The Research Of Underwater Binocular Stereo Dense Matching Algorithms

Ocean has great potential for development with its rich resources and vast space. Vision Technology is an important tool for underwater exploration, with high speed, high precision, large amount of information, etc., and is equipped with underwater exploration robot. The key to make underwater robot work is to obtain reliable underwater image and image processing, whereas the stereo matching is the core technology of underwater image processing. This paper conducts systematic researches on underwater binocular stereo vision and makes improvement to the existing stereo matching algorithm in the air to make it suitable for underwater environment. The main works are as follows:(1)The paper starts with the cameral projecting model, and introduces four coordinates as well as their transitions. Then it introduces the linear model and refraction model of underwater. Meanwhile, the parallel model in binocular stereo vision is established and the relationship between the parallax and depth is given. The basic principle, processes steps, constraints of disparity and the classification of stereo matching algorithms are introduced as well.(2) In terms of underwater image distortion, blurring due to light refracted, scattering and absorption, and the low correct rate of Daisy matching algorithms for gray image, we propose two improved Daisy methods based on color model. Method One, for RGB model, Daisy algorithm generates respectively the sub-descriptors in three channels R, G, B, and we will combine them into one larger descriptor, for increasing descriptor difference compared with the gray image. After that, we use the Euclidean distance similarity measure method to find the matching points; Second method, we convert the raw RGB color model into a Gaussian color model firstly for enhancing image illumination invariant and color invariance, then Daisy algorithm generates respectively the sub-descriptors in three channels E, E1, E2.After that, we computer the averaged value of them, and also choose the Euclidean distance similarity measure method to search the matching points. Experimental results show that our two methods achieve the same matching accuracy and reduced matching time compared with the SIFT feature matching algorithm and this can be used to perform dense disparity for underwater images.(3) In the light of the underwater images no longer meet the epipolar constraint in the air. An underwater stereo matching algorithm is presented based on color segmentation and curve constraint. Firstly, reference image is segmented by mean-shift algorithm. Then assign different weights to the window pixels according to falling on different divided regions, and region-based matching algorithm SAD is used to calculate the matching cost. Meanwhile, derive reference curve constraint of feature points in reference image is derived and used as the matching search area of SAD algorithm, which can reduce the amount of calculation. Experimental results indicate that the algorithm is better than SIFT feature matching algorithm and matching accuracy is improved obviously.