Mathematical problems of feature matching for vision-guided vehicles with limited resources

Abstract

Vision-based sensing plays a critical role for autonomous vehicles, where image data serve as the primary input for navigation, mapping, state estimation, and motion control. These operate under real-time and resource-constrained conditions, requiring feature detection and matching algorithms that are accurate and computationally efficient. A key component of such pipelines is the identification of keypoints – distinctive image locations x ∈ R^2 within a digital image function I : R^2 → R^c, where c is the channels number, and I(x) ∈ R^c denotes the local pixel intensity. Each detected keypoint is associated with a descriptor d ∈ R^D encoding a local visual structure in a form invariant to translation, rotation, and moderate scale changes. Keypoints between two images x_i and x′_i are matched by comparing descriptors yielding the correspondences x′_i ↔ x_i. Geometric verification is performed by estimating a transformation matrix H ∈ R^{3×3} and accepting ∥x′_i−Hx_i∥ < ε, where ε is a matching tolerance. The proportion of such inlier correspondences referred to as the inlier ratio serves as an accuracy metric. Classical keypoint methods, e.g. the Scale-Invariant Feature Transform (SIFT) versus learned methods, e.g. the SuperPoint (applied to a manually constructed dataset of satellite imagery), are mathematically evaluated. Performance is analyzed in terms of inlier ratio and computational efficiency, reflecting the trade-offs between robustness and resource use. The results pursue design guidelines for integrating lightweight yet accurate vision algorithms into platforms such as UAVs, enabling reliable visual odometry and the Simultaneous Localization and Mapping (SLAM) under constrained hardware conditions.

Pages of the article in the issue: 138 - 141

Language of the article: English