Solutions are not public but my solutions to the lab exercises that I'm currently working on can be found here github.com/conorhennessy/SLAM-Course-Solutions Let me know if you spot anything wrong
Thank you for this great material. Unscented transformation transfer points from state space to measurement space. In space update equation we summarise mean in state space with K multiplied delta z in measurement space. It's possible because K contains Pxy which contains information about transformation from measurement space to state space?
Hi Professor, You have mentioned that the unscented transform gives a way to transforming a Gaussian distribution through a non-linear function. I wonder if we can still apply the same technique to transform a NON-Gaussian distribution? In other words, if we want to non-linearly transform a non-Gaussian distribution, can we calculate the sigma point in the same way and obtain the same accuracy in estimating the moments of the transformed distribution? Thank You
Good lecture Prof Stachniss. Can you tell me in regards to the simple example of tracking a noisy sine wave with changing amplitude and frequency, will a UKF or particle filter perform better than the EKF in terms of accuracy and robustness against divergence? I have generally seen the EKF perform poorly at tracking a changing amplitude....
ปีที่แล้ว +1
Has anyone filtered quaternion measurement with ukf? Like measuring orientation in form of quaternion and having states of quat, velocity of quat, acceleration of quat?
It is good to explain easily, but I wonder if the UKF calculation amount is "Slightly slower than the EKF" compared to EKF even if the state dimension is large.
Is it true that the unscented Kalman Filter is unscented because it doesn't stink? The story I heard is that the graduate students that developed it thought their professor's EKF idea stunk and used "unscented" to distinguish their algorithm from their professor's.
You could have just google it, couldnt you? From wiki - " its creator Jeffrey Uhlmann explained that he came up with the name after noticing unscented deodorant on a coworker's desk."
At 34:17 in the Extended Kalman filter equations (3) and (4) the noise covariances seems that Rt represents the process noise and Qt represents the Measurement noise. Is this true? Otherwise they should be switched in these equations.
I don't understand the choice of the weightings. It's stated that the weights should sum to 1. Yet the weights taken (24:38) have both w_m and w_c the same, except w_c[0] is (1- a^2 + b) bigger, hence they both can't sum to 1. Am I misunderstanding something or do the weights not need to sum to 1?
Since UKF use unscented transform twice to propagate the sigma points throw function g and function h, one for the estimate step and another for the Kalman Gain step, is it possible to use only one unscented transform to propagate the original sigma points throw g*h to calculate the Kalman Gain?
As UKF pass sigma points through non-linear function,which is similar to Particle Filter passing particles through non-linear function , how is UKF compare to Particle Filter?
Can covariance matrix have an element which would be negative or the square root of covariance matrix will have imaginary value(this case will lead to sigma points coordinates with imaginary component)?
then the state transform function will be just treated as an identity function as the professor has stated and the observation function is treated as non linear.
yeah and when there is nothing on it like a hat or a bar then its our guessss?????? when you calculate sigma points you wrote square of a covariance,... which one is it hat bar ... some alien covariance.
Professor, you know what! Attach exercise materials, course notes, etc... and this would be the greatest course in the web.
+Msa Chehadah Couldn't agree more!
+Msa Chehadah Oh look what I just found: ais.informatik.uni-freiburg.de/teaching/ws13/mapping/
Do u have corrections for these exercises ?
Solutions are not public but my solutions to the lab exercises that I'm currently working on can be found here github.com/conorhennessy/SLAM-Course-Solutions
Let me know if you spot anything wrong
Thank you for this great material. Unscented transformation transfer points from state space to measurement space. In space update equation we summarise mean in state space with K multiplied delta z in measurement space. It's possible because K contains Pxy which contains information about transformation from measurement space to state space?
Hi Professor,
You have mentioned that the unscented transform gives a way to transforming a Gaussian distribution through a non-linear function. I wonder if we can still apply the same technique to transform a NON-Gaussian distribution?
In other words, if we want to non-linearly transform a non-Gaussian distribution, can we calculate the sigma point in the same way and obtain the same accuracy in estimating the moments of the transformed distribution?
Thank You
Amazing
Good lecture Prof Stachniss. Can you tell me in regards to the simple example of tracking a noisy sine wave with changing amplitude and frequency, will a UKF or particle filter perform better than the EKF in terms of accuracy and robustness against divergence? I have generally seen the EKF perform poorly at tracking a changing amplitude....
Has anyone filtered quaternion measurement with ukf? Like measuring orientation in form of quaternion and having states of quat, velocity of quat, acceleration of quat?
It is good to explain easily, but I wonder if the UKF calculation amount is "Slightly slower than the EKF" compared to EKF even if the state dimension is large.
Is it true that the unscented Kalman Filter is unscented because it doesn't stink? The story I heard is that the graduate students that developed it thought their professor's EKF idea stunk and used "unscented" to distinguish their algorithm from their professor's.
You could have just google it, couldnt you? From wiki - " its creator Jeffrey Uhlmann explained that he came up with the name after noticing unscented deodorant on a coworker's desk."
😂 thanks; i didn’t know that.
At 34:17 in the Extended Kalman filter equations (3) and (4) the noise covariances seems that Rt represents the process noise and Qt represents the Measurement noise. Is this true? Otherwise they should be switched in these equations.
Can UKF be implemented for state estimation of a partially unknown non-linear system?
I don't understand the choice of the weightings. It's stated that the weights should sum to 1. Yet the weights taken (24:38) have both w_m and w_c the same, except w_c[0] is (1- a^2 + b) bigger, hence they both can't sum to 1. Am I misunderstanding something or do the weights not need to sum to 1?
Since UKF use unscented transform twice to propagate the sigma points throw function g and function h, one for the estimate step and another for the Kalman Gain step, is it possible to use only one unscented transform to propagate the original sigma points throw g*h to calculate the Kalman Gain?
Can anyone please explain about "constrained unscented kalman filter" ? What the word constrained here denotes
one question, why the ukf use the square of the sum of dimensionality plus the scaling parameter? namely sqrt(n+lambda)
Which exactly is the nonlinear function g? Do we "create" it, based on what our data is?
intuitive
so ukf is a smart version of particle filter?
No. UKF is a Gaussian filter, not really related to a PF.
As UKF pass sigma points through non-linear function,which is similar to Particle Filter passing particles through non-linear function , how is UKF compare to Particle Filter?
One key difference is that the UKF always goes back to the Gaussian belief, the PF does not. There are several other differences as well ....
i am not clear in y=h(x) how can i map ?will i use any specific nonlinear model or not...if i use which is in your lecture...don't define it..
Subtitle please
Can covariance matrix have an element which would be negative or the square root of covariance matrix will have imaginary value(this case will lead to sigma points coordinates with imaginary component)?
Covariance matrices are positive (semi)-definite, that means that determinant is not negative
Awesome!
What if my state transform function is linear, but my observation function is non-linear?
then the state transform function will be just treated as an identity function as the professor has stated and the observation function is treated as non linear.
Do we have to apply inverse unscented transform after the nonlinear system to match sigma points with original sigma point characteristics?
No
what does bel mean in bel(x)?
bel(x) = p(x | data)
46:40 how do you get the 'true gaussian' on the left?
Watch 49:40 . Mean and Covar of the 'true gaussian' are determined by sampling "a lot" of points.
Yes
yeah and when there is nothing on it like a hat or a bar then its our guessss?????? when you calculate sigma points you wrote square of a covariance,... which one is it hat bar ... some alien covariance.
please provide subtitle with the videos its worth spreading, anyone have subtitle for this video ?
the auto generation also is not appear