• Underwater
• 3D image reconstruction
• point cloud
• ENU
• NED
• calibration camera
• stereo matching
• outlier removal
• mesh
• feature detection
• stereo vision

Underwater environments constitute a complex and dynamic setting, presenting numerous challenges to visual perception and 3D reconstruction. This thesis provides a review of various methodologies for addressing 3D reconstruction of underwater images.

3D reconstruction can be performed offline or in real-time using various types of cameras, including monocular and stereo cameras. The thesis focuses particularly on images captured by stereo cameras mounted on Autonomous Underwater Vehicles (AUVs). Issues related to image distortion, refraction effects, and underwater lighting variations are examined, and literature-based techniques to mitigate these issues and enhance the quality of the input images are discussed.

Subsequently, the thesis introduces algorithms and image processing techniques enabling the creation of accurate 3D models of underwater objects and environments. These approaches encompass methods such as feature detection, feature matching, and triangulation to obtain a 3D point cloud (PCL). The PCL typically contains outliers, and various methods for their removal to obtain a satisfactory mesh are analyzed. The most commonly used meshing methods are also described, with a critical analysis of their advantages and disadvantages. The obtained point cloud is typically in the camera's reference frame, leading to the challenge of converting the PCL into the underwater world's reference frame, the NED. This transformation is achieved by integrating data from sensors mounted on AUVs.

Additionally, the thesis attempts to apply a 3D reconstruction pipeline, based on the previously analyzed algorithms, to a dataset of simulated images contained in a ROSbag. OpenCV and Open3d libraries, popular in computer vision, are utilized for this purpose. The simulated environments are designed to provide a simplified yet realistic reconstruction, as they are not affected by the disturbances of the underwater world. The reconstruction process starts directly with feature detection, as pre-processing and image filtering are unnecessary. Various outlier removal algorithms are tested on the resulting NED-aligned point cloud, and their outcomes are analyzed. Finally, meshing is performed using two methods available in MATLAB, with their results commented upon.

The 3D reconstruction of underwater images significantly impacts various aspects of the underwater realm, including underwater cartography, underwater archaeology, marine biological research, and underwater infrastructure maintenance. Most importantly, it enables the execution of hazardous tasks and exploration in otherwise inaccessible regions for humans.