Multi-dimensional image feature attention-based stereo vision measurement for close-range scenes

Abstract

In recent years, binocular stereo matching has been widely adopted for three-dimensional object measurement. However, conventional methods often suffer from performance degradation in close-range scenarios, where the reduced distance to the target results in substantially larger disparities, requiring a wider search range and reducing the binocular overlap. These challenges are further exacerbated by severe geometric deformations and occlusions, which significantly increase the difficulty of feature extraction and correspondence matching. To overcome these limitations, we propose Close-Stereo, a novel stereo matching network driven by a multi-dimensional feature attention mechanism. By synergistically modeling consistency across textural, chromatic, and structural dimensions, this architecture strengthens feature discriminability and effectively broadens the range of disparity estimation. Additionally, to facilitate robust training and evaluation in large-disparity contexts, we create a dedicated close-range binocular dataset. Comprehensive evaluations demonstrate that our approach delivers superior matching and measurement performance across both public benchmarks and real-world applications.