Presentation

· Contributors · Organizations · Search Program · Flagged · Happening Now

Estimating 3D Dynamic External Hand Forces From Markerless Motion Capture Using Attention-Based Recurrent Neural Networks

SessionPoster Session 1

DescriptionPhysical exposure assessment are essential for identifying high-risk tasks, guiding the development and assessment of ergonomic interventions, and enhancing our understanding of exposure-risk relationships. Yet, accurately assessing three-dimensional, dynamic hand forces for physical exposure is challenging due to the need for specialized equipment. We explored using data from markerless motion capture systems to predict hand forces during two-handed manual material handling tasks, using attention-based deep learning architecture. Model architectures included an input layer, recurrent neural network layers, an attention layer, dropout functions, and output layers.

Our results were encouraging overall, but predictions were less accurate in both the proximal-distal and anterior-posterior directions during box pushing and pulling tasks. Overall, our findings indicate that the proposed approach has the potential to predict dynamic external hand forces, without direct measurement (e.g., load cell, instrumented gloves), offering a balance of simplicity and non-intrusiveness in quantifying physical exposure. Future work is needed, though, to improve the performance of the model in predicting hand forces during push and pull task conditions.

Contributors

Aanuoluwapo Ojelade

University at Buffalo

Mohammad Sadra Rajabi

Virginia Polytechnic Institute and State University