Overview: With the exploration of Marine resources, the underwater vision technology has a very broad application prospect. As a main branch of computer vision, binocular stereo vision has become a common realization form of underwater vision technology due to its relatively simple structure and high efficiency. Among the realization steps of image acquisition, camera calibration, image preprocessing, stereo matching and 3D reconstruction, stereo matching is a key technology and a hot spot in binocular vision field. How to get a matching algorithm with high precision and high speed is more difficult to study. At present, the underwater stereo matching method adopts adaptive search, determines the optimal search domain, and is based on color segmentation, but only takes into account the calculation accuracy. For example, in previous literature, SIFT feature matching and curve constraint were adopted to further improve accuracy, but the calculation speed was not discussed. In this paper, a matching method is proposed to improve the operation speed under the premise of ensuring accuracy.
According to different matching primitives, stereo matching is mainly divided into regional matching and feature matching. The region-based matching algorithm takes parallax for each pixel, and finally gets the dense parallax map. While the feature-based matching algorithm extracts feature points for matching, and finally gets the sparse parallax map. Due to problems such as low illumination and high noise in underwater images, which will cause large errors in solving dense parallax plan. Thus, feature matching algorithm that is insensitive to noise and environment is selected in this paper. Commonly used feature matching algorithms include SURF, KAZA, SIFT, ORB, etc., among which SIFT algorithm is the most classic one. This algorithm has many advantages, such as good stability, high precision and strong robustness, which are widely used in many scenes. But at the same time, there are problems such as large calculation and long time. After SIFT, an ORB algorithm is also proposed. ORB is an algorithm for rapid feature extraction and matching. Although there are some problems such as low accuracy and many mismatches in complex scenes with large changes such as scale and illumination, the computing speed is greatly improved compared with the previous algorithm, and some scenes can be two orders of magnitude higher than the SIFT algorithm. Although the velocity of ORB algorithm is improved, there are a series of factors affecting the image quality, such as refraction, lens distortion, etc., in the underwater environment. If it is used directly, there will be many mismatches. For the accuracy problem, the curvilinear polar constraint commonly used in underwater stereo matching is derived under the circumstance that underwater environment does not meet the polar constraint in air, which is applicable to underwater curve polar constraint, so as to eliminate the mismatched points and ensure the matching accuracy.
This paper adopts underwater stereo matching method combining ORB feature detection and curve polar line constraint. Firstly, it used ORB feature detection algorithm to match feature points of underwater image, and then reduced mismatched points according to curve polar line constraint, so as to achieve the goal of improving speed under the premise of ensuring accuracy.