Abstract:

In order to identify unsafe actions of construction workers accurately and timely, a real-time recognition method based on skeleton was proposed by using computer vision, image recognition technology and safety knowledge of buildings. Then, pose estimation algorithm and action recognition algorithm were combined to establish a model, and motion classification and unsafe motion recognition were realized by training the model with comprehensive data. Among them, AlphaPose was used for posture estimation to extract coordinate positions of key points of bones, and ST-GCN was adopted for action recognition. Finally, the model was verified through experiment. The results show that the method's accuracy rate of dangerous actions identification for ladder climbing can reach 98.48%. Meanwhile, ST-GCN has stronger generalization ability compared with support vector machine, random forest and LSTM. This method can improve traditional safety management mode and information application of safety management through real-time identification and early warning of unsafe actions of workers on site.

Key words: space-time graph convolutional networks (ST-GCN), construction workers, unsafe behavior, action recognition, AlphaPose, long short-term memory (LSTM)