| description abstract | Real-time understanding of on-site construction personnel and construction machinery input is beneficial to navigation-power junction construction management. However, navigation-power junction projects consist of numerous buildings, and the significant weather fluctuations make the background for resource identification complex. Additionally, the construction resources of different categories vary in scale. These factors contribute to redundant computations in existing resource identification algorithms and highlight the need for improvements in recognition accuracy and generalization capability. In this research, an improved YOLOv8 approach for construction resource identification is provided. Firstly, an efficient convolution module is introduced into the backbone network, improving feature extraction capabilities and reducing redundant calculations caused by complex backgrounds of resource entities through spatial reconstruction and channel reconstruction mechanisms. Then, in view of the different sizes of resource entities and the mutual occlusion of some resource entities, shuffle attention is embedded between the backbone network and the feature fusion network to reduce the loss of entity information of various construction resources and enhance the feature capturing capability of small target resources. Meanwhile, a new loss function is proposed to improve the generalization ability of the YOLOv8 model. Finally, monocular vision technology is used to determine location information. To validate the efficacy and superiority of the suggested approach, we use the real data from a navigation-power junction project in China as a case study for our model. | |
