骨格ベースの人物行動認識のための生成モデルを用いたデータ拡張

カテゴリ: 論文誌(論文単位)

グループ名: 【C】電子・情報・システム部門

発行日: 2024/12/01

タイトル(英語): Data Augmentation Method Using Generative Model for Skeleton-based Human Action Recognition

著者名: 濵部　翔太（名城大学），山田　啓一（名城大学）

著者名(英語): Shota Hamabe (Meijo University), Keiichi Yamada (Meijo University)

キーワード: 人物行動認識，データ拡張，生成モデル，拡散モデル　　human action recognition，data augmentation，generative model，diffusion model

要約(英語): Human action recognition from video images generally requires a large amount of training data, but often only a limited amount of data is available. This paper proposes a data augmentation method using a generative model and a selector for skeleton-based human action recognition. Skeletons are extracted from video images and encoded into a pseudo-image with time-joint point coordinates. A generative model and a selector are trained on these encoded images for data augmentation. We conducted experiments using a diffusion model as the generative model and an Isolation Forest as the selector for a 21-class action classification problem on the JHMDB dataset. The results showed that adding the generated encoded images to the training data of the discriminator significantly increased the accuracy by 0.53 percentage points, from 55.05% to 55.58%, when the number of training data was 10 sequences per class. Furthermore, adding only the generated encoded images selected by the selector to the training data improved the accuracy by 0.90 percentage points to 55.95%.

本誌: 電気学会論文誌C（電子・情報・システム部門誌） Vol.144 No.12 （2024） 特集：電気・電子・情報関係学会東海支部連合大会

本誌掲載ページ: 1209-1216 p

原稿種別: 論文／日本語

電子版へのリンク: https://www.jstage.jst.go.jp/article/ieejeiss/144/12/144_1209/_article/-char/ja/