Deep Learning-Based Cascaded Light Source Detection for Link Alignment in Underwater Wireless Optical Communication

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Obtaining the light source position from the image is an important solution for achieving link alignment in laser-based underwater wireless optical communication (UWOC) systems. However, in practical scenarios, the misalignment degree between the light source and camera is variable, and factors such as ambient light may introduce disturbances, leading to significant variations in the appearance of light spots in images. Existing research primarily relies on simple features like brightness, color, or shape, which makes it difficult to accurately obtain position information from these non-ideal images. In this paper, deep neural networks (DNNs) with strong feature extraction capabilities are introduced to automatically learn the patterns of the light source from diverse images. A detection architecture cascading an object detector and a keypoint detector is adopted, achieving better comprehensive performance in terms of accuracy and speed. To train and evaluate the deep learning model, we construct the UWOC Light Source Detection Benchmark (ULDB) dataset. This dataset comprises 2200 images captured in a standard swimming pool, covering a misalignment range far beyond existing studies. On the ULDB test set, the proposed detection method achieves an average precision (AP) of 99.1% and an average positioning error of 4.66 pixels, while the traditional method may frequently extract false light spots. To the best of our knowledge, the ULDB dataset is the first image dataset specifically designed for the task of link alignment between UWOC terminals. And the dataset is available at https://github.com/happycode123/ULDB-Dataset.
Funding Information
  • Key Research and Development Program of Hainan Province (ZDYF2023GXJS016)
  • Project of Sanya Yazhou Bay Science and Technology City (SCKJ-JYRC-2022-40)
  • National Key Research and Development Program of China (2022YFC2808200, 2022YFB2903403, 2022YFC2808100)

This publication has 29 references indexed in Scilit: