Publications Repository - Gdańsk University of Technology

This paper presents the development and preliminary testing of a fall detection algorithm that leverages OpenPose for real-time human pose estimation from video feeds. The system is designed to function optimally within a range of up to 7 meters from ground-level cameras, focusing exclusively on detected human silhouettes to enhance processing efficiency. The performance of the proposed approach was evaluated using accuracy values obtained from experiments conducted on the Town Centre Dataset, Multiple Cameras Fall Dataset, and MPII Human Pose Dataset. The results demonstrate high accuracy levels for the investigated approaches, with the Dense Neural Network achieving accuracy rates of 98% on both the Town Centre & MPII Human Pose Dataset and the Multiple Cameras Fall Dataset. These findings highlight the effectiveness of the proposed fall detection system in accurately identifying fall events based on estimated human poses. This study details the algorithm's design, challenges in implementation, and potential for future IoT device integration, aiming to significantly enhance public safety and community monitoring capabilities.

Authors

• Tomasz Ludwisiak,
• dr inż. Magdalena Mazur-Milecka link open in new tab ,
• dr inż. Tomasz Kocejko link open in new tab ,
• prof. dr hab. inż. Jacek Rumiński link open in new tab ,
• Jo Kang-Hyun