I am working on a project for my school, the projects are unique and based on real world problems so there really could not be a solution.
So here it goes. I have a bunch of photos of fields, some of which go back ~10 years all at the same locations and supposed to be in the same direction. There is a weather tower at each location and the technician were supposed to stand 10 feet from the tower North and take pictures in the direction of North, North-East, East, etc. Some of these images were taken with an old camera that had like a 4:3 aspect ratio and some are taken on cell phones from many different types in both landscape and horizontal mode. I will attach some of the photos to this for example.
The idea is that we need to find objects that are within 100 meters of the tower that might hinder it's ability to gauge the wind and other weather elements. We should be able to find Distance and Size of objects so we can rate how accurate a tower should be. I have looked up this quite a bit and can't find anything on how to actually do this. Someone pointed to ImageJ and another pointed to ImageAI but I can't figure out how it would find distance even if the images were all standard. From what I have been able to find it shouldn't be a problem to use AI to find distance if we have two pictures at the same angle and we know that distance between the two. But we only have one of each.
We went to one of the towers and took some with one person from our group standing 100 ft away from the camera to get an idea of what 100 feet from the tower should look like. and took pictures in the cardinal directions. I will post some of these as well.
Hi Trevor,
I'm not sure if I understand your problem properly.
The only thing that comes into my mind is to compute a homography in order to remove perspective distortion. After that you can apply it to compute a new image, where you are able to meassure lenghts (but only in that plane).
See screenshot of a lecture that I found in the internet and a link to that lecture, that explains how to compute it (page 35-36).
http://www.cse.psu.edu/~rtc12/CSE486/lecture16.pdf
Additional links:
https://docs.opencv.org/master/d9/dab/tutorial_homography.html
Regards
Olaf
No, so I need an AI to automatically try to find objoecta in the field and to give back the distance the object is from where I am standing and the size. Make sense?
O.K., I give it another try. If you don’t care about accuracy then you could just treat your camera as a simple pinhole camera. The pinhole camera model facilitates a simple geometric model that is central projection (see uploaded figure). From this model one can directly find similar triangles.
Furthermore under perspective projections the cross-ratio is fixed. First you could use the cross-ratio to compute the distance to an object (see this example on wikipedia, equation (2)).
https://en.wikipedia.org/wiki/Cross-ratio#Projective_geometry
After that – I just assume that you know the focal length of your camera (at image acquisition time) – you can use this similar triangle y / f = Y / Z, in order to compute the height of that object. In your case y is the length inside the image, f is the focal length of your camera, Z is the distance to the object, and Y is the height of it.
Don’t forget to use same units within the triangles. All measured pixel coordinates you have to multiply with the size of one pixel of your camera sensor. In case you only know the width and height of your sensor just devide it to the pixel resolution, respectively. This gives you the scale which can be multiplied to a vertical length in pixels, or horizontal length respectively.
For instance let’s assume your camera device has a width of 12 mm and a height of 8 mm, and your pixel resolution is 1024 x 768. Then your horizontal scale is:
sh = 12 / 1024 [mm / pixel] = 12 / 1000 / 1024 [m / pixel]
sv = 8 / 786 [mm / pixel] = 8 / 1000 / 786 [m / pixel]
I think this should work for you.
Cheers!
Olaf
Sorry it took so long. Surely you can understand with times like these. I appreciate your time and effort in trying to help me find a solution. I do believe that there may not be a solution to this problem after more research. The answer you provided though I believe is the closest thing to what would work. However, we do not know the focal length of the cameras. They started with a standard of a certain camera but a couple of years ago the technicians that took the photos started using their cell phones. Now there is a mixture of cameras that took pictures for this project.
Every picture should have a horizon line in the distance where land meets the sky, is there a way to estimate the distance to the horizon and therefore the distance of spots in between the camera and the horizon?
Just to give you a bit of background information that also explains, why, actually, there is no solution to your problem. The 3-dimensional (3D) world is imaged into a 2D image plane of a camera. So you lose one information namely the depth information. Now, in order to reconstruct the 3rd dimension one needs additional information. These information could be
- a second image from a different perspective, a stereo image pair
- or knowledge about the illumination condition (incident angle of sunlight), this is known as shape from shading
- or knowledge of certain geometric properties of the real world
- and others
However, if you want to use the cross-ratio like I proposed, then indeed you need to know (at least) one real world distance in order to calculate one unknown distance (see equation (2) in wikipedia figure). All of them have to lie on the same line in space (at least approximately). I uploaded the example from wikipedia.
Let's consider your picture 'westportrait.jpg' for example. So, in your case the tower is (roughly) at position A. I just draw a red line along the rope on the ground. And I draw the estimated position of the line of infinity (the horizon) in cyan color. The intersection of both lines gives the vanishing point V. The person stands at point B, and position C marks one desired object. So we have four image positions A, B, C, and V.
Since you do know the distance to the person (between real world points A'B') which is 100 yards. And we seek the unknown distance to the object between real world points A'C'. Since they are all on one line it holds:
A'C' = A'B' + B'C' = 100 yards + B'C' = 100 yards + w
Due to the fact that we can see the vanishing point V in the image, the cross ratio equation simplifies (see equation (2) of the wikipedia figure). So, we already have everything we need to know to solve it. See my solution in the uploaded picture 'westportrait.solution.jpg'.
My result for w is 238 yards. Thus the total distance A'C' = 100 yards + 238 yards = 338 yards. Done!
But, to be honest, I doubt that the result is correct. I would rather guess that it must be shorter. However, in principle this solution works.
If you have a GPS device you could check it, or with a tape measure.
Credits Wikipedia:
https://en.wikipedia.org/wiki/File:Cross_ratio_metrology_example.svg
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence