I read an ieee journal article over a decade ago claiming success with helmet mounted sensors. The system used the time of the sonic crack and the muzzle report as observed from each helmet. Its like GPS, if you have 5 or more observers you can solve for the time reference as well. Actually, it might be less since they had multiple observations from each helmet. It supposedly could then highlight the location of the shot using AR. It then went on to say they could tell what the firearm was from the muzzle report. Can't find the article now. Doesn't seem that outlandish, but as you say, accuracy would be the real question. The speed of sound can vary so much based on weather.
> It then went on to say they could tell what the firearm was from the muzzle report.
> over 95% caliber estimation accuracy for all shots, and close to 100% weapon estimation accuracy for 4 out of 6 guns tested.
That's pretty cool. Gotta read the whole thing now.
e: kinda relevant to the other subthread about time sync:
> Correlating ToA measurements requires a common time base and precise time synchronization in the sensor network. The Routing Integrated Time Synchronization (RITS) [15] protocol relies on very accurate MAC-layer time-stamping to embed the cumulative delay that a data message accrued since the time of the detection in the message itself. That is, at every node it measures the time the message spent there and adds this to the number in the time delay slot of the message, right before it leaves the current node. Every receiving node can subtract the delay from its current time to obtain the detection time in its local time reference. The service provides very accurate time conversion (few μs per hop error), which is more than adequate for this application. Note, that the motes also need to convert the sensorboard time stamps to mote time as it is described earlier.
Yeah I'm pretty sure that was it. They were getting really good classification results with simple statistical methods. Today I suspect someone would have thrown a hybrid CNN-LSTM at it and gotten similar results but no explainability and 100x the processor requirements. "Try the simplest thing that might possibly work".
A few μs for all that work seems high when gps could get you to ns, or adding extra observers could remove the need for prior synchronization. I guess it is a more robust answer at least, in case GPS isn't there and some of the observers got shot already.
