Computer Vision-Based RULA Assessment: A Comparison With IMU Sensors for Static and Dynamic Occupational Tasks

Abstract

Computer vision (CV)-based pose estimation offers a scalable approach to ergonomic risk surveillance using the Rapid Upper Limb Assessment (RULA), but workplace deployment requires validation against objective references and characterization of key implementation choices such as camera placement. Most prior validation studies compared only the overall RULA score, leaving component-level agreement and camera-angle effects underexplored. We evaluated agreement of RULA scores from MediaPipe-based CV against Xsens inertial measurement unit (IMU) data in 21 participants performing typing (static) and bottle packing (dynamic) tasks, recorded from 3 camera angles (0°, 45°, 90°). Across 35,142 synchronized observations, the 90° sagittal view produced the highest action-level agreement (typing: 88.7%; packing: 70.4%) and the lowest root mean square error (typing: 0.93; packing: 1.19). Score B (neck/trunk) produced larger disagreements than Score A (upper limb). CV-based RULA shows promise as a screening tool for static tasks; dynamic tasks require cautious interpretation.