An underwater binocular stereo matching algorithm based on the best search domain
aa r X i v : . [ c s . C V ] F e b An underwater binocular stereo matching algorithmbased on the best search domain
Yimin Peng, Zijing Fang, Yunlong Li
Abstract —Binocular stereo vision is an important branch ofmachine vision, which imitates the human eye and matchesthe left and right images captured by the camera based onepipolar constraints. The matched disparity map can be calcu-lated according to the camera imaging model to obtain a depthmap, and then the depth map is converted to a point cloudimage to obtain spatial point coordinates, thereby achieving thepurpose of ranging. However, due to the influence of illuminationunder water, the captured images no longer meet the epipolarconstraints, and the changes in imaging models make traditionalcalibration methods no longer applicable. Therefore, this paperproposes a new underwater real-time calibration method and amatching method based on the best search domain to improve theaccuracy of underwater distance measurement using binoculars.
Index Terms —underwater binocular, epipolar constraints, real-time calibration, underwater distance measurement
I. I
NTRODUCTION
Robots must ”see clearly” underwater, machine vision is themost important thing, and the camera is the ”eyes” of the robot,and its calibration is the basis of machine vision. The edge partof the image obtained by the camera will be distorted. Aftercalibration, the internal and external parameters of the cameraand the distortion parameters can be obtained to obtain thecorrect image. Most of the most advanced underwater visionsystems are manually calibrated in shallow water and canbe used in the high seas without modification. However, theunderwater situation is complex, and the refractive index of thewater changes adaptively according to salinity, temperature,depth or other underwater environment indicators. At thistime, the underwater vision system is prone to large errors.Therefore, in order to avoid the influence of nonlinear geo-metric transformation caused by underwater light refraction,this project is based on the machine learning method and theimu sensor to establish an underwater model of water depth,temperature, and refractive index transformation, which isanalogous and equivalent to the calibration model on land , Toachieve underwater real-time calibration. Generally, an IMUcontains three single-axis accelerometers and three single-axisgyroscopes. The accelerometer detects the acceleration signalof the object in the independent three-axis coordinate systemof the carrier, and the gyroscope detects the angular velocitysignal of the carrier relative to the navigation coordinatesystem. The angular velocity and acceleration of the objectin three-dimensional space, and the posture of the object iscalculated based on this. It has very important applicationvalue in navigation. Based on the machine learning method,learn the correspondence between the two-dimensional obser- vation and the three-dimensional posture from the trainingsamples obtained in advance under different postures, andapply the learned decision rules or regression functions tothe samples, and the result is used as the posture of thesample estimate. Learning-based methods generally use globalobservation features, do not need to detect or recognize localfeatures of objects, and have good robustness.II. P
ROBLEM DESCRIPTION
Binocular vision technology is generally divided intofive steps: image acquisition-image preprocessing-cameracalibration-stereo matching-three-dimensional reconstruction.In this paper, a sealed cabin is used to encapsulate andstabilize the binocular camera to achieve the purpose ofwaterproofing. The arranged environment is taken underwaterto obtain images. Since the vision system needs to work inan underwater environment, during the underwater imagingprocess, light must pass through water/glass/air/lens (they are avariety of different media) and finally reach the imaging planefor imaging; therefore, such as As shown in Figure 1, whenlight passes through these different types of media, refractionoccurs. Most of the most advanced underwater vision systemsare manually calibrated in shallow water and can be used inthe high seas without modification. However, the underwatersituation is complicated, and the refractive index of the waterchanges adaptively according to salinity, temperature, depthor other underwater environment indicators. At this time, theunderwater vision system is prone to large errors. There-fore, in order to avoid the influence of nonlinear geometrictransformation caused by underwater light refraction, thisproject is based on the machine learning method and theimu sensor to establish an underwater model of water depth,temperature, and refractive index transformation, which isanalogous and equivalent to the calibration model on land , Toachieve underwater real-time calibration. Generally, an IMUcontains three single-axis accelerometers and three single-axisgyroscopes. The accelerometer detects the acceleration signalof an object in the independent three-axis coordinate systemof the carrier, while the gyroscope detects the angular velocitysignal of the carrier relative to the navigation coordinatesystem. The angular velocity and acceleration of the objectin three-dimensional space, and the posture of the object iscalculated based on this. It has very important applicationvalue in navigation.II. E
STABLISHMENT OF UNDERWATER BINOCULARVISION MODEL
In the land-based binocular stereo vision system, the targetobject and the camera are all located in the air, but in thestereo vision system in the underwater environment involvedin this article, the above two are located in water and airrespectively, so the light will pass through different media,and finally Image on the image plane. During the experiment,this article stabilized the binocular camera in the airtight cabin,fixed the airtight cabin with clamps, so that it was immersedin water to take pictures of the target object.The optical axes of the left and right cameras are placed par-allel to the outside of the waterproof cover and perpendicularto the refraction plane. As shown in the figure, the distancebetween the optical center of the camera and the glass coveris , the thickness of the transparent waterproof glass is , andthe water object point passes through the waterproof cover.At and point , they are imaged as point in a linear camera.However, in actual situations, due to equipment factors and theangle of view, the optical axes of the two cameras cannot becompletely parallel and placed perpendicular to the refractionplane, and the cameras and the refraction plane will be inclinedat a certain angle. R