Application of improved PoseConv3D model in recognition of unsafe behaviors of construction workers near the edge
To accurately obtain the common hazardous behavioral actions of construction workers and reduce the incidence of fall-from-height accidents,a method based on improved PoseConv3D is proposed for the recognition of construction workers'unsafe behaviors near the edge.The method is improved based on a PoseConv3D behavioral recognition model and adopts a High-Reolution Network(HRNet)pose estimation model to capture the modal information of the workers'skeletal key points in the video map data to reflect the subtle transformations of the behavioral actions.Meanwhile,a 3D Convolutional Block Attention Module(3D-CBAM)is introduced into the 3D convolutional neural network for behavioral feature extraction to strengthen the adaptive learning and assignment capability of the key features,so as to overcome the problems of scene transformation and action similarity.In addition,six types of unsafe behaviors,namely,crossing the guardrail,throwing objects,smoking,standing leaning on the guardrail,sitting leaning on the guardrail,and making phone calls,are selected based on the risk level of the behaviors for image capturing and labeling,and the proximity unsafe behaviors dataset is constructed in this paper.A series of experimental results on the self-constructed unsafe behaviors dataset show that the Top-1 recognition and classification accuracy of this method for the six types of adjacent dangerous behaviors can reach 95.3％,which is 0.7 percentage point,11.5 percentage point,and 1.2 percentage point higher than that of the initial model,as well as the graph convolutional network-based Spatial Temporal Graph Convolutional Networks(ST-GCN)model,and the Adaptive Graph Convolutional Network(AGCN)model,respectively.The improved model also demonstrates enhanced recognition capability for unsafe behaviors with higher similarity,such as smoking.Therefore,the method proposed in this paper has good recognition accuracy and generalization,and can accurately identify the dangerous behavioral actions of workers in construction edge scenarios,which provides a certain reference for construction safety management to reduce the incidence of fall-from-height accidents.
NETL
NSTL
万方数据
