Event-Based Photometric Bundle Adjustment.

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Event cameras offer high dynamic range and temporal resolution, ideal for dynamic scenes.
• Simultaneous Localization and Mapping (SLAM) and bundle adjustment are crucial for accurate 3D reconstruction.
• Existing methods often require image-like data or struggle with rotation-only event camera setups.

Purpose of the Study:

• To develop a novel bundle adjustment method specifically for rotation-only event cameras.
• To directly utilize event data and its spatio-temporal properties for photometric bundle adjustment.
• To improve the accuracy and detail of scene reconstruction from event camera streams.

Main Methods:

• Formulated bundle adjustment as a non-linear least squares optimization problem using the event generation model.
• Defined photometric error based on camera rotations and scene brightness triggering events.
• Developed a sparse Levenberg-Marquardt solver tailored for large-scale event data.

Main Results:

• Achieved up to 90% reduction in photometric error, demonstrating significant improvement.
• Generated high-quality, detailed scene maps, revealing previously obscured features.
• Validated effectiveness on synthetic and real-world datasets with modern high-resolution event cameras.

Conclusions:

• Event-based Photometric Bundle Adjustment (EPBA) is a pioneering event-only method for accurate camera pose and map refinement.
• EPBA excels in panoramic imaging and integrates with other methods like IMU dead reckoning.
• The method's open-source availability facilitates further research and application in robotics and computer vision.