Study on motion pattern recognition of crane boom in construction process based on 3DResNet

Aiming to address the challenge of recognizing tower crane boom motion patterns during construction and dedicated to resolving the construction safety risks caused by crane collisions, this research proposes an innovative digital identification framework by integrating spatiotemporal behavior analysis and digital twin mapping. Firstly, by collecting 583 segments of on-site boom videos and annotating over 20000 tower crane images, a boom pattern dataset was established. Data augmentation techniques, such as image rotation, were employed to enhance the dataset's scale and improve the model's generalization capability. The You Only Look Once version 5 (YOLOv5) algorithm was utilized for boom target recognition and segmentation to accurately locate the boom area. On this basis, the 3D Residual Network (3DResNet) algorithm was applied for pattern classification, leveraging its 3D convolutional kernels to simultaneously capture spatial structure and temporal motion features. Experimental results indicate that this method achieved a recognition accuracy of 87.39% on the test set, with a warning response time not exceeding two seconds, significantly outperforming traditional temporal models (with an accuracy improvement 1.32 times higher than Long Short-Term Memory (LSTM)+ 2D Residual Network (2DResNet)). The research outcomes provide an efficient and reliable solution for construction safety management and control, offering technical support for the development of smart construction sites.