イベント情報
【GIR公開セミナー】Dr. Nicki Holighaus / オーストリア科学アカデミー(オーストリア）
| 日時 | 2025.11.17（13：00～14：00） |
|---|---|
| 会場 |
ミーティングID： 851 6447 6809 パスコード：501239 |
| 講演者 | Dr. Nicki Holighaus |
| 所属機関 | オーストリア科学アカデミー （オーストリア） |
| 講演タイトル | "Computer Vision and Acoustic Signal Processing for Understanding Animal Communication" ＜要旨＞ In this talk I will present some recent results and future plans regarding the computer vision and signal processing aspects of our research in the project ANIML, in which we study the acoustic communication of a colony of budgerigars (Melopsittacus undulatus) in the animal laboratory of the Acoustics Research Institute. Using artificial intelligence to analyze audiovisual footage of animal groups holds great potential for studying social behaviors where manual analysis of recordings is impractical or even impossible. Our work combines computer vision, machine learning for audio, and acoustic methods to automatically separate individual vocalizations from recordings of a group of animals. In an aviary housing 11 budgerigars, the approach starts with a computer vision system designed to localise and identify individual birds on video. This system was developed by recording each bird individually to generate training data with minimal manual effort and by fine-tuning YOLOv11 and MobileNetV2 models, enabling the detection and identification of budgerigars with 98% mAP and up to 94% accuracy, respectively. At the same time, audio recordings of the group’s vocalization are taken with a planar array comprising 112 microphones. The spatial position of individual birds, derived from video, is integrated with acoustic source localization techniques performed on the audio to separate vocalizations from recordings where up to 11 birds vocalize simultaneously, and to attribute them to specific individuals. This pipeline allows us to capture individual vocalizations difficult to record outside of a group, such as a female budgerigars'. Additionally, it will be used to study turn-taking behaviour and consonant-, and vowel-like elements in budgerigars. Importantly, such an automated system is adaptable across species and enables the creation of large datasets of individually labeled social vocalizations, which allows for deeper analysis of non-human vocal interactions in naturalistic settings on both a structural and behavioral level. |
| 言語 | 英語 |
| 対象 | どなたでも、ご参加いただけます。 |
| 主催 | グローバルイノベーション研究院 ライフサイエンス分野 矢田部チーム |
| お問い合わせ窓口 | グローバルイノベーション研究院・工学研究院 矢田部 浩平 e-mail: yatabe（ここに@ を入れてください）go.tuat.ac.jp |
| 備考 | 本セミナーは、オンライン・対面型の同時開催となります。 |
このページの上部へ