基於AI影像辨識的下背傷病風險評估與監測系統

Abstract

在當今的勞動環境中，人工物料搬運已是普遍及常見的工作項目，尤其常在物流業、貨運業及碼頭業等作業現場，而此類作業經常要求從事重複的抬舉、搬運、推拉等動作，這些活動有高風險導致肌肉骨骼的損傷。為了避免此類健康危害，「職業安全衛生法」明確要求雇主針對可能引發肌肉骨骼疾病的重複性作業，制定並執行預防措施，並針對擁有100人以上員工的事業單位，規定必須進行「人因性危害預防計畫」。 傳統的「人因性危害預防計畫」主要使用問卷調查法抑或是量表觀察法，然而這些方法普遍存在著評估成本高、主觀性強等問題，因此本研究旨在利用影像偵測技術整合現有的影像辨識骨架模型，並引入自行設計能夠適用於不同身高、體型的個人骨架校正方式，最後取用關鍵指標法（Key Indicator Method, KIM）的評量精神與其部分的評估方式，實現簡易的下背傷病風險評估演算法。 此演算法於模擬實驗中的結果顯示，針對16種不同的搬運姿勢所得出的精確率皆達到 90% 以上，並且透過我們的人員骨架校正方法，也顯著改善搬運時腰部彎曲角度的觀測精度，並且在最後我們將演算法與圖形介面整合，開發出一套基於AI影像辨識的下背傷病風險評估與監測系統。該系統不只可以生成下背評估的量表，還能在監測系統上實時給予作業人員警示。 本研究所提出的技術能夠準確預測和分析作業人員的姿勢變化及其潛在風險，從而預防不良姿勢、高重複動作和過度負荷引起的職業性肌肉骨骼傷害健康風險，同時有效避免相關職業傷害。

In today's work environment, manual material handling has become a common and prevalent task, particularly in industries such as logistics, freight, and dock operations. These tasks often require repetitive lifting, carrying, pushing, and pulling, activities that carry a high risk of musculoskeletal injuries. To mitigate these health hazards, the Occupational Safety and Health Act explicitly mandates that employers implement preventive measures for repetitive tasks that could lead to musculoskeletal disorders. For organizations with over 100 employees, a "Prevention Plan for Ergonomic Hazards" is required. Traditional prevention plans for ergonomic hazards primarily rely on questionnaire surveys or observation-based scales. However, these methods are often associated with high assessment costs and strong subjectivity. This study aims to integrate existing skeleton recognition models using image detection technology and introduce a self-designed personal skeletal correction method adaptable to different heights and body types. It also incorporates the evaluation principles and partial assessment methods of the Key Indicator Method (KIM) to develop an algorithm for lower back assessment. The algorithm's results from simulated experiments show that the accuracy for 16 different handling postures exceeded 90%. Furthermore, the proposed skeletal correction method significantly improved the observational accuracy of lower back bending angles during handling tasks. Finally, we integrated the algorithm with a graphical user interface to develop an AI-based image recognition system for lower back injury risk assessment and monitoring. This system not only generates lower back assessment scales but also provides real-time alerts through the monitoring system. The proposed technology can accurately predict and analyze workers' posture changes and potential risks, thereby preventing occupational musculoskeletal injuries caused by poor posture, high repetition, and excessive load. This effectively reduces the occurrence of work-related injuries.