A data-driven approach to classifying manual material handling tasks using markerless motion capture and recurrent neural networks

Abstract

Work-related musculoskeletal disorders (WMSDs) are prevalent problems that encompass a range of conditions affecting muscles, tendons, and nerves due to repetitive strain, non-neutral postures, and forceful exertions. These disorders lead to pain, reduced productivity and substantial healthcare costs. Effective physical exposure assessment tools are needed in the workplace to quantify WMSD risks and the association between exposure and risks. While several tools are available, they are often limited in scope and lack the ability to assess physical risks continuously. In this study, we evaluated a data-driven approach to continuously classify manual material handling tasks and specific task conditions using different feature sets and machine learning algorithms. Specifically, kinematic data from markerless motion capture (MMC) system was used as input for various recurrent neural networks to classify among eight distinct manual material handling tasks: box lifting, asymmetric box lifting, box carriage, box pushing, box pulling, cart pushing, overhead lifting, and box lowering. The models we tested include bidirectional long-short term memory, gated recurrent units, and bidirectional gated recurrent units. We also classified specific task conditions, such as hand configurations and initial lifting height. Overall, using the MMC's kinematic data led to satisfactory results (e.g., accuracy of 80–94 %) in classifying the tasks and the task conditions. Our results, though, also emphasize that classification performance varied across different feature sets, tasks, and between males and females. Nonetheless, use of MMC demonstrates clear potential for physical exposure assessment.