Underwater bio-inspired optical navigation based on deep learning and ambient light polarization characteristics

Traditional atmospheric navigation methods are unsuitable for underwater environments due to the skin effect of electromagnetic waves. However, using natural polarization pattern distribution for navigation offers advantages like resistance to interference, low cost, and no cumulative errors. The stability of underwater polarization patterns makes them ideal for underwater navigation. The polarization angle, degree of polarization and light intensity are utilized to comprehensively represent underwater polarization patterns. A squeeze-and-excitation ResNet50 model is used to estimate solar position information (zenith and azimuth angles) based on these features. Compressed sensing is then applied to predict latitude and longitude. Experimental results show average errors of 0.189° for solar zenith, 0.158° for azimuth, and 0.117° latitude, 0.034° longitude for a single location. For multiple locations, errors are 0.622° (zenith), 0.295° (azimuth), 0.238° (latitude), and 0.023° (longitude).