Updating state based on multiple sentences about gps.js HOT 6 CLOSED

Hey,
I am glad you find it useful! It is an interesting use case, what receiver are you using? Or are you trying to merge 4 receivers at 10Hz?

The thing is, that the state object should serve as a small layer on top of each NMEA sentence to help in the average case, which makes the idea of multiline message passing a bit of an overkill to add.

Furthermore, the question is, if averaging is the natural thing to do. It does not take the deviation in each update step into account, and if you do, you naturally come to Kalman Filters instead of averaging. I was thinking a lot to add a Kalman Filter to clean the state object, but decided to add it only as an example to the map example since in the general case you can't assume a constant timing of the sentences (like in your case) and the more important part: The KF would be too simple, to just filter the noise instead of using a proper model.

All in all, I would suggest to ignore the state object for your use case, subscribe to the data event and create an own state object, maybe in a similar fashion like the library itself does.

Robert

from gps.js.

Kalman Filters are in a way an overkill, and it is also true that Kalman Filters are hard to grasp. I said it leads naturally to Kalman Filters, since an "average filter" that you would implement by averaging is one assumption about what to do with deviating data. Such a filter is good if you want the result converge to a value, like calibrating a scale. An alternative would be a low pass filter, where you say

pos = prev_pos * alpha + (1 - alpha) * new_pos

The Kalman Filter now can be seen in a similar way, that alpha, which is constant in low pass filters, depends on the variance. So yes, you can use Kalman Filters to filter against a model (and therefore fuse information from different input) or you treat it only in the basic form of filtering with a varable alpha. It is still a lot of tuning, and at such a high rate a low pass filter could be enough (with a manually tuned alpha between 0 and 1), but the best way would be a KF, as you operate in a dynamic environment.

from gps.js.

Hey,

Thanks for taking the time not only to comment on the idea but providing me with suggestions for how to handle my case. Unfortunately I don't have the specs of the receiver, I just receive a package with sentences over websocket so some processing has already been done.

I agree with your assessment that the addition of noise filtering and collection support is probably not needed in the library itself.
Thanks for the hints, I'll look into Kalman Filters more closely.

from gps.js.

Hmm, Kalman filters might be a bit overkill for my case as well, the implementation and examples I can find seem to deal with predicting the next location and correcting based on the prediction and the input (and some other stuff). What I really need is a good enough and fast enough calculation of the "correct" location based on the data in each package.

I'm still thinking averaging because with 10 sentences any reasonable outliers' effect will be mitigated. I'll probably still need a way to filter significant outliers.

from gps.js.

If you want an reference implementation of a KF, have a look at the maps example that is shipped with gps.js

from gps.js.

Thanks again for going out of your way to help me out! I guess I just have to bite the bullet and learn KF if I want accurate data.

from gps.js.