深層学習法を用いた鳥類の動画自動検出追跡システムの開発と響灘ビオトープでの運用

抄録

In the Hibikinada reclaimed land, where the rare ecosystem is preserved, the development of wind power generation will continue in the future, so it is important to understand the living behavior of birds through constant monitoring. In the Hibikinada biotope in this area, network cameras are installed to constantly observe birds, but it is difficult to manually analyze the huge amount of video data. In this study, we developed a real-time automatic bird detection and bird tracking system for 4K videos using the deep learning method YOLO. Through constant operation, verification for ecological analysis of birds was conducted to clarify the problems of the system. By optimizing the YOLO parameters, we were able to detect birds with a minimum size of 0.002% or less in the 4K video. By creating learning data from images obtained by constant operation of the system and developing a unique learning model, we were able to improve the detection accuracy of birds and significantly reduce false detections.