Abstract
The performance of physical workers is significantly influenced by the extentof their motions. However, monitoring and assessing these motions remains achallenge. Recent advancements have enabled in-situ video analysis forreal-time observation of worker behaviors. This paper introduces a novelframework for tracking and quantifying upper and lower limb motions, issuingalerts when critical thresholds are reached. Using joint position data fromposture estimation, the framework employs Hotelling's $T^2$ statistic toquantify and monitor motion amounts. A significant positive correlation wasnoted between motion warnings and the overall NASA Task Load Index (TLX)workload rating (\textit{r} = 0.218, \textit{p} = 0.0024). A supervised RandomForest model trained on the collected motion data was benchmarked againstmultiple datasets including UCF Sports Action and UCF50, and was found toeffectively generalize across environments, identifying ergonomic risk patternswith accuracies up to 94\%.