Thursday, October 2, 2008

GHC08: Fighting Crime using Gunshot Location Systems

This is a very interesting talk on the ShotSpotter technology by Elecia White and Sarah Newman. This technology has been installed in several major cities, helping to solve crimes when the shooter can be pinpointed quickly.  In one example, a sniper shot someone from a roof, and actually stayed on the roof, relaxing and smoking a cigarette - thinking he was out of the expected shooting area.  But, the ShotSpotter technology had pinpointed him and the police were able to make an arrest.


Of course, this technology needs to attempt to differentiate between firecrackers, hammers, backfiring cars and gunshots. The technology takes a first pass at guessing what it was hearing (and gives a level of confidence), but then asks the police dispatcher to make a judgment call on whether police action is required or not.


They find this gets faster and better reporting than actually relying on people calling 911 (as there is a longer delay before they call and only about 50% of gun crimes are called in).  The system isn't perfect, but seems that it can definitely help!


Valerie Fenwick

2 comments:

  1. Qualify: Safety Dynamics sells gunshot location systems based on temporal pattern recognition of gunshot.
    This technology definitely needs to attempt to differentiate target sounds from similar noise. False positives, especially in urban environments, are an enormous challenge during installation and testing with clients watching. The York PA Daily published a 6/29/2008 report on ShotSpotter performance over one weekend which showed the system had a 86% rate of false positives. Police confirmed of the over 657 events detected, less than 100 were actual gunshots.
    When Safety Dynamics participated in Operation Disruption in Chicago, there were a number of sounds in that environment that could easily be mistaken for gunshot. Since the SENTRI gunshot location system is based on mathematical models of human brain function, it can be trained to listen for target sounds and ignore non-target sounds. Downtown Chicago was a real challenge, but also the perfect place to obtain recordings of gunshot-like sounds for input to a neural network with negative feedback, so the model learns the difference between a bus backfire and a gunshot and can implement that discriminatory function in the field. This is a better solution to lowering error rates in gunshot location system performance.

    ReplyDelete
  2. Hi. James from ShotSpotter here. The good Dr. Berger (the founding scientist of Safety Dynamics, by the way), isn't quite providing the full story. Our systems *do* detect far more fireworks than gunfire, but SHOTSPOTTER SYSTEMS AUTOMATICALLY DIFFERENTIATE FIREWORKS FROM GUNFIRE. Thus saying that the system has an 86% false positive rate is misleading: the system identifies both gunfire and fireworks *and tells the police which are which*. (On that June weekend--right before July 4th, as you will notice--there simply were many more fireworks set off than gunfire in the City. Detecting each one and telling police it is a firework does not constitute a false positive; the fireworks are a fact of the community. We don't blindly report all of them; we differentiate them!)
    Are we always correct? No, not always. There is an error rate, but it's substantially less than 20% of detected incidents, and often less than 10%. (Right now, York Commissioner Whitman says we are detecting 94% of the known gunfire in the City, and additionally that 2/3 of the incidents we detect aren't reported to 911.) But the chance that there's a machine classification error is why we also return the audio of that actual incident so that dispatchers can review any questionable incidents and provide additional intelligence. Over time, we re-train the algorithms which classify gunfire and fireworks on a city-by-city basis, using the dispatcher's feedback (and our own expertise) to improve ongoing performance. Also, this audio is used for forensic analysis. Just last week, for example, our audio data was played to a jury in a criminal trial of two defendants, one of whom claimed he didn't have a gun. The jury heard the two guns for themselves, and we were able to plot each location and prove how far apart the two defendants were standing (thus proving it wasn't one defendant holding two guns).
    (Why detect fireworks in the first place? Because people call 911 about them all the time, usually thinking they are gunfire, and by providing information about such incidents to police dispatchers, they can *avoid* diverting precious resources to deal with fireworks. Just last night, for example, I watched five fireworks incidents come in to an East Coast system, all of which were correctly identified as fireworks by ShotSpotter automatically, and any one of which could otherwise have been a waste of officer's time had someone called them in to 911 as gunfire.)
    The article Dr. Berger is referring to actually provides a nicely balanced picture, and I recommend it for those who want the facts: http://www.inyork.com/ci_9822975?IADID=Search-www.inyork.com-www.inyork.com

    ReplyDelete