Visual Odometry via Antipodal Transform Rigid Structure Detection

The patent extracts and utilizes only the essential geometric constraints from triangulation (epipolar geometry and fundamental matrix) without performing full 3D reconstruction. By working directly with 2D image correspondences and applying the eight-point algorithm to compute the fundamental matrix, the method eliminates the computationally intensive triangulation step while preserving the necessary geometric information for accurate trajectory estimation.

The patent segments the visual odometry process into distinct stages: feature detection, feature matching, fundamental matrix computation, and trajectory estimation. By separating these steps and applying appropriate algorithms to each (such as RANSAC for robust fundamental matrix estimation), the method reduces overall computational complexity while maintaining accuracy through specialized processing at each stage.

The patent performs preliminary computation of the fundamental matrix and epipolar geometry constraints from feature correspondences before trajectory estimation. By pre-establishing these geometric relationships, the method ensures consistent scale maintenance across frames without requiring repeated triangulation operations, thereby reducing processing time while preserving trajectory consistency.

Instead of performing actual 3D triangulation to create reconstructed features, the patent uses 2D image correspondences and fundamental matrix computations as a simplified copy of the essential geometric relationships. This approach maintains the necessary constraints for uniform scale without the computational burden of full 3D reconstruction in each frame.

The patent extracts only the essential 2D geometric correspondences and epipolar constraints needed for visual odometry, discarding the computationally intensive 3D reconstruction step. By working directly with 2D image plane relationships and fundamental matrices, the method maintains complete feature tracking information while significantly reducing computational energy requirements.

The patent uses lightweight 2D feature descriptors and matching algorithms instead of heavy 3D reconstruction pipelines. These simplified 2D representations serve as disposable intermediaries that capture the necessary geometric information for trajectory estimation without the sustained computational burden of maintaining full 3D models throughout the processing pipeline.