HOME: Robust Hough-space Matching Method for Structured and Textureless Videos

Visual front-ends for robotics, such as SLAM and visual odometry, typically rely on point-based features like ORB. While effective in textured environments, these methods often fail in structured, man-made settings (e.g., corridors or rooms with pillars) where linear features dominate. Existing line-based alternatives are often computationally prohibitive for real-time edge robotics. HOME (Hough-space One-dimensional Matching of Extrema) addresses this bottleneck by reformulating the line-matching problem into a highly efficient 1D point-matching task within Hough space.

HOME transforms input images into Hough space, where global linear structures manifest as stable local extrema. By using these extrema as keypoints, the system avoids the heavy computational overhead of traditional line segment detection and description. The framework introduces a 1D radial descriptor that is inherently invariant to rotation and translation, eliminating the need for complex orientation estimation. The matching process is further optimized by a deterministic homography estimation scheme that replaces the unpredictable runtime of traditional RANSAC-based approaches, making it suitable for real-time video processing.

Evaluations on the TUM dataset demonstrate that HOME maintains high registration accuracy in structured environments where point-based methods struggle. Because the Hough transform is based on summation, it naturally suppresses noise, leading to more stable keypoint detection than standard image-space methods. HOME operates significantly faster than existing line-based methods, providing a reliable, deterministic alternative for resource-constrained edge devices that require robust localization in structured or textureless scenes.