Real-time risk assessment for work-related musculoskeletal disorders (MSD) has been a challenging research problem. Previous methods such as using depth cameras suffered from limited visual range and wearable sensors could cause intrusiveness to the workers, both of which are less feasible for long-run on-site applications. This document examines a novel end-to-end implementation of a deep learning-based algorithm for rapid upper limb assessment (RULA). The algorithm takes normal RGB images as input and outputs the RULA action level, which is a further division of RULA grand score. Lifting postures collected in laboratory and posture data from Human 3.6 (a public human pose dataset) were used for training and evaluating the algorithm. Overall, the algorithm achieved 93% accuracy and 29 frames per second efficiency for detecting the RULA action level. The results also indicate that using data augmentation (a strategy to diversify the training data) can significantly improve the robustness of the model. The proposed method demonstrates its high potential for real-time on-site risk assessment for the prevention of work-related MSD. A demo video can be found at https://github.com/LLDavid/RULA_2DImage.

与工作有关的肌肉骨骼疾病（MSD）的实时风险评估一直是一个具有挑战性的研究问题。以前的方法（例如使用深度相机）会受到视觉范围的限制和可穿戴式传感器的损坏，这可能会对工作人员造成干扰，这两种方法对于长期的现场应用都是不太可行的。本文研究了一种基于深度学习的快速上肢评估（RULA）算法的新颖端到端实现。该算法将正常的RGB图像作为输入并输出RULA动作级别，这是RULA大分数的进一步划分。在实验室中收集的抬高姿势和来自Human 3.6（公共人类姿势数据集）的姿势数据用于训练和评估算法。总体，该算法在检测RULA动作水平方面达到了93％的精度和每秒29帧的效率。结果还表明，使用数据增强（一种使训练数据多样化的策略）可以显着提高模型的鲁棒性。所提出的方法证明了其实时实时风险评估在预防与工作有关的MSD方面的巨大潜力。演示视频可以在https://github.com/LLDavid/RULA_2DImage中找到。