Region Detection Geometry Prediction Convex Polytopes

The patent segments the region geometry representation into a hierarchical structure: first identifying candidate regions at a coarse level, then refining geometry only for detected regions using convex polytopes. This avoids processing every pixel in the scene, reducing computational complexity while maintaining geometry accuracy for relevant regions.

Instead of performing full per-pixel labeling across the entire scene, the patent applies geometry prediction only to detected candidate regions. This partial action approach focuses computational resources on relevant areas, reducing overall complexity while preserving necessary geometry accuracy for navigation decisions.

The patent changes the geometric representation parameters from simple axis-aligned bounding boxes to convex polytopes defined by multiple hyperplanes. This allows representation of complex, rotated, and irregular shapes while maintaining computational efficiency through parametric definitions rather than pixel-level descriptions.

The patent uses a composite representation combining multiple hyperplanes to form convex polytopes. This composite structure can approximate complex geometries by combining multiple simple planar surfaces, achieving better geometry accuracy than single bounding boxes while remaining computationally tractable.

The patent replaces heuristic-based mechanical labeling rules with a learned neural network model. The network learns scene-independent geometric patterns from training data, providing reliable predictions across diverse scenes without requiring hand-crafted heuristics that may fail in novel situations.

The neural network model provides a universal solution that works across different scene types, object categories, and environmental conditions. Unlike scene-specific heuristics, the learned model generalizes to new situations, improving reliability while maintaining implementation simplicity through a single unified approach.