Visual SLAM is essential for resource-constrained embedded systems, yet global
feature-based approaches often struggle with high computational loads and instability
in repetitive environments. Conventional methods process entire images without
structu...
Visual SLAM is essential for resource-constrained embedded systems, yet global
feature-based approaches often struggle with high computational loads and instability
in repetitive environments. Conventional methods process entire images without
structural discrimination, causing resource waste and frequent false matches in
texture-poor regions, which hinders pose estimation stability.
To address these challenges, this paper proposes ROI-ORB, a method that defines
Regions of Interest (ROI) based on line-segment detection and selectively extracts
features within these areas. The algorithm restricts ORB operations using line-based
masks to reduce redundant computation and reinforces matching accuracy through a
filtering process that integrates disparity cues with epipolar geometric constraints.
Experiments on the KITTI Stereo 2015 dataset demonstrate that ROI-ORB achieves 107.73
FPS, a 5.5× speed improvement over global ORB, proving its efficiency. It also
records an Inlier Ratio of 91.68%, confirming strong structural reliability. In
conclusion, ROI-ORB is validated as an effective front-end solution balancing
real-time performance and matching quality in resource-constrained environments.