ORB-SLAM

ORB-SLAM extends monocular visual SLAM to omnidirectional (fisheye and catadioptric) cameras by integrating the enhanced unified camera model (EUCM) to enable localization and mapping with wide field-of-view imagery.

Key Features:

• ORB-SLAM2 foundation: Builds on the ORB-SLAM2 framework for core SLAM components.
• Enhanced Unified Camera Model (EUCM): Uses the EUCM as the projection function to handle catadioptric systems and wide-angle fisheye cameras with up to 195 degrees field of view.
• Full-image omnidirectional processing: Processes images without rectification or cropping, utilizing the full image area despite significant distortion.
• Omnidirectional map initialization: Implements a novel map initialization technique specifically designed for omnidirectional cameras.
• Analytical Jacobians and reprojection error: Analytically derives Jacobian matrices to compute reprojection errors with respect to camera pose and 3D point positions for optimization.
• Real-time performance and accuracy metrics: Demonstrates real-time operation with reported positioning errors below 0.1% in small indoor environments and below 1.5% in larger settings, improving accuracy and robustness over pinhole camera models.

Scientific Applications:

• Autonomous driving: Supports localization and mapping for autonomous driving using fisheye or catadioptric sensors.
• Augmented reality: Enables precise localization and mapping for augmented reality systems that use wide-field-of-view cameras.
• Robotic navigation and mapping: Applies to robotic navigation and large-scale mapping tasks requiring wide-angle omnidirectional imagery.

Methodology:

Integrates the enhanced unified camera model (EUCM) as the projection function, implements an omnidirectional-specific map initialization, and analytically derives Jacobian matrices to compute reprojection errors with respect to camera pose and 3D point positions.