Abstract
Detection of underwater structures using a multi-beam forward looking sonar poses significant challenges due to various factors such as sonar noise and multi-path acoustic reflections. In this context, machine learning techniques have proven to be valuable tools. However, they require a substantial dataset of real sonar samples to effectively train the network for structural object detection from sonar images. Notably, even untrained structures or structures that closely resemble one another may fail to be correctly detected by the network. To address this, we propose an innovative algorithm that combines computer vision and statistical probability models for the detection of underwater structures. Our proposed algorithm begins by employing OTSU and connected component labeling and analysis to identify potential regions of interest within individual single-frame sonar images. It then takes advantage of the spatial continuity present across multiple consecutive frames to pinpoint key structural areas in each frame. A probability model based on the Kalman filter is applied to precisely delineate the contours of the structure within each frame. The objective of this process is to reliably generate the edges of underwater structures from sonar images. The effectiveness and reliability of this algorithm are demonstrated through rigorous real-world sea trials conducted at a pier, further emphasizing its practical applicability and suitability for addressing the complexities of underwater structural detection.
