2.1. Structure-From-Motion Technique and Pipeline
When using 3D modelling techniques based on structure-from-motion, the most important parameter to be understood and quantified in surveys is the ground sampling distance (GSD). This is because this is the parameter from which models are interpreted. The GSD is the distance between two consecutive pixel centres, with respect to the real ground measurements as shown in
Figure 4.
Furthermore, it details the smallest observable objects within the image and subsequently the replicated model [
52]. The greater the value of the GSD is, the fewer details are measurable. Therefore, it important to understand the size of details that are required in a survey and the subsequent final model. The GSD will provide the resolution of the generated models. The size of the smallest visible details should generally be two to three times the value of the GSD during the survey [
53]. This means that it will not be possible to accurately measure and decipher any details on the model that have a real-world measurement that is smaller than 2–3 times the value of the GSD. Given this, it is important to establish what measurements are required for typical pavement distresses. In past studies, it has been shown that the most common distress severities are generally no smaller than 10 mm [
54]. For other studies, related to this, a 3 mm resolution is utilized which corresponds to the metric values of distresses within global survey manuals [
10]. Given this, and using the rule of thumb previously identified, if a resolution of 3 mm would be enough to identify and analyse these distresses then a GSD one-third of this value would be sufficient for the survey. Therefore, the GSD should have a value no larger than 1 mm. This is however a benchmark and if the actual GSD is lower, then the smallest details would be more visible and the models more clearly represented. The GSD is related to the camera parameters and is given by Equation (1) below:
where
D = object distance,
ƒ = focal length and
pxsize = pixel size (as defined by the ratio of the camera’s sensor size to the image size). The camera parameters of focal length and pixel size can be adjusted to yield an appropriate GSD. For this study, a GSD of 1 mm was aimed for based on the requirements seen in previous work linking industry standards to models [
28]. With this GSD in place, the resolutions of the resulting model would allow measurements to be made that could explain the severity assessments of the distresses present on the pavement. Given this, the object distance was manipulated to ensure this value was obtained for the survey. Based on this the equation could be rewritten as shown in Equation (2) with the focus on the object distance:
The object distance will not remain exactly constant during the surveys but an approximate value should be determined beforehand which surveyors should try and not exceed during the survey. To apply Equations (1) and (2), the internal parameters of the camera have to be considered. For the survey, the commercially available DJI Mavic 2 Pro drone was used for the surveys (
Figure 5). The cost of this drone or one with similar specifications is negligible in comparison to the cost of a survey vehicle equipped with lasers and other tools.
The specifications of this device are given in
Table 2, which includes the exact camera specifications utilized for the surveys. The weight is particularly important as this is a dimension that is consistently mentioned in regulations and must be adhered to in order to allow for flights.
Given the specifications of the camera, the object distance was calculated. This resulted in a required object distance of 11.81 m to produce the associated GSD value. Given this value, it was then decided to survey a maximum flying height of 10 m to ensure that at no point the object distance exceeded the recommended value. This was done by monitoring the altitude of the drone throughout the survey with respect to the ground at the beginning of the survey.
The main survey was taken over a distance of approximately 1 km on a roadway that was relatively flat and contained sections suffering from cracking, depressions and rutting. These distresses represent key distresses that appear on roadways and are typical of the most common distresses practitioners have to deal with in real situations. It was especially important to have a road section with cracks in it, as this is the distress that occurs the most in the region of the study [
55]. An image of the overview of the pavement studied is shown in
Figure 6.
The survey roadway was straight and surrounded by trees and open greenery areas. The survey was also carried out during a time when there was minimal traffic to avoid cars blocking the view of the roadway. As mentioned previously, this is a limitation of the process and it would be difficult to survey a heavily trafficked section, as the drone would not be able to visualize the road and the distresses. The survey was also done during a period where there was minimal wind. The drone itself has a maximum wind speed resistance of 29–38 km/h so this also represents a limitation that must be considered during surveys. With heavy winds, getting stable images could be a problem and would result in poor models. The trees that are present in the section also present a challenge to the process as they will cast shadows across the road section and the colour of the resulting model will have a different colour at those points. However, for the metric evaluation, the presence of the dark points will not have an effect. For the survey, a typical SfM workflow was utilized and this is illustrated in
Figure 7. The survey itself took approximately twenty minutes and GPS points on the ground taken by the road authority were used as ground control points for the metric scaling of the subsequent models. This time frame is important as it does represent a limitation of the technique as surveys over longer sections will require recharging the battery of the drone or using multiple batteries as the battery has an approximate life of thirty minutes.
During the survey of the pavements, the flight was pre-set using specific flight settings of the drone. One trip was made across the surface. Whilst multiple trips across the surface could produce more images and possibly a higher accuracy with more details, the survey is limited by the drone battery, which limits the time possible for one survey. Additionally, by using one strip, the workflow becomes more practically applicable and more easily adapted for practitioners. The use of one long burst of images across the surface could produce a deformed 3D model so care must be made to use control points along the surface. This was done using eight coordinated points in every 200 m section of the pavement, which can be considered appropriate based on previous studies [
56]. The errors of these residuals are given in
Section 3.
Furthermore, by controlling the camera specifications and settings, the settings would not change per image and therefore there would be consistency in the image dataset. This is important because if an automated camera setting is used for the camera, images would have different environmental effects and matching images could cause issues for the 3D replication. The images were taken in sequence moving horizontally across the pavement with the camera focusing downwards at the pavement surface allowing for an estimated overlap of 70% by an image with the GPS of the drone being utilized for location assessment. Within the settings, the drone was instructed to automatically take images every 2 s to allow the overlap of images needed for the process. The images were also captured in their RAW format to take advantage of the number of colours and complexities that can be represented in this format as opposed to the typical jpeg format which has sixteen times less colour and uses lossless compression, which allows them not to suffer from image-compression artefacts. The images captured could then be transferred to the software maximizing the colours available in the original files.
Following this, the images were transferred to the SfM software, Agisoft Metashape [
57] where the SfM pipeline as depicted in
Figure 7 was used to replicate 3D models of the surveyed pavement section. Within the software, there are options for calibration and compensation for the rolling stock using pre-calibrated value but these would not be applicable for such a low-level camera and survey and could result in worse results. Any self-calibration would also result in an elongated and difficult workflow that would affect the practicality of implementing it in a road authority. The use of ground control can be considered satisfactory for this type of workflow. In the process, images and corresponding point clouds were filtered to ensure that the model would only depict the pavement and not the side elements of the pavement such as the sidewalk or trees. Once the 3D models were generated, the models were transferred to the open-source software, CloudCompare [
58] to examine the defects and identify the distresses existing on the pavement sections. This methodology employed to examine the defects is represented in
Section 2.2.