Visual Localization Domain for Accurate V-SLAM from Stereo Cameras.

Area of Science:

• Computer Vision
• Robotics
• Signal Processing

Background:

• Visual Simultaneous Localization and Mapping (V-SLAM) is crucial for robot navigation but faces challenges in accurate trajectory estimation.
• Traditional V-SLAM methods often struggle with precise keypoint matching in stereo image sequences.

Purpose of the Study:

• To propose a novel approach for V-SLAM by introducing a transformed domain that emphasizes visually significant features.
• To enhance the accuracy of trajectory estimation in V-SLAM using a new domain and filtering techniques.

Main Methods:

• Developed a VIsual Localization Domain (VILD) based on information-theoretic principles for V-SLAM.
• Utilized Circular Harmonic Function (CHF) filters to obtain transformed coefficients for image representation.
• Employed a first-order approximation using first-order CHF filters for direct coefficient computation.
• Applied VILD for keypoint matching and tracking across stereo video sequences.

Main Results:

• The VILD provides a theoretically grounded and visually relevant image representation.
• Keypoint matching and tracking in VILD demonstrated improved performance over spatial domain methods.
• Experimental results on real-world datasets confirmed significantly enhanced trajectory estimation accuracy.

Conclusions:

• The proposed VILD offers a robust framework for V-SLAM by focusing on visually salient features.
• Integrating visually-driven filtering in VILD substantially improves the accuracy of trajectory estimation in stereo vision systems.
• This approach advances the field of V-SLAM by providing a more effective method for localization and mapping.