sfwa / ukf

Unscented Kalman Filter library for state and parameter estimation
http://au.tono.my/log/20130531-kalman-filter.html
MIT License
461 stars 172 forks source link

Is there some principled way of utilising confidence of detection? #60

Open LukeAI opened 2 years ago

LukeAI commented 2 years ago

I have a stream of sets of measurements, and an implementation of this library's UKF which filters out the noise in these measurements. The measurements come with confidence scores, indicating how likely the measurements are to be accurate. I would like to improve the UKF output by utilising this information would it be reasonable/efficient to do something like this before every step:

// if score is low, eg. 0.1 then measurement covariance will increase from baseline by a factor of 100
// if score is very high eg. close to 1 then measurement covariance will be approx. baseline 
my_ukf.measurement_covariance = measurement_covariance_ / score^2
my_ukf.step(time_step, measurement)