Camera offset survey
- Thread starter Jim Campi
- Start date
Yesterday I used photogrammetry on my first real job. I was cutting out 6 acres from a 20 acre pasture. There was a small house, a few cross fences, a power line and a storage building, all of which would have needed to be located by an offset technique or by total station. I divided this into 3 camera offset survey projects: House, Power line and Storage. I shot 10 photos around the house locating the deck, steps and yard fence in that scene. Then I shot only 3 photos of the power line. This is the minimum number of photos required and provides no redundancy nor residuals. I have aerial images to confirm the power poles though. And for the storage building I had four images. I was very pleased with the results.
Today I wanted to test a theory based on observation with the steel structure survey from my video. It appears that the relative position of points in close range photogrammetry with the LS exceeds the accuracy of RTK alone. To demonstrate this, I place five quarters on my concrete driveway. Four of them were roughly in the shape of a square with the fifth being in the middle, roughly at the intersection of the diagonals. I then took four photo points around the quarters. My photo points were between 15-17 feet from the center of the box - very close. I only recorded with verify on set to a confidence of 10, with it set to "by time" for 10 epochs (at 2Hz) and validate on with a minimum of 2 engines. So each point was only collected for about 20 seconds from a base that was only 130 feet away. I would expect these points to be centimeter level horizontal and perhaps as much as 2 centimeters vertical. Given this sort of accuracy for RTK, what would we expect the accuracy of the five points to be? From my earlier, brief experience, I would expect it to be much better. I taped the distance from quarter to quarter and noted it. I then inversed the distances between the quarters determined from photogrammetry. Here is how it shaped up:
Q1-Q2
1.825' (taped)
1.829' (photo)
Q1-Q3
2.875' (taped)
2.872' (photo)
Q1-Q4
2.240' (taped)
2.234' (photo)
Q1-Q5
1.545' (taped)
1.547' (photo)
Q2-Q3
2.025' (taped)
2.021' (photo)
Q2-Q4
2.875' (taped)
2.873' (photo)
Q2-Q5
1.310' (taped)
1.307' (photo)
Q3-Q4
2.040' (taped)
2.043' (photo)
Q3-Q5
1.330' (taped)
1.325' (photo)
Q4-Q5
1.575' (taped)
1.575' (photo)
Maximum difference between taped and inverse: 0.006'
I did not take the time to determine vertical differences to verify results from photogrammetry, but the driveway is fairly level here. The heights of the five points determined by photogrammetry were:
Q1 9.051'
Q2 9.060'
Q3 9.074'
Q4 9.074'
Q5 9.061'
These results are amazing, but must be viewed in context. The relative position of this 5 point cluster to the rest of the project will be limited to the accuracy of RTK. These points are not 0.005' accurate to the Earth or even to my base. But, they do show incredible relative accuracy within the photo offset project. I would expect that the position of these points is more accurate than one of the single RTK positions used to locate the camera, as the adjustment uses the combined positions of the various photos - an average of the 4 photo points.
Also, I was able to use very good photogrammetric techniques. I surrounded these points with photos. Had I been limited to a narrow field of view, the results would not have been as good. Sometimes this limitation is very real, as it is impossible to work entirely around a scene of interest. I was very close to the points, this is also a limitation that is not always practical. Although, I will say I've had good results at distances much farther than 20 feet from the scene. Point geometry makes a difference in those cases. Having points that vary in depth on longer distances makes a big difference.
Today I wanted to test a theory based on observation with the steel structure survey from my video. It appears that the relative position of points in close range photogrammetry with the LS exceeds the accuracy of RTK alone. To demonstrate this, I place five quarters on my concrete driveway. Four of them were roughly in the shape of a square with the fifth being in the middle, roughly at the intersection of the diagonals. I then took four photo points around the quarters. My photo points were between 15-17 feet from the center of the box - very close. I only recorded with verify on set to a confidence of 10, with it set to "by time" for 10 epochs (at 2Hz) and validate on with a minimum of 2 engines. So each point was only collected for about 20 seconds from a base that was only 130 feet away. I would expect these points to be centimeter level horizontal and perhaps as much as 2 centimeters vertical. Given this sort of accuracy for RTK, what would we expect the accuracy of the five points to be? From my earlier, brief experience, I would expect it to be much better. I taped the distance from quarter to quarter and noted it. I then inversed the distances between the quarters determined from photogrammetry. Here is how it shaped up:
Q1-Q2
1.825' (taped)
1.829' (photo)
Q1-Q3
2.875' (taped)
2.872' (photo)
Q1-Q4
2.240' (taped)
2.234' (photo)
Q1-Q5
1.545' (taped)
1.547' (photo)
Q2-Q3
2.025' (taped)
2.021' (photo)
Q2-Q4
2.875' (taped)
2.873' (photo)
Q2-Q5
1.310' (taped)
1.307' (photo)
Q3-Q4
2.040' (taped)
2.043' (photo)
Q3-Q5
1.330' (taped)
1.325' (photo)
Q4-Q5
1.575' (taped)
1.575' (photo)
Maximum difference between taped and inverse: 0.006'
I did not take the time to determine vertical differences to verify results from photogrammetry, but the driveway is fairly level here. The heights of the five points determined by photogrammetry were:
Q1 9.051'
Q2 9.060'
Q3 9.074'
Q4 9.074'
Q5 9.061'
These results are amazing, but must be viewed in context. The relative position of this 5 point cluster to the rest of the project will be limited to the accuracy of RTK. These points are not 0.005' accurate to the Earth or even to my base. But, they do show incredible relative accuracy within the photo offset project. I would expect that the position of these points is more accurate than one of the single RTK positions used to locate the camera, as the adjustment uses the combined positions of the various photos - an average of the 4 photo points.
Also, I was able to use very good photogrammetric techniques. I surrounded these points with photos. Had I been limited to a narrow field of view, the results would not have been as good. Sometimes this limitation is very real, as it is impossible to work entirely around a scene of interest. I was very close to the points, this is also a limitation that is not always practical. Although, I will say I've had good results at distances much farther than 20 feet from the scene. Point geometry makes a difference in those cases. Having points that vary in depth on longer distances makes a big difference.
Now for a trickier example, one that I'm sure that surveyors will want to employ. Locating a point under a tree. Today I attempted locating a point by photogrammetry such as this. The point was in the wide open, but I pretended it was under canopy, this way I could also directly tie the point with RTK to compare the results. I limited myself to about 90° field of view and I stayed at a distance of 50 feet from the point with my photo points, supposing that if the point were under a tree, I would need to be some distance away to get a good position with the photo point. I took four photos.
I am finding it necessary to have a cross of points for best results the points don't need to be equidistant, but you need to have three levels of depth. Near, Middle, Far. Preferably these points should be visible from 90° angles. See photo below. More to follow.
I am finding it necessary to have a cross of points for best results the points don't need to be equidistant, but you need to have three levels of depth. Near, Middle, Far. Preferably these points should be visible from 90° angles. See photo below. More to follow.
Notice in my sketch that the distant object becomes the right object, the near becomes left and the right and left become near and far (respectively) as we rotate around the scene. This helps us get good horizontal coordinates. In my example I didn't have a good near mark to work with.
All of these photos are 45-50 feet from the pole. I put a sharp point on the top of the pole. A spherical target would be better, but this allowed met to precisely pick the tip of the point. The photo point was 0.127' from the RTK position I determined for this point. The height was 0.054' from the height I determined. There are things I would likely do to improve the accuracy, but I wanted to display a realistic field condition - longer range, narrow field of view. As a surveyor, I would not rely on this for boundary. I like to chase hundredths of a foot. But I believe there are times that this could be used for boundary under more favorable conditions. Better geometry, closer to the point. I also would not want to use photo points that have degraded accuracy due to multipath. If you are under canopy, attempting to photo offset a point against a tree, I think you'll be frustrated by the results, but I have no real experience with that yet. Furthermore, image resolution, not necessarily pixel count but the ability to resolve things in photos is important. If your targets are in the shade and the camera is looking into the sun, it will be very difficult or impossible to see the object. Look for opportunities to get good lighting for this process.
All of these photos are 45-50 feet from the pole. I put a sharp point on the top of the pole. A spherical target would be better, but this allowed met to precisely pick the tip of the point. The photo point was 0.127' from the RTK position I determined for this point. The height was 0.054' from the height I determined. There are things I would likely do to improve the accuracy, but I wanted to display a realistic field condition - longer range, narrow field of view. As a surveyor, I would not rely on this for boundary. I like to chase hundredths of a foot. But I believe there are times that this could be used for boundary under more favorable conditions. Better geometry, closer to the point. I also would not want to use photo points that have degraded accuracy due to multipath. If you are under canopy, attempting to photo offset a point against a tree, I think you'll be frustrated by the results, but I have no real experience with that yet. Furthermore, image resolution, not necessarily pixel count but the ability to resolve things in photos is important. If your targets are in the shade and the camera is looking into the sun, it will be very difficult or impossible to see the object. Look for opportunities to get good lighting for this process.
One more low-tech sketch. When working with verticals, think of your high school geometry days and create similar triangles. Also, changing the height of the camera helps give better geometry for resolving elevations.
Note: I'm posting this because I'm learning, just like I'm sure many of you will be as you buy the upgrade. My three weeks of OTJ training probably doesn't qualify me as an expert, but I'm finding that it's not terribly difficult to get good results.
One thing to note in the above examples, I used a control point with the quarters, although I don't think it was necessary. I did not use a control point with the pole. My base was in one photo and I added it to see if there was any benefit. In this case there was little benefit. Horizontal improved slightly to 0.121' from RTK, vertical was a bit worse at -0.17'. Error estimates did improve.
One thing to note in the above examples, I used a control point with the quarters, although I don't think it was necessary. I did not use a control point with the pole. My base was in one photo and I added it to see if there was any benefit. In this case there was little benefit. Horizontal improved slightly to 0.121' from RTK, vertical was a bit worse at -0.17'. Error estimates did improve.
Mikhail Drakin
Developer
Note for the test with quarters: when assessing the relative accuracy as opposed to exact georeferenced accuracy, take into account that the whole scene may be a bit rotated due to georeference deficiency, so it's more correct to speak only about 3D distance accuracy, not horizontal or vertical one. That is, having some bad RTK position (or artificially editing a point) you could, in principle, obtain result when the concrete plane with quarters was 45 degree slant, but 3d distances between quarters still closely matched taped ones.
Mikhail Drakin
Developer
As for the "canopy" test: we wouldn't expect hundredths of a foot accuracy with our built-in camera even at bright day, but in our model situation when one has only LS handy and the need to measure a point under a tree, the multipath effects could to some extent be levelled by taking more photos (say, 6-8) to help filter the worst positions with photogrammetry adjustment.
I considered this as well. I think you are right about more photos in a bad location possibly improving the accuracy of an offset point, I just haven't done it yet.
I need to consider how to do this, but I would like to test the rotation of close range photo points.
Probably I will use a total station on a long baseline with known direction and orient the edge of a small plate to the base line. I'll have to think about it.
Probably I will use a total station on a long baseline with known direction and orient the edge of a small plate to the base line. I'll have to think about it.