田邉修斗

深層学習による画像認識を用いた低波浪時の離岸流発生場所把握のための研究

犬飼直之 准教授


Every year, drowning accidents occur at beaches, with offshore currents being one of the causes. According to the Japan Lifesaving Association, about 50% of drowning accidents at beaches are attributed to offshore currents. Offshore currents are strong flows directed seaward, generated by the interaction of waves and coastal topography. Identifying their origin is generally challenging, contributing to a decrease in the number of lifeguards and consequently failing to reduce accident rates. Recent years have seen the development and practical implementation of AI-based methods for detecting offshore current occurrences, with Dumitriu et al. focusing on instance segmentation in "A Novel Benchmark and YOLOv8 Baseline Results," and Ishikawa et al. working on constructing models capable of detecting offshore currents in open areas and near breakwaters. However, distinguishing offshore currents during low wave heights, when wave breaking phenomena are minimal, has proven difficult in previous studies. Given that half of beach accidents occur during low wave heights, as per prior research by our group, it's imperative to enable detection during such conditions.
Hence, in this study, we leveraged insights from offshore current research to attempt detecting offshore current occurrences during periods of minimal wave breaking. We employed image recognition of surface wave deformation patterns at offshore current generation sites and constructed a model using deep learning techniques. The training data comprised offshore current occurrence videos captured at Fujitsuka-hama Beach in 2015. These videos were decomposed into frames, annotated with prominent surface wave deformation patterns at offshore current generation sites, and used to build the model. We utilized the YOLOv8 model, a state-of-the-art algorithm for object detection capable of real-time execution, for model construction. Additionally, we assessed the applicability of detecting offshore currents during low wave heights at Fujitsuka-hama Beach using Dumitriu et al.'s publicly available large dataset.
The results confirmed the inability to detect offshore currents during low wave heights using Dumitriu et al.'s dataset. Conversely, our model utilizing the dataset created in this study recognized offshore currents in 92.5% of similar-angle offshore current videos. Furthermore, by combining Dumitriu et al.'s dataset with Fujitsuka-hama Beach offshore current videos, we constructed a model capable of identifying offshore current occurrence locations during low wave heights with over 90% accuracy, even with changes in camera angles.