Hybrid continuous construction worker pose estimation under occlusion and missing data using ST-GCN and Kalman filtering

Frequent occlusions on construction sites hinder continuous worker pose estimation, limiting the reliability of automated monitoring systems. This research addresses this critical gap by proposing and validating a novel hybrid framework integrating Spatial-Temporal Graph Convolutional Network (ST-GCN) inference with Kalman filter-based smoothing and refinement. The objective was to develop a robust methodology that overcomes the limitations of prior approaches, particularly during extended occlusions. Experimental results demonstrate that the proposed ST-GCN + Kalman filter method significantly reduces pose discontinuities and achieves substantially lower estimation errors (approx. 28 % lower Mean Position Error) compared to baseline methods (linear interpolation and Kalman filter), even under simulated occlusions up to 5 s. The major conclusion is that this validated hybrid approach enhances the continuity and accuracy of pose tracking under challenging site conditions, thereby providing more reliable data crucial for advancing construction automation applications such as safety monitoring and ergonomic analysis.