Skeleton-based action recognition for biomechanical risk condition assessment

This paper addresses the challenge of biomechanical risk assessment in human lifting tasks, a common issue in fields such as ergonomics, healthcare, and robotics. We propose a novel dataset tailored to assess lifting risk factors based on skeletal motion extracted from with two different approaches: a marker-based motion capture system and a video-based technique. Additionally, we benchmark state-of-the-art methods such as Graph Convolutional Networks (GCNs), Transformers, and State-Space Models (SSMs) for biomechanical lifting analysis. Our results highlight the strengths and limitations of these approaches in the context of human motion analysis. The key contributions of this work include a new dataset for biomechanical risk assessment, a comprehensive performance evaluation of contemporary methods, and a proposed framework for integrating action recognition with safety-critical factors informed by the Revised NIOSH Lifting Equation (RNLE). This research aims to enhance the safety and efficiency of lifting tasks by paving the way for more accurate and interpretable risk assessment systems.