Fast Underwater Optical Beacon Finding and High Accuracy Visual Ranging Method Based on Deep Learning

Abstract

Visual recognition and localization of underwater optical beacons is an important step in autonomous underwater vehicle (AUV) docking. The main issues that restrict the use of underwater monocular vision range are the attenuation of light in water, the mirror image between the water surface and the light source, and the small size of the optical beacon. In this study, a fast monocular camera localization method for small 4-light beacons is proposed. A YOLO V5 (You Only Look Once) model with coordinated attention (CA) mechanisms is constructed. Compared with the original model and the model with convolutional block attention mechanisms (CBAM), and our model improves the prediction accuracy to 96.1% and the recall to 95.1%. A sub-pixel light source centroid localization method combining super-resolution generative adversarial networks (SRGAN) image enhancement and Zernike moments is proposed. The detection range of small optical beacons is increased from 7 m to 10 m. In the laboratory self-made pool and anechoic pool experiments, the average relative distance error of our method is 1.04 percent, and the average detection speed is 0.088 s (11.36 FPS). This study offers a solution for the long-distance fast and accurate positioning of underwater small optical beacons due to their fast recognition, accurate ranging, and wide detection range characteristics.