Homography Based Egomotion Estimation with a Common Direction

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Egomotion estimation is crucial for robot navigation and augmented reality.
• Existing methods often require numerous feature points or complex iterative algorithms.
• Leveraging the gravity vector and homography offers a promising geometric approach.

Purpose of the Study:

• To investigate minimal solutions for camera egomotion estimation using homography.
• To analyze the impact of reference plane prior knowledge on solution uniqueness and complexity.
• To develop efficient and robust egomotion estimation techniques.

Main Methods:

• Exploration of minimal solution sets for egomotion estimation.
• Derivation of closed-form and Gröbner basis solutions based on reference plane properties.
• Utilizing the gravity vector and calibrated images for geometric constraints.
• Analysis of solution dependency on the number of matched points (two or three).

Main Results:

• Identified distinct minimal solutions for egomotion estimation.
• Demonstrated that solutions vary based on prior knowledge of the reference plane.
• Showcased the derivation of direct closed-form or Gröbner basis solutions.
• Validated the approach's efficiency and robustness on diverse datasets.

Conclusions:

• The proposed method provides efficient and robust egomotion estimation.
• Reference plane knowledge significantly influences the minimal solution space.
• The approach is effective in both synthetic and real-world scenarios.