Recent advancements in crack detection research have seen notable improvements, particularly in methods utilizing image classification, object detection, and semantic segmentation. Despite these advancements, accurately describing the three-dimensional location of damage remains challenging, as current periodic inspections predominantly record results in a two-dimensional format. Conventional inspection reports often include photographs of damaged areas but rarely capture undamaged sections, leading to potential oversight of damages in unphotographed areas. To address this issue, a method integrating inspection results into a 3D model using Structure from Motion (SfM) has been proposed, allowing for an objective assessment of damage scale and inspection omissions based on the 3D model. However, SfM requires extensive image data and meticulous camera calibration for optimal results.
In this study, we introduce a method for crack detection utilizing the YOLO object detection technique, combined with the 3D scene creation technology of instantNeRF. This approach enables the visualization of detected cracks in red within a three-dimensional scene. Unlike SfM, instantNeRF delivers high-quality results even with fewer images from limited viewpoints, leveraging the power of neural networks. Furthermore, instantNeRF's capability to generate free-viewpoint images facilitates the development of systems that can evaluate crack progression during each inspection cycle. This proposed method not only enhances the three-dimensional representation of cracks but also improves the accuracy and efficiency of damage assessment in structural inspections.
As a result, Our proposed method effectively creates a 3D scene where cracks are accurately detected and highlighted using YOLOv7. This approach illustrates the capability of integrating object detection with sophisticated 3D rendering techniques to generate comprehensive and informative visual representations of structural damage. Despite these advancements, some cracks are still missed, suggesting that the method has not yet achieved the precision of a close-up visual inspection. To mitigate this issue, we intend to improve accuracy by utilizing transfer learning to incorporate YOLO’s inference results into the training dataset. This iterative process aims to enhance the system’s detection performance. Furthermore, given the recent advancements in various 3D scene creation techniques, we plan to investigate and adopt the most effective methods to further refine the system's accuracy and efficiency. By continually enhancing both the detection algorithm and the 3D rendering process, we believe our system will become a crucial tool for automated structural inspection and monitoring.