A stereo vision approach to scour pit detection using lightweight YOLOv8n-pose

Abstract Sonar detection and diver surveys represent prevalent methods utilized in current offshore wind practices to monitor scour pits around offshore wind turbine foundations. This monitoring is indispensable for ensuring the safety and longevity of these critical structures. Sonar detection, renowned for its versatility across various water depths, offers the advantage of high-resolution 3D mapping of the seabed. However, it entails significant equipment and operational expenses, with its efficacy susceptible to water turbidity and rough sea conditions. On the other hand, diver surveys enable meticulous inspections, providing invaluable insights. Nonetheless, they are constrained by water depth and weather conditions, rendering them time-consuming and financially burdensome due to labour costs. To overcome these limitations, this paper introduces and validates a new accurate scour pit visual monitoring technique, slated for integration into a foundation maintenance AUV operating on the seabed. The research of this new technique begins with the underwater image enhancement model. Subsequently, YOLOv8n-pose is proposed for detecting the deepest point of the scour pit, achieving 84.1% AP while maintaining lightweight properties. Furthermore, the incorporation of SGBM introduces a novel idea for reconstructing the three-dimensional model of scour pit. Finally, experimental verification tests were conducted in the laboratory. The experiment results demonstrate that the proposed technique can swiftly and accurately detect scour pit depth, achieving a Root Mean Square Error (RMSE) of 5.02 mm. This laboratory-scale validation provides a strong foundation for subsequent testing in real-world offshore environments.