Hey, at 60 km/h, after 1 hour you are 60 km away, not one! 🙂 Perhaps you meant 1 minute later... Anyway, the animation is nice, but the voice sounds a bit "echoey"/ metallic. Perhaps is my phone, though. Also, the left-side terms in the first equation seem a bit misaligned vertically. Can't finish the video now. More comments (perhaps) later.
Thanks for the detailed feedback! - "at 60 km/h, after 1 hour you are 60 km away, not one!" Good catch! - "... the voice sounds a bit "echoey"/ metallic ..." I will try to fix that in the future. - "the left-side terms in the first equation seem a bit misaligned vertically." Yeah ... that's because of the animation with matrices that follows, I didn't think it was worth spending time on it to fix it. But it's good to know that it is noticeable.
XD omg. next time i will jump to the comments faster. Been beating myself up for 15 min thinking I forgot basic motion principles. :D I even tried converted to miles since he mentioned California (thought maybe he got the unit wrong) :D. The video is top quality though.
"The surprise term is the difference between what you have predicted and what you have measured. The larger this is, the more surprised you are - and the more you question everything that you learned at University."
Nice. You indicate that the Kalman gain term can be between -1 and 1, but only explain the case when it is 0 and when it is 1. I would add a note in the video to a reference describing why it could be -1 or edit the audio to remove this statement.
You make a valid point! While a Kalman gain like "Kx,xdot" can take negative values, "diagonal" Kalman gains like Kxx are always between 0 and 1. I will add this to the video description.
Hey, It was great explanation. I have a question at video time 10:52, where the predicted state x_(t+1) + h has a common term K Which is multiplied with (y_t+1 - Hx_(t+1)), and my question is how significant the measurement state i.e. y_(t) in this calculation. I did not get where is that being considered. And also could you clarify how the current state system is considered i.e. x_(t) as mentioned at video time 5:14.
I just came to your channel, great videos. Have you ever considered making a video about the connection between kalman filters and what happens if you design a luenberger observer with the riccati controller formula?
It's difficult to keep track of what each equation refers to -- please name your variables descriptively. This concept was not visually explained: It used pictures, but it was explained for those who have a strong math background. When advertised as a visual explanation, I hoped for less of a mathematical explanation and more of a logical breakdown of the concept, with the math as a secondary support. Instead It feels like the math is the primary descriptor, and the explanation is very difficult to follow without a strong mathematical background.
Hey there, thanks for watching! I understand where you are coming from. It's hard to calibrate the video to all backgrounds. This video's target are 1st and 2nd year engineering undergrads that have some familiarity with linear algebra and dynamical systems. If you don't come from a math background, it might be hard to follow all the equations in the video, especially towards the end. But hopefully you still get a sense of what Kalman Filters do, how they can be useful, and maybe have a clearer idea of what you should look up next related to Kalman filters. In any case, thanks a lot for leaving a comment, it helps me improve for the next videos.
Sadly the car can not start moving at constant speed from time = 0+. \\\ In the beginning the speed starts from 0 then accelerates (Delta V / Delta t) until it reaches constant speed.
You are absolutely right! Realistically, a car cannot go for a speed of 0 to V in no time (even for Tesla cars that are known for some insane accelerations). However, I have picked a stylized dynamics model here that only involves two state variables (position and velocity only, without acceleration) in order to make the problem two dimensional, and therefore easier to visualize and explain.
@@VisuallyExplained That's what they taught us in the University in Physics and the real life is something else. In electrical engineering we have something called electric current at t = 0 and t = 0+ and t = 0- funny but true
Why dont you take into account inaccurate initial conditions? I mean if initial position and velocity information is incorrect,besides perturbations,inaccuracy coming from initial conditions should propogate as well.Does that included in sigma_t ?
thank you it is very clear !! But .. Excuse me, maybe I didn't understand something! It is obvious that counting one hour and at a speed of 60 km / h, the two distances 0.8 km and 1 km are false and far away! Don't you agree with me that the position should be around 60 km ?? I hope you mean t = 1.00mn and not t = 1.00h
Would it be possible to use the algorithm for tracking the vehicle position in real time. And I'm talking every second rather than hour? Thanks a lot for the great video!
Thanks a lot! We are trying to track the position using a mobile device. So we were looking for an algorithm that uses minimum processing power/ resources and now considering this as an option.
Hi, thank you for sharing your knowledge on this difficult topic! May I ask for the derivation of next sigma = P + A sigma A transpose? I am struggling with the integrals xD Also, A needs to be invertible right? Thank you!!
Sure! This is coming from the fact that x_next = Ax + u, so assuming u is independent from x, we get var(x_next) = var(Ax) + var(u) On the one hand, var(u) = P by assumption On the other hand, var(Ax) = A var(x)A^T (This is always true by the way, as long as the matrix A is deterministic) Since var(x) = Sigma_t, we get the result you stated. Note that we didn't need the fact that A is invertible. (A^T denotes the transpose of A, not the inverse). Hope this helps.
@@VisuallyExplained Thank you for your response!!! Worked through the hassles with monstrous integrals and baffling matrix lemmas (inverse & determinant), truly frightening, but at the same time exciting. Got it, A needs not to be invertible! Also, do you recommend any texts on the history of inventing the Kalman Filter? That must be very fun to read.
I figured it out just the way you speak xD, I am delighted there are such amazingly explained videos on youtube done by a Moroccan. I will be honoured to contact you on Facebook. I have already sent you a friend request.
Why is background music "nice"? Should you not be listening to the lecture? You would catch the 1km vs. 60km error described above more easily without audio disturbances.
Thats weird....the mechanical speedo in my can isnt noisy at all....Maybe if youre using noisy sensors you need to invest in better designed sensors instead of those amazon cheapos. Also, before applying to Tesla please for the love of god, learn to program for failure modes.... The amount of videos of vehicle 'AI' systems freaking out at the moon because it thinks its a traffic light. Or a traffic light, being transported and freaking out because it cant see the color indicator, or at cones again on a truck....what if some youngin programmer decided to have the car slam on the brakes when one of these false positives come through? Yeah, the state of these systems arent actually road safe. 'It works' is NOT a end point. 'We cannot get it to screw up' is. Unfortunately too much has been invested into the 'it works, ship it' industry, that when they discover what we already knew since the 70s (computers only got faster since then, we're still computing in binary kiddies) that TEACHING A COMPUTER TO DRIVE IS ALMOST IMPOSSIBLE!!!! Sleep well knowing these 'it works, good enough, ship it' 3 ton machines are mindlessly following lines, not actually conceptualizing their environment...... Maybe when it runs over a kid instead of some bum, again, you people will listen....but I doubt it....
Hey, at 60 km/h, after 1 hour you are 60 km away, not one! 🙂
Perhaps you meant 1 minute later...
Anyway, the animation is nice, but the voice sounds a bit "echoey"/ metallic. Perhaps is my phone, though.
Also, the left-side terms in the first equation seem a bit misaligned vertically.
Can't finish the video now. More comments (perhaps) later.
Thanks for the detailed feedback!
- "at 60 km/h, after 1 hour you are 60 km away, not one!" Good catch!
- "... the voice sounds a bit "echoey"/ metallic ..." I will try to fix that in the future.
- "the left-side terms in the first equation seem a bit misaligned vertically." Yeah ... that's because of the animation with matrices that follows, I didn't think it was worth spending time on it to fix it. But it's good to know that it is noticeable.
@@VisuallyExplained also your avatar is 1 millimeter too far to the left, unacceptable. xx
XD omg. next time i will jump to the comments faster. Been beating myself up for 15 min thinking I forgot basic motion principles. :D I even tried converted to miles since he mentioned California (thought maybe he got the unit wrong) :D. The video is top quality though.
@@joegriffith1683 (perhaps)
@@Elmaxsivu lol. Star trek, modern SciFi and NASA use kilometres :)
"The surprise term is the difference between what you have predicted and what you have measured. The larger this is, the more surprised you are - and the more you question everything that you learned at University."
This kind of dry humour is peak engineering. Makes the video a bit easier to enjoy, too
I agree that 1 hour 1 km is very confusing, which is slower than the walking speed
Saying something is ‘visually explained’ because equations were displayed visually might be missing the spirit of the title.
this kind of videos remind me of what internet is all about: sharing knowledge. Thanks for the content. I hope internet stopped here
Nice. You indicate that the Kalman gain term can be between -1 and 1, but only explain the case when it is 0 and when it is 1. I would add a note in the video to a reference describing why it could be -1 or edit the audio to remove this statement.
You make a valid point! While a Kalman gain like "Kx,xdot" can take negative values, "diagonal" Kalman gains like Kxx are always between 0 and 1. I will add this to the video description.
I was going to say this. How can you have something that is the ratio of two variances to be negative.
If last year is anything to go by, the world is far from perfect.
Very well explained - Bravo to you
NIce! Can you also make one for Extended Kalman Filters?
i love such videos when i start a new topic. thanks!
Good job. Very clear explanation
Wow thank you. What an exposition!
Best short video for KF so far
Nice :) happy to see your again via your videos 🙂
Fantastic video. Really explained this topic well
Hey, It was great explanation. I have a question at video time 10:52, where the predicted state x_(t+1) + h has a common term K Which is multiplied with (y_t+1 - Hx_(t+1)), and my question is how significant the measurement state i.e. y_(t) in this calculation. I did not get where is that being considered. And also could you clarify how the current state system is considered i.e. x_(t) as mentioned at video time 5:14.
great explanation to establish a general understanding of the topic. Thx!!
I just came to your channel, great videos. Have you ever considered making a video about the connection between kalman filters and what happens if you design a luenberger observer with the riccati controller formula?
thank you to making this video, it's clear enough
It's difficult to keep track of what each equation refers to -- please name your variables descriptively.
This concept was not visually explained: It used pictures, but it was explained for those who have a strong math background. When advertised as a visual explanation, I hoped for less of a mathematical explanation and more of a logical breakdown of the concept, with the math as a secondary support. Instead It feels like the math is the primary descriptor, and the explanation is very difficult to follow without a strong mathematical background.
Hey there, thanks for watching! I understand where you are coming from. It's hard to calibrate the video to all backgrounds. This video's target are 1st and 2nd year engineering undergrads that have some familiarity with linear algebra and dynamical systems. If you don't come from a math background, it might be hard to follow all the equations in the video, especially towards the end. But hopefully you still get a sense of what Kalman Filters do, how they can be useful, and maybe have a clearer idea of what you should look up next related to Kalman filters. In any case, thanks a lot for leaving a comment, it helps me improve for the next videos.
Sadly the car can not start moving at constant speed from time = 0+. \\\ In the beginning the speed starts from 0 then accelerates (Delta V / Delta t) until it reaches constant speed.
You are absolutely right! Realistically, a car cannot go for a speed of 0 to V in no time (even for Tesla cars that are known for some insane accelerations). However, I have picked a stylized dynamics model here that only involves two state variables (position and velocity only, without acceleration) in order to make the problem two dimensional, and therefore easier to visualize and explain.
@@VisuallyExplained That's what they taught us in the University in Physics and the real life is something else. In electrical engineering we have something called electric current at t = 0 and t = 0+ and t = 0- funny but true
Superb explanation
Thank you !
Why dont you take into account inaccurate initial conditions? I mean if initial position and velocity information is incorrect,besides perturbations,inaccuracy coming from initial conditions should propogate as well.Does that included in sigma_t ?
thank you
it is very clear !!
But .. Excuse me, maybe I didn't understand something!
It is obvious that counting one hour and at a speed of 60 km / h, the two distances 0.8 km and 1 km are false and far away!
Don't you agree with me that the position should be around 60 km ??
I hope you mean t = 1.00mn
and not t = 1.00h
Hey there, thanks for watching carefully! You are right, it’s a typo. It should be 1 minute instead of 1 hour exactly as you pointed out.
Very good video. Thank you!
Would it be possible to use the algorithm for tracking the vehicle position in real time. And I'm talking every second rather than hour?
Thanks a lot for the great video!
Absolutely! The algorithm is super fast, you can even use at a higher frequency than once every second (if your measurement instruments can keep up).
Thanks a lot! We are trying to track the position using a mobile device. So we were looking for an algorithm that uses minimum processing power/ resources and now considering this as an option.
Nice video. May we know if P and Q should be given, or are they calculated online?
the first few seconds were true for me, except the whole Tesla thing >_
Feedback. Please give the variables in the equation full names and descriptive names instead of single letter names such as “Q”. Thanks.
Q is just the covariance of the noise, before learning KF you need to have some probability knowledge otherwise its gonna be hard
@@ayasera_9501 Is is that difficult to explain sth from first principles in a simple manner?
@@satellite964 i think the presenter just assumed the watcher might have this previous knowledge
very nice , thank you
Thank you very much!
Hello sir, I would like to cite this video in my research paper can you kindly explain me how to? thank you.
I am sorry to tell you that, but it is probably smth wrong with your research methods if you want to cite the youtube video
Great video!
Hi, thank you for sharing your knowledge on this difficult topic! May I ask for the derivation of next sigma = P + A sigma A transpose? I am struggling with the integrals xD
Also, A needs to be invertible right?
Thank you!!
Sure!
This is coming from the fact that
x_next = Ax + u,
so assuming u is independent from x, we get var(x_next) = var(Ax) + var(u)
On the one hand, var(u) = P by assumption
On the other hand, var(Ax) = A var(x)A^T (This is always true by the way, as long as the matrix A is deterministic)
Since var(x) = Sigma_t, we get the result you stated.
Note that we didn't need the fact that A is invertible. (A^T denotes the transpose of A, not the inverse).
Hope this helps.
@@VisuallyExplained Thank you for your response!!!
Worked through the hassles with monstrous integrals and baffling matrix lemmas (inverse & determinant), truly frightening, but at the same time exciting.
Got it, A needs not to be invertible!
Also, do you recommend any texts on the history of inventing the Kalman Filter? That must be very fun to read.
Fantastic
Thanks! :)
Ideal situation 1 km in one hour..far from ideal
nice video
thanks
Cool
Good explanation, i bet you are Moroccan 😅, let me know if you are.
Spot on! What gave it away?
I figured it out just the way you speak xD, I am delighted there are such amazingly explained videos on youtube done by a Moroccan.
I will be honoured to contact you on Facebook.
I have already sent you a friend request.
Thanks for the awesome explanation 👏👏
I would suggest adding some nice background music, it would be great
Thanks for watching! I will try to incorporate your feedback in next videos.
Why is background music "nice"? Should you not be listening to the lecture? You would catch the 1km vs. 60km error described above more easily without audio disturbances.
Thats weird....the mechanical speedo in my can isnt noisy at all....Maybe if youre using noisy sensors you need to invest in better designed sensors instead of those amazon cheapos. Also, before applying to Tesla please for the love of god, learn to program for failure modes.... The amount of videos of vehicle 'AI' systems freaking out at the moon because it thinks its a traffic light. Or a traffic light, being transported and freaking out because it cant see the color indicator, or at cones again on a truck....what if some youngin programmer decided to have the car slam on the brakes when one of these false positives come through? Yeah, the state of these systems arent actually road safe. 'It works' is NOT a end point. 'We cannot get it to screw up' is. Unfortunately too much has been invested into the 'it works, ship it' industry, that when they discover what we already knew since the 70s (computers only got faster since then, we're still computing in binary kiddies) that TEACHING A COMPUTER TO DRIVE IS ALMOST IMPOSSIBLE!!!! Sleep well knowing these 'it works, good enough, ship it' 3 ton machines are mindlessly following lines, not actually conceptualizing their environment...... Maybe when it runs over a kid instead of some bum, again, you people will listen....but I doubt it....
I didnt get anything, really bad explanation
bad audio
Hey! Thanks for the feedback! I tried to improve the audio in the more recent videos, let me know what you think
0:04 Worst outcome of graduation.
:-)