Development of AI-based Multi-action Detection System for multiple patients using Day-Night Omnidirectional Cameras

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K18109
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 90150:Medical assistive technology-related
• Research Institution: Japan Atomic Energy Agency
• Principal Investigator: Imabuchi Takashi 国立研究開発法人日本原子力研究開発機構, 福島研究開発部門 福島研究開発拠点 廃炉環境国際共同研究センター, 研究職 (90845471)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: Deep learning / Human pose estimation / Behavior analysis / Omnidirectional camera

Outline of Final Research Achievements

In this study, we developed an automatic monitoring system for detecting potential accidental behaviors of patients and residents at medical care facilities using AI (Artificial Intelligence)-related technologies. Our system is characterized using an omnidirectional camera to acquire a wide range of video images and the application of a 3D joint coordinate extraction algorithm to predict specific accidental behaviors using calculated joint angle variations. In evaluation of prototype system that implemented the proposed methods on an edge-AI computer, we shown our system can detect multiple accidental behaviors with high accuracy and in real time.

Free Research Field

コンピュータビジョン

Academic Significance and Societal Importance of the Research Achievements

全方位カメラで取得した2次元画像に対して人物の3次元関節座標抽出および関節変動角度の学習を組み合わせることで特定行動の高精度推定を実現した．本研究では，単眼の全方位カメラおよび小型エッジAIデバイスの組み合わせによりプロトタイプシステムを開発し，最小限のシステム構成でありながら広範囲をリアルタイムにモニタリング可能な性能を有していることを実証した．これは類似システムと比較して安価な設備投資で実用可能であり，医療介護現場のみならず他の分野においても応用・貢献できる．