Relative motion between Quaternions
Show older comments
I'm working on Quaternions. I wanna find out the relative motion between quaternions in MATLAB.
For example, I've
QuatA = [1283842729 -831786718 1290396116 1271562329] and
QuatB = [1313190417 1274713520 -845615766 1306114950]
How can I find how much QuatB is rotated with respect to QuatA?
Any help is appreciated.
Thanks.
5 Comments
James Tursa
on 24 Sep 2015
What do these quaternions represent?
BAN
on 24 Sep 2015
James Tursa
on 24 Sep 2015
Edited: James Tursa
on 24 Sep 2015
Angular motion? You mention rotation, which normally means I would have expected unit quaternions, but your quaternions are not unit. Hence I am trying to understand what exactly they represent.
BAN
on 24 Sep 2015
BAN
on 24 Sep 2015
Answers (2)
James Tursa
on 24 Sep 2015
Edited: James Tursa
on 24 Sep 2015
3 votes
In general, for unit quaternions you would multiply the conjugate of one times the other, and then extract the angle and rotation axis from that. Be sure to do a quaternion multiply, not a regular multiply. Also be sure the quaternion multiply routine you are using assumes the scalar is in the same spot (1st or 4th) as the quaternions you have. You would have to divide your quaternions above by their respective norms to apply this technique, but not knowing what these quaternions represent I can't advise if this is the correct procedure or not.
8 Comments
BAN
on 24 Sep 2015
James Tursa
on 24 Sep 2015
Edited: James Tursa
on 24 Sep 2015
E.g., assuming quaternion scalar is in the 1st spot to match the Aerospace Toolbox:
QuatA = [1283842729 -831786718 1290396116 1271562329];
QuatB = [1313190417 1274713520 -845615766 1306114950];
QuatAu = QuatA / norm(QuatA);
QuatBu = QuatB / norm(QuatB);
QuatDiff = quatmultiply(quatconj(QuatAu),QuatBu);
Then QuatDiff will be equivalent to [cos(ang/2) sin(ang/2)*eigx]
Where ang is the angle you are looking for, and eigx is a 1x3 unit vector about which the rotation is taking place. So this can recover the angle:
ang = 2 * atan2(norm(QuatDiff(2:4)),QuatDiff(1));
Again, this all assumes that normalizing the quaternions has meaning in your setup, and that the scalar is in the 1st spot.
BAN
on 3 Oct 2015
James Tursa
on 3 Oct 2015
Edited: James Tursa
on 3 Oct 2015
Sorry I was not more clear. QuatDiff(1) is equal to cos(ang/2), and QuatDiff(2:4) is equal to sin(ang/2)*eigx. So, only the last three elements of the 1x4 vector are used for the sin(ang/2)*eigx part. I could have written it like this (with a comma):
[ cos(ang/2) , sin(ang/2)*eigx ]
BAN
on 4 Oct 2015
manikya valli
on 16 Feb 2016
I am having similar problem. I want to find difference between reference quaternion and measured quaternion. I want to verify by converting into euler angles. But the difference quaternion when converted back to euler angles is not expected euler angles. Here is the code: q1=angle2quat(deg2rad(30),deg2rad(40),deg2rad(50));
q2=angle2quat(deg2rad(10),deg2rad(10),deg2rad(10));
q3=quatmultiply(quatconj(q2),q1);
[r,p,y]=quat2angle(q3);
rad2deg(r);rad2deg(p);rad2deg(y); result I am getting is
r,p,y=11.6921,33.0794,34.6602
but I have to get 30-10,40-10,50-10=20,30,40 respectively. Why I am not getting this result?
Aitor Burdaspar
on 17 Dec 2019
Hello Manikya,
I have exactly the same problem as yours. Did you finally solve it?
Best regards,
James Tursa
on 17 Dec 2019
Edited: James Tursa
on 17 Dec 2019
There is a fundamental misunderstanding of how to work with quaternions and Euler Angle sequences. In the first place, Euler Angle sequences do not behave linearly. That is, if you have a (yaw1,pitch1,roll1) sequence followed by a (yaw2,pitch2,roll2) sequence, you should not be expecting the result to be the same as a (yaw1+yaw2,pitch1+pitch2,roll1+roll2) sequence. Sequential rotations simply do not behave that way. So the subtractions you are doing (30-10, 40-10, 50-10) to get your "expected" result is simply not true with regards to sequential rotations.
In particular, I will use a couple of example coordinate frames to illustrate the issue.
Suppose q1 and q2 are both ECI_to_BODY quaternions. I.e.,
q1 = ECI_to_BODY1 (BODY1 being the BODY frame orientation for q1)
q2 = ECI_to_BODY2 (BODY2 being the BODY frame orientation for q2)
Then
q3 = conj(ECI_to_BODY2) * ECI_to_BODY1
= BODY2_to_ECI * ECI_to_BODY1
= BODY2_to_BODY1
So q3 represents a rotation between the two end BODY frames for the quaternions. Note that q3 is not another ECI_to_BODY quaternion. Since it is not another ECI_to_BODY quaternion, you should not expect the angles associated with it to be the same as what you get from a straight subtraction of your q1 and q2 angles.
Side note: The default order for the angles is y,p,r and not r,p,y.
Mo
on 3 Jun 2018
0 votes
How is this method used for a matrix? It seems "norm(A)" for a quaternion matrix is different. I appreciate any help.
1 Comment
Luca Pozzi
on 11 Jun 2018
If A is an nx4 matrix (a 'column' of quaternions) the following code should work:
for i=1:size(A,1) % for each row of A
Anorm(i,:)=norm(A(i,:)); % dividing a row by its norm
end
norm(A) returns the norm of the whole matrix A, the for cycle here above allows you to normalize every single row of the matrix (i.e. every quaternion) instead. Hope this helps!
Categories
Find more on Quaternion Math in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!