In our 3D laser scanning class at Oregon State University this fall, we spent a lot of time discussing various scan alignment methodologies, techniques, and strategies. Slight deviations in computing scan transformation parameters can have significant influence on scan accuracy, particularly in long range scanning.
One of the first exercises consisted of students picking the centers of targets out of the point cloud for each scan position. Figure 1 shows the pick point from a target where the point cloud is intensity shaded to highlight the differences between the black and white sections of the target. The targets were scanned in separately at higher resolutions. These pick points were then used to register the scans to control coordinates representing the centers of the targets that had previously been established with a total station. It should be noted that the targets were generally within 10-20 m of each scan position.
While grading the student projects, I thought it would be interesting to compare their results to see how much of an impact the difference in pick points by each operator would have on the overall scan transformation. Table 1 shows a comparison of the computed X,Y,Z locations and roll, pitch, and yaw rotations for each scan position obtained by all 8 students. Highlighted values represent outliers where the student had another problem with the data. These outliers were not used to compute the averages and standard deviations.
Note that the pick points had a relatively small influence on the translation parameters (std dev < 5 mm), but had more significant influence on the rotation parameters (std dev up to 0.022 degrees). This is because of the proximity of the targets to the scanner. Had this project covered larger extents with the targets spaced farther, most likely there would be less variation in the rotation angles.
Inclination sensors, in my opinion, are important for validation of scan alignments. An upcoming article in the ASCE Journal of Surveying Engineering discusses their utility in more detail and presents a case-study of how they can be used to spot problems with control. The scanner used to collect the data analyzed here specifies a 0.008 degree standard deviation for inclination sensor measurements. Table 2 shows a comparison of the inclination sensor values to those derived from the targets. Note that these differences are fairly high.
There are a couple reasons for this. 1) A slight difference in the pick point, particularly in the Z-direction, makes a large difference in the rotation angle at close distances. 2) The scanner was setup inside of a concrete tank and all of the targets were a few meters higher in elevation above the scanner. Hence, the targets were somewhat oblique to the scanner in the Z-direction, which would have a large impact on the roll and pitch.
A few things should be noted. First, Leica Cyclone has a feature to automatically extract black and white targets from scan data. However, the targets used for this dataset were not the same type of target. Second, better results would be obtained by least square fitting a plane (or circle) to the target to remove some of the scanner noise. Third, with a black and white target, there will be a slight range bias between the black and white parts of the target. Finally, positioning of the targets and visibility of the targets from each scan position can have an influence on the results.
So the question arises, are these results good? The answer is – it depends! To answer this, we would need to know what the job requirements were. In this case, the research team who were using the data only needed cm-level accuracy, which is satisfied since only close range data (<40 m) would be used. The transformation parameters are reasonably close despite different operators selecting the pick points. However, the differences between the inclination sensors and target values would be of large concern for more precise work. If that were the case, I would recommend constraining the data to the inclination sensor values since the targets were so close to the scanner.
Another important part of this excercise was to help students see the importance of a consistent naming convention for the targets and how that can make processing simpler. In this case when the field crew did find scans for the targets, they did not always capture the TargetIDs (and mixed them up in some cases!). While some software will find the best combination of target pairs between scan positions automatically regardless of name, consistent naming always makes it simpler for the person processing the data when things do not work out as expected.
Acknowledgements
We thank Leica Geosystems and David Evans and Associates for their generous software and hardware support for OSU students. I thank the students who worked diligently in the class!