Abstract:Creating 3D models through the Structure from Motion technique is a recognized, efficient, cost-effective structural monitoring strategy. This technique is applied in several engineering fields, particularly for creating models of large structures from photographs taken a few tens of meters away. However, discussions about its usability and the procedures for conducting laboratory analysis, such as structural tests, are rarely addressed. This study investigates the potential of the SfM method to create submillimeter-quality models for structural tests, with short-distance captures. A series of experiments was carried out, with photographic captures at a 1-meter distance, using different quality settings: camera calibration model, Scale Bars dispersion, overlapping rates, and the use of vertical and oblique images. Employing a calibration model with images taken over a test board and a set of Scale Bars (SB) appropriately distributed over the test area, an overlap rate of 80 percent, and the integration of vertical and oblique images, RMSE values of approximately 0.1 mm were obtained. This result indicates the potential application of the technique for 3D modeling with submillimeter positional quality, as required for structural tests in laboratory environments.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: **Evaluating the feasibility of achieving sub - millimeter accuracy through Structure from Motion (SfM) technology in a close - range capture environment**. Specifically, the author explored whether SfM technology can generate 3D models with sub - millimeter - level position accuracy when conducting structural tests in a laboratory environment.
### Problem Background
SfM technology is a computer vision method based on the principles of photogrammetry, which extracts high - resolution three - dimensional data by combining two - dimensional images. It is usually used to create 3D models of large - scale structures, especially when taking photos from several meters away. However, there is less discussion about its application in laboratory analysis (such as structural testing).
### Research Objectives
This research aims to explore whether the SfM method can generate 3D models that meet the sub - millimeter accuracy requirements under short - distance capture conditions (1 - meter distance). To this end, the author conducted the following experiments:
1. **Camera Calibration Models**: Different camera calibration models were used, including On - the - Field Calibration (OtF), Computer Vision Calibration (CVC), and Pre - Self Calibration (PSC).
2. **Scale Bar Dispersion**: Different numbers and distributions of Scale Bars (SB) were used.
3. **Overlap Rate**: An 80% overlap rate was set.
4. **Vertical and Tilted Images**: Vertical and tilted images were combined for capture.
### Experimental Results
Through a series of experiments, the author found that:
- When using the PSC model, the RMSE value was the lowest, reaching an accuracy of approximately 0.17 mm, which was significantly better than other models.
- As the number of reference elements (SB) increased, the value of the covariance matrix gradually improved, indicating an improvement in model quality.
- Among all configurations, when using the maximum number of SB (8), the best sub - millimeter - level accuracy could be obtained.
### Conclusions
The research shows that by optimizing the camera calibration model, reasonably distributing Scale Bars, and adopting an appropriate overlap rate and image combination, SfM technology can achieve sub - millimeter - level 3D modeling accuracy in a laboratory environment, thus being suitable for high - precision application scenarios such as structural testing.
### Formula Presentation
When evaluating the quality of 3D modeling, the root - mean - square error (RMSE) formula was used:
\[ \text{RMSE} = \sqrt{\frac{1}{n} \sum_{i = 1}^{n} (E_i - R_i)^2} \]
where:
- \( n \) is the number of samples,
- \( E_i \) is the estimated value at position \( i \),
- \( R_i \) is the actual measured value at position \( i \).
This formula is used to measure the deviation between the estimated value and the actual measured value, thereby evaluating the quality of the 3D model.