1
$\begingroup$

I am using a camera and a Deep Neural Network to predict one angle. This network received as input the frame, and calculates the mean and the variance associated to the prediction (which is basically its uncertainty), so that the output of the systen consists of a normal distribution.

I want to use this extra information (the uncertainty of the prediction) to better aproxmate the angle I want to predict. Since I have no experience with this, I don't know which filters could be better:

I have implemented a Kalman Filter that works really well with linear environments, as expected. However, if the angle starts to change non-linearly, which filter could use the information I have obtained (mean and variance for each sample) to improve the accuracy of the system?

$\endgroup$
  • $\begingroup$ Have you looked at the extended Kalman filter? Otherwise maybe a particle filter could be an option. $\endgroup$ – fibonatic Jul 19 at 23:28
  • $\begingroup$ I have heard about these non linear filters, but all the examples I found on internet usually implement the dynamics of he system or some extra information that it is not possible in my case. Maybe I am not understanding them well. Would it be possible to implement them (EKF or PF) by only having the mean and the variance of each output? $\endgroup$ – Manu Jul 19 at 23:44
  • $\begingroup$ So you only know that that the angle changes, but you have no idea how (so maybe nonlinear)? $\endgroup$ – fibonatic Jul 20 at 1:29
  • $\begingroup$ If there is any information remaining in the output of the neural network that can be used to improve that output, then the neural network needs more training. Is it possible to talk a little bit more about the problem you are trying to solve? $\endgroup$ – A_A Jul 20 at 8:26
  • $\begingroup$ Well, the angle is related to how a manipulator interacts with a surface (for example, maintaing the actuator normal to the surface). Right now the entire project uses a closed loop algorithm to interact with the surface by controlling the angle predicted by the network. Therefore, how the angle changes will also depend on the controller response (PI controller). Maybe this one could provide extra information to filter the output signal of the network, although I don't know how I could implement the filter that fuses all this info. $\endgroup$ – Manu Jul 20 at 11:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.