水下双目立体视觉对应点匹配与三维重建方法研究

针对水下双目立体视觉成像稠密立体匹配因不满足空气中极线约束问题,提出一种水下对应点匹配与三维测量方法, 可将水下双目相机采集的立体图像校正为符合共面行对齐原则的图像对,再套用空气中成熟的立体匹配方法得到水下左右相 机图像视差图,从而实现水下目标的三维重建。 首先,将进入相机的所有光线总和看成光场,采用四维光场参数表达对每一条 光线建模,据此建立相机的折射成像模型和双目立体视觉模型并计算光线的方向向量;根据光线的光场表达将光线转化为点矢 量的形式,计算方向图像上任意像点对应原图像的像素坐标并确定位置映射关系。 通过插值即可快速得到符合行对齐原则的 左、右方向图像,并最终获得每条光线对应的视差图。 仿真结果表明,方向图像的行对齐误差小于 0. 8 pixel。 水池实验采用事 先标定的靶球作为目标物,利用随机散点主动投射以增加目标物表面的纹理信息,对靶球多次测量的均方根误差为 2. 8 mm,具 有较高的测量精度。

Research on corresponding point matching and 3D reconstruction of underwater binocular stereo vision

The dense stereo matching in underwater binocular stereo vision imaging does not meet the epipolar constraint in air. To address this issue, an underwater corresponding point matching and the 3D measurement method are proposed. The stereo image collected by the underwater binocular camera can be corrected to an image pair conforming to the principle of coplanar line alignment. Then, the display image of the underwater left and right cameras can be obtained by applying the mature stereo matching method in the air to realize the 3D reconstruction of the underwater target. Firstly, the sum of all the light entering the camera is regarded as a light field, and each light is modeled by a 4D light field expression. Based on this, the plane refraction imaging model and the binocular stereo vision model of the camera are established, and the direction vector of the light is calculated. According to the light field expression, the light is transformed into the form of a point vector, the pixel coordinates of any image point on the direction image corresponding to the original image are calculated, and the position mapping relationship is determined. Through interpolation, the left and right images in line with the line alignment principle can be quickly obtained. Finally, the display map corresponding to each light is obtained. Simulation results show that the line alignment error of the direction image is less than 0. 8 pixel. In the pool experiment, the ball calibrated in advance is used as the target, and the random scattered point active projection is used to increase the texture information of the target surface. The root mean square error of multiple measurements of the target ball is 2. 8 mm, which has high measurement accuracy.