Construction safety monitoring method based on multiscale feature attention network
Shuxuan Zhao, Yin Li, Shuaiming Su, Chuqiao Xu, RunYang ZHONG · Scientia Sinica Technologica · 2023
Safety management is one of the most important tasks in the construction process. The automatic construction safety detection method plays an important role in ensuring workers’ safety and reducing construction accidents. This paper proposes a construction safety detection method based on a multiscale feature attention network. First, the instance segmentation network which integrates multi-scale feature attention mechanisms is designed. The high-dimensional feature with rich semantic information is used to guide the network to learn low-dimensional features to enhance the focusing ability of the network in distinguishing the unique features of different construction systems and achieve accurate segmentation of workers and construction equipment in a complex environment. Second, a three-dimensional bounding box reconstruction method based on pose estimation is proposed. This method employs neural networks to extract the orientation of construction equipment in a three-dimensional space. Then, we combine this method with the geometric constraints reflected by the two-dimensional bounding box to reconstruct the three-dimensional bounding box. Finally, a dynamic warning method for construction sites is designed to divide construction sites into dangerous areas, early-warning areas, and safe areas in terms of production safety requirements. Through real-time tracking and the three-dimensional constraint reconstruction of equipment and workers on the construction site, a warning can be given when workers enter dangerous and early-warning areas. The experimental results show that the proposed method achieves >90% classification accuracy and 59% segmentation mAP on the MOCS dataset, which is higher than those achieved by other deep learning and attention-based algorithms. A safety detection case also demonstrates that the proposed method can accurately judge the safety status of workers on construction sites.