A Monocular Visual Odometry Method Based on Virtual-Real Hybrid Map in Low-Texture Outdoor Environment

Area of Science:

• Robotics
• Computer Vision

Background:

• Visual odometry (VO) is crucial for robot navigation in unknown environments.
• Mainstream VO struggles in low-texture environments lacking sufficient point features.
• Man-made scenes often contain abundant line features, offering an alternative for navigation.

Purpose of the Study:

• To develop a robust monocular visual odometry algorithm for low-texture environments.
• To improve the stability and accuracy of VO by utilizing line segment features.
• To create a virtual-real hybrid map for enhanced camera pose estimation.

Main Methods:

• Reprocessing line segment features to generate virtual intersection matching points.
• Constructing a virtual map using these virtual intersection points.
• Combining the virtual map with a real map (built on point features) to form a hybrid 3D map.
• Solving continuous camera pose estimation using the hybrid map.

Main Results:

• The proposed algorithm demonstrates improved stability in low-texture environments.
• Experimental results confirm the robustness and effectiveness of the virtual-real hybrid map approach.
• Successful camera pose estimation was achieved in diverse challenging scenes.

Conclusions:

• The virtual-real hybrid map based monocular visual odometry algorithm significantly enhances performance in low-texture environments.
• Reprocessing line features offers a viable solution for VO challenges in environments with limited texture.
• The proposed method provides a robust and effective approach for robot navigation and exploration.