LET-SE2-VINS: A Hybrid Optical Flow Framework for Robust Visual-Inertial SLAM.

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Visual-Inertial Simultaneous Localization and Mapping (VI-SLAM) systems are crucial for autonomous navigation.
• Existing frameworks like VINS-Mono face challenges in accuracy and robustness in complex environments.

Purpose of the Study:

• To enhance the VINS-Mono framework for improved localization accuracy and robustness.
• To develop a hybrid system integrating neural network-based feature extraction and SE2 optical flow tracking.

Main Methods:

• Integration of Lightweight Neural Network (LET-NET) for feature extraction.
• Implementation of Special Euclidean Group in 2D (SE2) for optical flow tracking.
• Utilizing IMU and camera data with pre-integration and RANSAC for feature matching.

Main Results:

• Achieved up to 43.89% improvement in localization accuracy on the EuRoc dataset with loop closure.
• Demonstrated error reductions of 29.7%, 21.8%, and 24.1% in no-loop scenarios (MH_04, MH_05, V2_03).
• Experimental results confirm superior robustness and accuracy through trajectory visualization and analysis.

Conclusions:

• SE2-LET-VINS provides a robust and accurate solution for visual-inertial navigation.
• The system shows significant performance gains in challenging environments.
• Paves the way for advanced real-time autonomous applications.