Optimal UAV camera position for automated computer vision-based inspection of bolt looseness in steel structures based on 4D BIM

The manual inspection of bolts in steel structures is time-consuming and labor-intensive and poses significant safety risks. While recent studies have focused on using computer vision for bolt inspection, there is a clear gap in the automated collection of construction site data. Previous research has identified camera position as a crucial factor in the accuracy of vision-based models. This paper introduces a framework that addresses these issues by identifying optimal camera positions for inspecting bolts during construction. By integrating vision-based inspection techniques with 4D building information modeling, this approach meets data-gathering requirements such as safety distance and collision avoidance. It enhances the accuracy of deep-learning model predictions by optimizing inspection distance, view coverage, and view angle. The effectiveness of this framework was demonstrated through its implementation in an actual construction project, where a YOLOv8 architecture was trained using a new dataset, and images collected from the construction site were analyzed using the developed deep learning model. The results show that the deep learning model, which achieved 87.7 percent mean average precision on the test dataset, successfully predicts all the bolt statuses from the image captured using the optimum camera position.