# GPS+IMU sensor fusion not based on Kalman Filters

Do you know any papers on or implementations of GPS + IMU sensor fusion for localization that are not based on an EKF (Extended Kalman Filter) or UKF (Unscented Kalman Filter)?

1. I've found KFs difficult to implement
2. I want something simpler (less computationally expensive)
3. rlabbe's book on Kalman Filters suggests they aren't ideal for this use case. In this chapter it says:

Kalman filters for inertial systems are very difficult, but fusing data from two or more sensors providing measurements of the same state variable (such as position) is quite easy.

Further in the "Can you Filter GPS outputs?" part

Hence, the signal is not white, it is not time independent, and if you pass that data into a Kalman filter you have violated the mathematical requirements of the filter. So, the answer is no, you cannot get better estimates by running a KF on the output of a commercial GPS. [...] This is a difficult problem that hobbyists face when trying to integrate GPS, IMU's and other off the shelf sensors.

If the above is true, other approaches should be out there. Yet answers to gps-imu fusion and implementation questions only point to EKFs. So far, I've only found a Master's thesis implementing an EIF (Extended Information Filter) and a paper with an Adaptive filter I don't quite understand. Do you know of alternative approaches?

• Could you elaborate on your application? Why do you need fused data and what platform are you using it for? Jul 27 '21 at 14:06
• The kalman filter works really well (and that it why it has stood the test of time). There are several libraries that have implemented it and you should be using one of those. That being said, it is just a bayesian filter and there are alternatives like the particle filter. Jul 27 '21 at 14:09
• Overall the short answer is no. rlabbe's book said it is difficult, not "not ideal". Almost all real-world applications use a Kalman filter. The papers you linked to EIF and the adaptive filter are just Kalman filter variants. The particle filter is one the exception I can think of, but that is even more complicated to implement, and more computationally expensive. Jul 27 '21 at 17:08

So the result will be a distribution shape that could estimate the expected point with some confidence '$$\alpha$$' (let's say 95% zone, or 99% confidence in this zone)