Real-time Underwater 3D Reconstruction Based on Monocular Image

Abstract

Real-time 3D reconstruction is essential for unmanned autonomous robot operation. In this work, we present a real-time reconstruction algorithm for underwater environments based on monocular image. First, the depth of the sparse features is calculated by performing feature optical flow tracking. Then, Delaunay triangulation is performed based on these features and the depth image is densified using triangular interpolation with the assumption that the seabed is close to the plane. Moreover, the densified depth image is used as the initial value, and the matching cost of the pixels in the current image and the neighboring images are computed according to the depth information. The SGM (semi-global matching) algorithm is used for cost aggregation to smooth and optimize the depth image, which can help to improve the accuracy of the depth image; finally, the current depth image is fused with the previous depth image using depth filter to further improve the accuracy and remove outliers. The experimental study shows that the algorithm can obtain better visualization on real-world underwater datasets. By leveraging the GPU, the algorithm can get real-time processing speed on the laptop.