Buoy Light Detection and Pattern Classification for Unmanned Surface Vehicle Navigation

Junseok Lee, Taeri Kim, Seongju Lee, Jumi Park, Kyoobin Lee

Abstract: Buoys and beacons indicate information about dangers in coastal navigation. At night, owing to challenging visibility, buoy lights are employed instead of buoys and beacons. For safe navigation, it is crucial to comprehend these lights, and autonomous vessels require algorithms capable of classifying buoy lights without human intervention, particularly during the night. To address this, we propose a Buoy Light Detection and Classification Network (BLDCNet), which combines buoy light detection and pattern classification. BLDCNet is applied to the Temporal Shift Module (TSM), known for its excellent performance in video understanding, to achieve precise classification based on the continuous light patterns in sequential images. We evaluate the performance of BLDCNet using a synthetic dataset generated to resemble real maritime environments and a real-world dataset obtained by capturing buoy light pattern videos onshore. BLDCNet achieved a classification performance of 89.21% for 11 different buoy light patterns.