Multi-scale parallel gated local feature transformer.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Visual Simultaneous Localization and Mapping (VSLAM) is essential for autonomous robots.
• Existing VSLAM methods struggle with scale variations and low-texture environments, leading to poor accuracy and robustness.

Purpose of the Study:

• To propose an improved multi-scale local feature matching algorithm, MSpGLoFTR, to address limitations in current VSLAM techniques.
• Enhance feature extraction and matching precision for VSLAM in complex scenarios.

Main Methods:

• Introduced a Multi-Scale Local Attention Module (MSLAM) for feature fusion and resolution alignment.
• Developed a Multi-Scale Parallel Attention Module to capture multi-scale features.
• Incorporated a Gated Convolutional Network (GCN) for dynamic weight adjustment and noise suppression.

Main Results:

• MSpGLoFTR demonstrated superior matching precision and relative pose estimation compared to the baseline LoFTR.
• The algorithm showed enhanced adaptability in complex environments with illumination changes, scale variations, and viewpoint shifts.
• Achieved improved robustness and efficiency in feature matching for vision tasks.

Conclusions:

• MSpGLoFTR offers an efficient and robust solution for feature matching in VSLAM.
• The proposed multi-scale approach effectively handles challenging environmental conditions.
• This advancement contributes to more reliable autonomous mobile robot navigation.