EKF Localization (Nived Chebrolu)

Рет қаралды 12,370

Күн бұрын

Пікірлер: 21

@sdpayot 10 ай бұрын

Thanks for sharing all this great content, Prof @CyrillStachniss and Nived Chebrolu. What a pleasure to follow along your courses! I'd be curious to hear more about how some of the "hyper-parameters" for those models tend to be estimated in practice: how would practitioners estimate Rt, Qt, or alpha_i in the example presented in this session for instance? also, are there some established python libraries practitioners tend to rely on to apply EKF (or some of the other algorithms introduced in this course) to their problems? thx again!

@sylvaingeiser1317 3 жыл бұрын

Thank you for this very clear lecture :) I'm using dead reckoning as motion model and I'm wondering how to add the yaw noise described in the video about motion models in the equations of EKF. Since the control command consists of only 2 variables for dead reckoning, noise can only be added to these 2 variables through M matrix. But as you pointed in motion models video, something seems to be lacking because the robot's state is defined by 3 variables. This is the reason of the introduction of a yaw noise. According to my comprehension of EKF, I think that this yaw noise can be taken into account by splitting the R matrix into a sum of 2 matrices, the first being VMV as you described in the video and the second being a 3x3 matrix with only a non-zero value at the bottom-right which corresponds to the yaw noise. Could you please tell me if it works or if it should be done another way ? Thank you in advance.

@modyskyline 3 жыл бұрын

many thanks for dr.Cyrill and dr.navid I have just one question how to get the Mt matrix??

@mych5713 2 жыл бұрын

same question here.

@iskalasrinivas5640 3 ай бұрын

Wonderful Thanks for the lecture

@hobby_coding 3 жыл бұрын

loved this thank you

@surajsapkal1293 4 жыл бұрын

Very good presentation. Thank you so much.

@S-Innovation 3 жыл бұрын

Thank you. This is a very detailed lecture. Love it.

@partha95123 4 жыл бұрын

Crisp and precise explanation.

@BrunoSantos-ov1sw 3 жыл бұрын

Maybe is not relevant but there is a superscript i missing on equation 16 for z_t (our predicted measurement). Referring to the correction step.

@notmyproudest 2 жыл бұрын

Is the video at the end the expected output for the ekf localization exercise given in the course website? or is there a difference in the Q and R matrix values provided in the exercise?

@gustavovelascoh 3 жыл бұрын

Thanks Nived

@apppurchaser2268 2 жыл бұрын

Great

@rahul122112 3 жыл бұрын

Can someone please help with this: As I understand, for the odometry motion model, the controls (deltaTrans, detaRot1, and deltaRot2) are not readily available from the source of odometry. They need to be derived based on the pose at time t-1 and the current pose at time t. That is, we use the (x,y,theta) information at time t and t-1 to get the (detlaTrans, deltaRot1, deltaRot2) control inputs. If this is correct, then when we set up the motion model equations, aren't we just reversing the above derivation? i.e. just using (deltaTrans, deltaRot1, deltaRot2) to get the (x,y,theta) at time t? The only difference I see is that in the motion model, we end up adding gaussian noise to the prediction. Is my understanding correct? What is the point/advantage of doing so?

@CyrillStachniss 3 жыл бұрын

Typically, your odometry (encoders counting the wheel ticks) will generate an INTERNAL (x,y,theta) coordinate. As this is not aligned with your frame, you compute the (r1,t,r2) representation between current and last pose in that internal frame and then concatenate it to YOUR frame.

@rahul122112 3 жыл бұрын

@@CyrillStachniss Ah, thanks for the explanation! The way I have understood is: The odometry would/could be generating the pose (x,y,theta) in the internal frame (or lets say its odom frame). This frame may or may not be aligned with our local reference frame. Therefore, the control inputs (deltaTrans, deltaRot1, deltaRot2) are generated which are independent of the frame of reference and can be concatenated or used as inputs in the local reference frame.