Upper Body Postural Assessment of Forklift Operators using Artificial Intelligence

Abstract

Observational postural assessment methods which are commonly used in industry suffer from low inter- and intra-rater reliability, as well as high time consumption. Postural assessment systems based on motion capture have been researched, but they have mainly been tested inside lab environments. This thesis aims to develop and evaluate an upper body postural assessment system using a depth camera and OpenPose, a human pose estimation library, for forklift driving in an industrial setting. The results from the computer vision system were compared to XSens, an IMU-based system. Data from three operators that performed seven indoor and outdoor forklift driving tasks in total was recorded with a depth camera and XSens sensors. The angles calculated using the OpenPose keypoints generally followed the trend of the XSens angles and showed small errors. However, the results after applying RAMP thresholds significantly differed. Thus, the direct applicability of the computer vision system for postural assessment could not be confirmed. Limitations that affected the performance were the small scale of the study, as well as the camera field of view and perspective, self-occlusions, and angle calculation formulas. Further study should be conducted to assess the reliability of XSens in industrial environments. In addition, the computer vision system should be compared to manual evaluations to assess its applicability. Further work should also focus on alternative camera positions for improving the field of view and self-occlusion problems.