A review of binocular vision stereo matching algorithms

Binocular stereo vision features low cost and non-contact sensing, making it a mainstream approach to acquire 3D scene information in fields such as 3D reconstruction and autonomous driving. As its core module, stereo matching is the key determinant of system sensing accuracy and efficiency. This paper presents a comprehensive review of binocular stereo matching. It systematically elaborates the imaging principle of binocular stereo vision and clarifies the pivotal role of stereo matching in 3D reconstruction. According to technical development trends, existing stereo matching algorithms are classified into traditional methods and deep learning-based methods. This paper also summarizes two mainstream categories of datasets: real-scene datasets and synthetic datasets. Combined with mainstream evaluation metrics including end-point error and bad pixel rate, a standardized evaluation system for stereo matching algorithms is established. Finally, the current challenges are analyzed, such as poor scene generalization and insufficient datasets for complex working conditions. Future research directions are further prospected, including multi-modal dataset construction and cross-domain model optimization, aiming to provide a solid reference for related theoretical research and engineering applications.