ScholarWorks@숙명여자대학교

초록

In manufacturing environments, human pose estimation is important in monitoring workers’ safety in the workplace. Usually, a single camera struggles to capture a worker’s entire body clearly, which is necessary for precise movement analysis. To improve this, estimating human poses from multiple camera views significantly enhances worker safety by reducing the problem of not seeing all parts of the body. However, most multi-view methods require camera calibration, which is impractical in manufacturing environments. This study introduces a framework for detecting anomalous 3D poses using multi-view cameras without camera calibration. Our approach simplifies the process and reduces costs, involving four main steps: estimating 2D human poses, filling in occluded body parts with affine transformations, converting these to 3D human poses using a Lifting Network, and detecting anomalous poses using a graph convolutional network. By filling in occluded body parts using the affine transformation, we improve the accuracy of 3D pose detection by 1.54% in the term of the Percentage of Correct Keypoints. Furthermore, our graph convolutional network method achieves a 0.943 F1-score in detecting anomalous poses. This framework aims to boost the usefulness of safety systems that rely on accurate human pose data. Our findings confirm that this method is effective and practical for real-world applications.

키워드