Low-latency real-time IMU sensor fusion for precise indoor positioning systems

If you can take the IMU sensor fusion post-processing – take it. It will always be more accurate and more reliable than the real-time IMU sensor fusion:

• More data can be taken into account with post-processing because more data is available – you accumulate it for long and then use it – higher confidence and robustness to jumps due occlusions or noise or other interference
• More data to be taken into account – higher accuracy and higher confidence and higher robustness

With the real-time IMU sensor fusion, you simply don’t have the privilege of more data. To accumulate more data, you need more time. More time = higher latency. Higher latency – IMU sensor fusion post-processing. Works brilliantly, but later.

With the real-time low-latency IMU sensor fusion, you have:

• Historical ultrasound-based location data. Typical update rate 4-20 Hz, i.e., an average latency from the latest ultrasound location update = 1/(location update) = 1/20.. 1/4 sec = 50 … 250 ms, which is OK for the majority of real life industrial applications. However, the if you use only the historic data, you don’t need the IMU – you can rely much easier and often more reliable extrapolation approach, which beautifully works for some types of the mobile objects, but only some 
• You have the latest ultrasound location update. Which is great, but you can never trust any data with 100% of confidence. But you don’t have enough information – yet! – to verify whether that the sudden change in the ultrasound-measured location was a real change because our mobile object was kicked by something strong and heavy or was it just a jump due to noise or occlusion. But you can’t be sure, because you don’t have additional data to verify it. When do you have it, it is too late to use it, since you already have new data and the old one is just outdated – you must rush ahead all the time. So, yes, we use it – indirectly – to correct rules, assumptions, weights, but it cannot be used to verify the data now, because we can’t have that verification data now
• Raw IMU data. Which helps a lot: if you think there was sudden change in the location, it means a sudden change of speed as well, which means high acceleration, that the IMU can easily detect that, in fact, there was no jump. Which solves some of the problems …  – OK, it was phantom/jump/occlusion or something. IMU shows, for example, no drastic acceleration that would explain the jump in the location. So, we can rule out that that measured ultrasound-based location is correct. It is not. But what is the correct one? – that is a tough question. That is where the difficulty starts. What shall we do with the raw IMU data to calculate the position based on the IMU? … – trivial double-integration of the acceleration? – so easy?! – not really: perfectly perpendicular IMU chip lying on the surface, will show very notable acceleration in X, Y – read the raw data and measure, because they are not perfect. Calibration? – helps and recommended. It is very useful. But, taking into account that the double integration is time in power of 2, the win with better calibration is short-living literally. Instead of 1/4 of sec, you can be within +-2cm accuracy 1, 2, 5 seconds! But that is it. After that time, your uncancelled integration error will start prevailing and you need some non-drifting source of data to cancel the drift: optical positioning, UWB, odometry, Marvelmind ultrasound positioning system, or, better a combination of all. That is the real sensor fusion. And that is what we do with a twist of self-checking and self-verifying algorithms that are driven by a single mantra: none of the standalone sources can be trusted. But when efficiently fused together, they are a very trustful, very reliable, and very accurate source of indoor positioning data