Towards Model-Based 3D Underwater Object Detection and Pose Estimation for Sonar Data

Underwater is one of the harshest yet critical environments for visual perception. Considering the poor illumination, limited range, and back-scattering, ordinary 2D cameras fail to generate usable data, especially in the deep ocean. Therefore, acoustic sonars that produce point cloud data are used in such conditions. Still, it remains infeasible to acquire a sufficient amount of underwater data for training learning-based methods such as artificial neural networks. Therefore, we propose a model-based (training-less) approach for 3D object detection and pose estimation for man-made and known underwater structures. Moreover, many existing techniques rely on down-sampling the point cloud to images, for instance, which naturally results in information loss. Therefore, the proposed approach directly uses the point cloud data without down-sampling to preserve information. In this context, we systematically modified and adapted the simple RANSAC method to detect complex known underwater objects. Through a proof-of-concept example, we outlined the requirements, advantages, and shortcomings of such an approach compared to the other 3D object detection variants. The results were first demonstrated on the simulated data and later on the underwater sonar data.