclusions:

• The integration of global and local features provides a more resilient approach to loop closure detection.
• This method significantly enhances the accuracy and reliability of SLAM systems.
• The algorithm offers a promising solution for real-world robotic applications requiring precise localization.