1
$\begingroup$

I have the orientation of a 3D spatial object represented by a unit quaternion:

$$ q = a_1 + a_2 i + a_3 j + a_4 k $$ $$ \|q\| = (a_1^2 + a_2^2 + a_3^2 + a_4^2)^{1/2} = 1 $$

I'd like to perturb this orientation slightly. If this displacment is small, I think I can get away with

$$ q'_2 = q + \Delta (u_1 + u_2 i + u_3 j + u_4 k) $$

where $\Delta << 1$ and the $u$ are uniform random numbers in the range of $[-1,1]$. $q'_2$ is obviously not a unit quaternion anymore so I would normalize it by

$$ q_2 = \frac{q'_2}{\|q'_2\|} $$

As $\Delta$ gets larger, I feel that this will become an increasingly worse approximation. Is there a better way to get a random orientational displacement?

$\endgroup$
  • $\begingroup$ You haven't specified what sort of distribution you want. Do you want (i) a uniformly distribution over all rotations with angle less than $\Delta$, (ii) a Gaussian-like falloff with some scale parameter $\Delta$, (iii) something else? In any case, you can w.l.o.g. pick from a distribution centered at the unit quaternion, and then multiply it with your $q$. $\endgroup$ – Rahul Aug 22 '13 at 18:16
  • $\begingroup$ @RahulNarain The end goal is a rotation step in a Monte-Carlo simulation, so the emphasis would be on the ease of implementation. I worry however, that my method above wouldn't sample evenly around my starting $q$. In regards to your other comment, I guess my question could be reformulated to, "how does one generate a distribution around the unit (did you mean identity?) quaternion?" $\endgroup$ – Hooked Aug 22 '13 at 18:26
  • $\begingroup$ Well, the easiest thing might be to sample uniformly from a hyperspherical cap around the identity quaternion (that is indeed what I meant, thanks). That is, pick $\theta$ from $[0,\Delta]$ with the appropriate distribution, pick $(x,y,z)$ uniformly from the unit $2$-sphere, and use the quaternion $\cos\theta + (ix+jy+kz)\sin\theta$. $\endgroup$ – Rahul Aug 23 '13 at 2:03
  • $\begingroup$ @RahulNarain That's a good idea. If you write it as an answer I'll accept it. Just to be clear, given your $q_R(\Delta)$ sampled from the hyperspherical cap, would the distribution of new perturb quaternions simply be $q q_R(\Delta)$? $\endgroup$ – Hooked Aug 23 '13 at 13:37
2
$\begingroup$

You would like to sample a unit quaternion $q'$ from a normal distribution with mean $q$ and some variance $\omega$?

Background:

A rotation matrix $R_{a,b}$ can be seen as transformation which rotates a point $p$ from reference frame $b$ to the reference frame $a$: $p_a = R_{a,b}p_b$. Obviously, the same holds for unit quaterions representing rotations, so we could write $q_{a,b}$ to emphasis that this quaternion represents a rotation from frame $b$ to frame $a$ (as in $p_a = q_{a,b}\cdot p_b \cdot q_{a,b}^{-1}$).

The first question, we have to ask ourselves in which frame we'd like to sample, i.e. perturb $q_{a,b}$: frame $a$ or frame $b$? This really depends on the application. Let us simply assume $a$ for now.

Approach:

A good approach would be to sample in the tangent space at frame $a$. To be more precise, we would sample in the tangent space around the identity $\exp(\omega)$ with $\omega$ been drawn from a 3-dimensional zero-mean Normal distribution. Then, we apply this quaternion $\exp(\omega)$ to $q_{a,b}$. Thus the resulting quaterion is:

$$ q' = \exp(\omega)\cdot q_{a,b} \quad (\text{with} \quad \omega \in \mathbb{R}^3)$$

Here $\cdot$ is the quaternion multiplication, and $\exp$ the unit quaternion exponential: $\exp(\omega) = (\cos(\theta), \frac{1}{\theta}\sin(\theta)\omega )$ with $\theta = ||\omega||$.

(If we'd like to sample in frame $b$, we would do $q' = q_{a,b}\cdot \exp(\omega)$

$\endgroup$
  • $\begingroup$ This is essentially the same idea voiced by @RahulNarain, correct? If so, I'll accept this so we can mark it closed. If not, please let me know the difference. $\endgroup$ – Hooked Aug 27 '13 at 13:29
  • $\begingroup$ Well yes, RahulNarain approach appears to be similar, though not sure whether there are subtle differences. In terms of implementation, it is probably easier to sample $\omega$ from a 3-dim Gaussian than to implement a uniform sampling on the 2-Sphere. Btw, no rush in accepting the answer. Take your time in digesting, and/or implementing it! $\endgroup$ – B0rk4 Aug 28 '13 at 5:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.