This is such a good series by MATLAB. Thanks alot. A video about Multisensor data fusion for LiDAR, Ulrasonic and Infrared with the help of Kalman Filter implementation for Ranging Application in MATLAB/Simulink would be helpful alot.
Nice video. Generally very practical. However, you ignore the topic of how to define the matrices Q and R. You say very broadly what they are. But, thru Part 4 of this video series, it is impossible to know how to calculate either matrix.
Error Alert at t=5:30.They accidentally swapped a minus for a plus when substituting 1/C for K, which if left as is would result in x_k = 2*x_k(pred)+ y_k.
I think at 5:43 the sign for Cx_hat- (the last term of the equation on the second line) was erroneously flipped to "+". It seems to cancel a + term later too so I'm think it should've stayed "-".
Wow! So much easier to understand Kalman Filters by listening to your video, than reading chapter 6 on State Space Models, in Time Series Analysis and its Applications by Shumway and Stoffer.
In the formulas appearing at 5:30 shouldn't it be "... - C\hat{x}_k^-)" and "... - C^{-1}C\hat{x}_k^-)" (the difference being the minus instead of plus symbol)?
At 1:39....I think in the state observer equation ....it should be y(k+1) instead of y(k) and also u(k+1) instead of u(k) and also Cx(k+1) instead of Cx(k). Correct me if I am wrong
it looks like an error in the sign of C*xhat_k predicted after the distribution of K_k. The K_k should be distributed in the multiplication without a change in the sign. So, it should read: xhat_k = xhat_k predicted + (K_k)( y_k - C*xhat_k predicted) xhat_k = xhat_k predicted + (K_k)y_k - (K_k)C*xhat_k predicted xhat_k = xhat_k predicted + (C^-1)y_k - (C^-1)C*xhat_k predicted *the (C^-1)C cancels and the (C^-1)y_k is effectively just y_k which leads to: xhat_k = xhat_k predicted + y_k - xhat_k predicted therefore: xhat_k = y_k **they mess up the signs a good bit. the previous part had corrections all over the place.
Thank you for the video. They are concise and helpful. Can someone let me know, how these animations within video are created? Could be helpful for broader teaching purposes.
There is a wee typo in the video, say, limiting the R approaching the none, and will cancel the prior state estimate. The sign should be plus rather than the minus.
Thank you for the nice video. At 5:32 how do summing the two x hat_k are cancel each other? They are on the same side of the equation? And both of the x hat_k are "+". One of them should be "-" for the cancel each other?
Is the predicted state not generated from the IMU? And then the measurement is from the GPS? Or do you use the velocity to predict then the measurement is from both the GPS and IMU?
I wonder if Kalman Filters can be applied to estimating whether a youtube video will force you to watch an advertisement or not. I watched this whole series without having to skip or mute a single commercial.
Yes. The Kalman filter is a tool. With enough "massaging" you can make it solve very complex problems in novel ways. But the question often is, is it the best tool for solving such problem? RU-vid has 300 hours worth of videos uploaded to it 60 seconds. Analyzing such data can be mind boggling. Likely machine learning may be more useful.
One thing I am confused by is that it seems like between the equations for the predicted estimate covariance matrix (P-), the Kalman gain matrix (K), and the updated estimate covariance matrix (P) that the Kalman gain will have a pre-determined trajectory, which feels odd. Seems like it should be affected by the feedback error in some way.
I got confused by this: At 1:40, for the state observer, should x_hat (on the left of the equal sign) have a dot on top? That's what it is in the previous video, no?
Hello I have a question concerning the process noise: If I have a distance signal which can change at maximum 5 mm between two measuring points, can I use those 5 mm as process noise? Or did I understand it wrong?
Why is the State Observer allowed only the previous state estimate, the previous input, and the previous measurement to estimate the current state WHEREAS the KF is allowed the previous state estimate, CURRENT input, and CURRENT measurement?
@MATLAB I'm using an IMU which has an accelerometer as well as a gyroscope and I am double integrating the acceleration data to get distance. should I be using sensor fusion to do this? or can i achieve this using only accelerometer data with a kalman filter?
2:00 Previous time step + current input... does it make sense?? shouldn be rather - Current state = previous state + previous input... same is Next step = current state + current input... So this is really confusing ..
