Computer vision based manual concrete vibration quality monitoring

Vibration is essential for achieving maximum concrete density and ensuring long-term structural performance. Manual vibration remains widely used in complex construction environment but relies heavily on workman skills and judgment. Automatic monitoring provides an effective means for on-site workmanship and product quality assessment. This study applies computer vision to assess workman behaviours and concrete surface quality for defect prevention. Workman behaviours are recognized through a skeleton-based approach, integrating object detection (Faster R-CNN), pose estimation (HRNet-w32), and action recognition (PoseC3D) with a global accuracy of 97.55%. Concrete surface state is classified through ShuffleNet V2 with a global accuracy of 99.95%. A prototype system is developed to provide dynamic feedback and operation instructions during vibration. The proposed method is applied in a metro station construction project in Suzhou, China to validate its applicability. This approach offers a practical alternative to manual supervision, enabling intelligent, dynamic quality management in complex construction environments. • A computer vision-based framework is proposed for manual concrete vibration monitoring. • Both workman behaviour and concrete surface state are detected and analyzed. • Pose-based action recognition achieves 97.55% accuracy for workman behaviour detection. • Surface state classification achieves 99.95% accuracy using lightweight neural networks. • The framework is validated through a real metro project.