Real-Time Ocean Waves Three-Dimensional Reconstruction Research Based on Deep Learning

Obtaining precise space-time data is crucial for marine research on wave fields. Stereo vision techniques, which can generate high-resolution three-dimensional (3D) point clouds, have gained popularity for reconstructing small to medium-scale wave fields. However, accurate reconstruction often requires extensive computation time, limiting the applicability of stereo-based analysis approaches in wave forecasting and offshore engineering design. This paper introduces a real-time 3D reconstruction algorithm for ocean wave sequences based on deep learning. Firstly, we incorporate the prior knowledge of disparity common area during the network model training, to enhance the quality of the resulting disparity maps. Then, instead of employing complex point cloud filtering procedures, the disparity masks are leveraged to eliminate outliers and reduce processing time. Finally, we evaluate our algorithm on the Acqua Alta dataset and compare it with the WASS algorithm. The results demonstrate the superior reconstruction effectiveness of our method, with a 75.03% improvement in effective wave height estimation accuracy and a 50.00% improvement in peak wave period estimation accuracy. Additionally, the processing speed of our algorithm remains at 40 frames per second (fps).