Yuki  Inoue

Research on the Shape Recognition of Structures Using Omnidirectional Cameras

Takaaki IKEDA, Masataka SHIGA

As three-dimensional open data, such as PLATEU and VIRTUAL SHIZUOKA, become more prevalent, the generation of frequent three-dimensional data over time, not just at a single point in time, enables comparisons of multiple periods. This facilitates assessments of crack progression from infrastructure inspections and comparisons of building subsidence and tilt due to earthquake damage. Two widely employed methodologies for generating 3D data are LiDAR and SfM-MVS. SfM-MVS reconstructs 3D data from multiple images, and numerous studies have applied it to conventional frame cameras.
The present study focuses on the use of omnidirectional cameras, which have been shown to shorten shooting time and maximize the model generation area for rapid investigation during the aftermath of severe earthquakes. However, there is a paucity of research addressing the impact of substantial distortion at the image edges, which can result in inaccuracies in the 3D model. The objective of this study is twofold: first, to assess the factors that influence the quality and processing speed of point cloud data generated by omnidirectional cameras in real-world settings; and second, to propose optimal conditions within a specific environment.
Firstly, the Information Systems Building at Nagaoka University of Technology was targeted for the purpose of conducting video recordings. To this end, omnidirectional cameras were utilized, and commercially available software, Metashape, was employed to generate point cloud data. The resolution, shooting distance interval (equivalent to frame per second [FPS]), and camera types were systematically varied, and three perspectives were analyzed: processing time, accuracy, and detail. Furthermore, we examined the underlying causes of certain outcomes by employing color histograms to facilitate a thorough investigation of the details.
The findings revealed that enhancing the resolution resulted in an increase in processing time and detail. Conversely, accuracy remained constant, irrespective of the resolution. Furthermore, it was observed that the shooting distance interval significantly increased processing time and detail, while accuracy remained unaffected. In the context of camera type, it was observed that when selecting shooting equipment or environments that widen the brightness range during shooting, an increase in the number of tie points was noted. Consequently, the study determined the optimal parameters for the shooting conditions and targets to be 3840×1920, with a shooting distance interval of 0.5m (FPS = 2). However, it remains to be seen whether these parameters are equally applicable under broader environmental conditions.