I need to estimate a camera orientation given only its relative orientation with regard to some other cameras, and where the relative orientations are not exact. For example, I would like to arrive at a linear system of the form shown below, so that I can solve for C2 via least squares: C1 = C0 + R01 C2 = C1 + R12 C2 = C0 + R02 here Cx is a 3D rotation describing the orientation of camera x and Rxy is the relative orientation between cameras x and y (e.g. the first equations says that the orientation of camera 1 is given by camera 0 rotated by R01). Note that this is only a small example, in the real case, I need to solve for 10+ cameras. Now if I could form equations like this, I would be able to solve them easily. The trouble is, if I use, say, quaternions to describe the rotations, then multiplication rather than addition should be used to combine rotations, so the equations would instead look like: C1 = C0 * R01 C2 = C1 * R12 C2 = C0 * R02 and this is no longer a set of linear equations that I can solve with least-squares (e.g. SVD). (I think) I'm hoping that someone can suggest how to 'linearize' this problem. For example, is there is a way of expressing rotations so that combined rotations are expressed as additions? Euler angles kind of do this, but I'm pretty sure they will misbehave and are not viable for this problem. Thanks for any tips, Jason.

Hi Ksero,

I never said anything about absolute orientations - I only want to know C2 with respect to, say, C1.

Given that the R's carry error, the solution for C2 is not unique, and that's why I want to apply least squares to find an optimal solution for C2. (think of the case where I have 20 equations, not three).

When the equations are expressed as products, the system is non-linear, and so is not (?) amenable to a least-squares solution.

So the original question remains: is there a way to express compound rotations as additions, without the perils of Euler angles?

Thanks,

Jason.

Hi Ksero,

Thanks for your efforts. I'm a little annoyed that I didn't work this out for myself now. 8-).

Probably a silly question, but why do you pre-multiply both sides by A^T?

Also, redarding your comment about adapting the solver, I was hoping to simply apply the same form of this system of equations, but with four times the number of rows. i.e. one row each for the unit-quaternion parameters. I figure this is OK, since these parameters are independent. Is this what you meant?

Thanks again,

Jason.

Well, it's not a game, but these game development sites seem to have quite a lot of active discussion about 3D geometry...

Familiar with photogrammetry? I'm coding up an application that creates a 3D model from a bunch of photos.

Well, I coded this up, only to realise too late that it won't work. To use a sample equation:

C3 = R13 * C1

We could solve for each of the (x, y, z, w) parameters of C3 in the way described above if it were true to say that

C3.x = R13.x * C1.x
C3.y = R13.y * C1.y

etc., but this is not the way that quaternions multiply. Even if I expand out the quaternion multiplications and include it all in one big set of equations, the rows will be coupled since a row which referenes a quaternion x will also reference a y, etc.

Alas, need to rethink this again.