Center-of-Gravity-Aware Graph Convolution for Unsafe Behavior Recognition of Construction Workers

Area of Science:

• Construction Safety
• Computer Vision
• Human-Computer Interaction

Background:

• Falls from height pose a significant risk in the construction industry.
• Accurate identification of high-risk worker behaviors near edges is crucial for accident prevention.

Purpose of the Study:

• To develop an improved action recognition model for identifying unsafe worker behaviors.
• To enhance proactive accident prevention strategies in construction environments.

Main Methods:

• Proposed a novel dynamic spatio-temporal graph convolutional network (CoG-STGCN).
• Incorporated a center of gravity (CoG)-aware mechanism using anthropometric priors.
• Utilized a Multi-Layer Perceptron (MLP) to dynamically adjust the skeleton graph's adjacency matrix based on CoG features.

Main Results:

• CoG-STGCN achieved a Top-1 accuracy of 95.83% on a dataset of hazardous behaviors.
• Demonstrated an average accuracy of 94.17% in fivefold cross-validation.
• Showed significant improvements in recognizing actions with rapid CoG shifts compared to baseline ST-GCN.

Conclusions:

• The CoG-STGCN offers a more effective and physically informed approach for recognizing unsafe behaviors.
• This model provides a valuable tool for intelligent early warning systems in construction.
• Enhances safety by improving the detection of stability-related risks near edges.