Accurate 3D reconstruction of underwater infrastructure using stereo vision

Publication Type:
Issue Date:
Full metadata record
Modern vehicle and pedestrian bridges over water are built on concrete piles; foundations that penetrate the soft soil of the bed of the body of water to sit on the solid rock below. Like all built structures these concrete piles require regular inspection to determine if any preventative maintenance is needed. Stereo vision and Structure from Motion techniques offer a cost effective method of creating a 3D reconstructions of a scene, but the underwater environment around a bridge pile has unique challenges. Poor visibility, strong and varying sunlight, and floating material in the water create difficulties for computer vision. This thesis evaluates exposure control, image enhancement, and feature detection and description algorithms, for the purpose of localising images captured around a bridge pile. Stereo correspondence algorithms are evaluated and used to create a single viewpoint 3D reconstruction of a scene, then a visual SLAM system is used to localise the single viewpoint reconstructions, so that they can be merged together to create a 3D reconstruction of a bridge pile. Visual odometry, using KAZE with CLAHE image enhancement for feature detection, was successfully performed in the underwater environment. ORB-SLAM2 can also perform well, and 3D reconstructions from a single viewpoint (created with block matching, semi-global matching, or ELAS) were merged to create 3D reconstructions of submerged bridge piles.
Please use this identifier to cite or link to this item: