Incremental fusion of Structure-from-Motion and GPS using constrained bundle adjustments

Area of Science:

• Computer Vision
• Robotics
• Geospatial Analysis

Background:

• Bundle Adjustment (BA) in Structure-from-Motion (SfM) faces challenges with long image sequences, including high computational cost and accumulated drift.
• Existing methods reduce computation time using local BAs incrementally but may not sufficiently mitigate drift.
• Current GPS-SfM fusion methods often minimize a weighted sum of errors, lacking guarantees on reprojection error and requiring manual weight tuning.

Purpose of the Study:

• To develop novel constrained Bundle Adjustment (BA) techniques for fusing GPS and Structure-from-Motion (SfM) data.
• To address the limitations of existing fusion methods by enforcing upper bounds on reprojection errors.
• To integrate these new fusion BAs into an incremental SfM pipeline and evaluate their performance.

Main Methods:

• Introduced two constrained Bundle Adjustment (BA) formulations for GPS-SfM fusion.
• These methods enforce an upper bound on reprojection error, unlike existing weighted sum approaches.
• Integrated the proposed constrained BAs and a standard fusion BA into an incremental SfM framework utilizing local BA.

Main Results:

• The constrained fusion BAs provide alternatives to existing methods, guaranteeing bounded reprojection errors.
• Evaluated the performance of three fusion BA methods (two proposed, one existing) on long monocular image sequences with low-cost GPS data.
• Demonstrated the effectiveness of the proposed constrained BAs in reducing drift and improving accuracy in SfM.

Conclusions:

• The proposed constrained Bundle Adjustment methods offer improved accuracy and reliability for GPS-SfM fusion in long image sequences.
• These techniques provide a more robust alternative to existing fusion BAs by directly controlling reprojection errors.
• The integration into an incremental SfM pipeline shows promise for real-world applications requiring accurate visual-inertial odometry.