Depth-based sparse bundle adjustment.

Area of Science:

• Computer Vision
• Photogrammetry
• Geomatics Engineering

Background:

• Traditional bundle adjustment methods assume consistent error models.
• Applying these methods to datasets with error-free observations (e.g., reference points) leads to accuracy loss due to inconsistent error models.

Purpose of the Study:

• To propose a novel depth-based object point model and sparse bundle adjustment method.
• To address the accuracy degradation issue in bundle adjustment when dealing with implicitly error-free observations.
• To enhance computational efficiency in sparse bundle adjustment.

Main Methods:

• Developed a depth-based object point model representing point positions by 1D depth relative to a reference image.
• Derived a corresponding projection model.
• Analyzed sparse block structures of normal equations based on shared image parameters.
• Derived sparse solutions for normal equations and parameter covariance matrices.

Main Results:

• The proposed depth-based sparse bundle adjustment method aligns with the error model of scenarios containing error-free observations, preventing further accuracy loss.
• Simulated experiments confirmed the accuracy improvement compared to traditional methods.
• Both simulated and real-world data experiments demonstrated significant improvements in computational efficiency.

Conclusions:

• The novel depth-based approach effectively resolves the accuracy issues associated with traditional bundle adjustment in specific scenarios.
• The method offers a more accurate and computationally efficient alternative for sparse bundle adjustment, particularly in photogrammetry and computer vision applications.