Development of ergonomic posture recognition technique based on 2D ordinary camera for construction hazard prevention through view-invariant features in 2D skeleton motion

• Published in: Advanced engineering informatics Vol. 34; pp. 152 - 163
• Main Authors: Yan, Xuzhong, Li, Heng, Wang, Chen, Seo, JoonOh, Zhang, Hong, Wang, Hongwei
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01-10-2017
• Subjects: 2D skeleton; Construction worker; Ergonomics; Person posture recognition; RGB camera; View-invariant; Person posture recognition; Ergonomics; Construction worker; 2D skeleton; View-invariant; RGB camera
• ISSN: 1474-0346; 1873-5320
• DOI: 10.1016/j.aei.2017.11.001

•Ergonomic posture recognition is a novel solution to prevent hazard in construction.•We use ordinary RGB camera to improve non-intrusiveness and applicability in EPR.•We developed EPR classifiers based on 2D ordinary camera for hazard prevention.•Our novel solution is through view-invariant features in 2D skeleton motion.•Our solution demonstrated robust accuracy for on-site postural ergonomic assessment.
Outdoor tasks operated by construction workers are physically demanding, requiring awkward postures leading to pain, injury, accident, or permanent disability. Ergonomic posture recognition (EPR) technique could be a novel solution for ergonomic hazard monitoring and assessment, yet non-intrusiveness and applicability in complex outdoor environment are always critical considerations for device selection in construction site. Thus, we choose RGB camera to capture skeleton motions, which is non-intrusive for workers compared with wearable sensors. It is also stable and widely used in an outdoor construction site considering various light conditions and complex working areas. This study aims to develop an ergonomic posture recognition technique based on 2D ordinary camera for construction hazard prevention through view-invariant features in 2D skeleton motion. Based on captured 2D skeleton motion samples in the test-run, view-invariant features as classifier inputs were extracted to ensure the learned classifier not sensitive to various camera viewpoints and distances to a worker. Three posture classifiers regarding human back, arms, and legs were employed to ensure three postures to be recognized simultaneously in one video frame. The average accuracies of three classifiers in 5-fold cross validation were as high as 95.0%, 96.5%, and 97.6%, respectively, and the overall accuracies tested by three new activities regarding ergonomic assessment scores captured from different camera heights and viewpoints were 89.2%, 88.3%, and 87.6%, respectively. The developed EPR-aided construction accident auto-prevention technique demonstrated robust accuracy to support on-site postural ergonomic assessment for construction workers’ safety and health assurance.