Waiting for the Gun
USC / Eric Mankin | November 30, 2004
A USC engineer uses his expertise with nerve cells to create a surveillance system that can recognize the sound of a nearby gunshot - and identify the shooter. In a unique pilot program, L.A. and Chicago will deploy test units in high-crime areas.
"You start with a relatively small number of sounds you have to distinguish with high accuracy - gunshots, for example; or diesel engines for border patrol crossings; or chainsaws to listen for outlaw loggers. This vocabulary is quite manageable," said Berger.
A USC biomedical engineer's pioneering brain cell research has led directly to a patented system that is now being rolled out to stem gun violence on the streets of Chicago and Los Angeles.
The engineer is Theodore Berger, director of the USC Center for Neural Engineering, whose life's work has deciphered the way in which nerve cells code messages to each other.
Berger is also a key researcher in the USC Viterbi School of Engineering's Biomimetic MicroElectronic Systems Engineering Research Center.
A microphone surveillance system now is using his insights to recognize - instantly, and with high accuracy - the sound of a gunshot within a two-block radius.
The system can then locate, precisely, where the shot was fired, turn a camera to center the shooter in the camera viewfinder and make a 911 call to a central police station.
The police can then take control of the camera to track the shooter and dispatch officers to the scene.
The city of Chicago is installing the first five of a planned 80 devices in high-crime neighborhoods, supplementing existing cameras. In Los Angeles County, Sheriff Lee Baca is now soliciting community involvement and participation to deploy 10 of the units in a pilot test, to be followed by more if the results are successful.
Algorithms devised by Berger, who holds the David Packard Chair in the USC Viterbi School of Engineering's department of biomedical engineering, are at the heart of the SENTRI system built by an Oak Brook, Ill.-based firm named Safety Dynamics, a company in which Berger serves as chief scientist.
SENTRI uses acoustic recognizers, posted in trios or larger groupings on utility poles or other listening posts, which are tuned to certain specific warning sounds with extremely high accuracy.
"A simple loud noise, even an explosive noise, won't set them off," Berger said.
The device is listening for the entire sound pattern of the gunshot, not just the initial explosion, which makes it much less likely to mistake other loud noises for shooting.
A specially configured computer system (a "directional analyzer") accurately calculates any authenticated gunshot's location - using the difference in the time the sound arrives at the different microphones on a SENTRI acoustic unit.
Field tests with real weapons have shown 95 percent accuracy with respect to gunshot recognition, and 100 percent accuracy with respect to centering an attached camera on the shooter for those recognized gunshots.
SENTRI is an acronym for "Smart Sensor Enabled Neural Threat Recognition and Identification." The "neural" in the title refers directly to Berger's work, which was based on analysis of the "language" nerve cells, or neurons, use to convey information, and specifically on his modeling of the way the brain forms memories of sounds.
The neurons' only way of distinguishing signals is to fire repeatedly, either faster or slower, in different temporal patterns.
"It is the time difference between pulses that carries the information," Berger said. "This is a coding completely unlike that used by computers, which are collections of ones and zeros, changing to the beat of a constant clock."
Working with computer specialists, however, Berger has created neural-like computer systems that can model the neural time coding and make distinctions the way nerves do.
Four years ago, he and a colleague used the technique to demonstrate the first speech recognition system that could pick words out of ambient noise as well as humans can.
While work continues on speech-recognition applications, the systems need training to learn individual signals. For language, this is very time consuming because the system has to learn each individual word.
"But for alarm signals," Berger said, "you start with a relatively small number of sounds you have to distinguish with high accuracy - gunshots, for example, or diesel engines for border patrol crossings or oil pipeline thieves, or chainsaws to listen for outlaw loggers. This vocabulary is quite manageable."
Machine sounds are the only ones in SENTRI's vocabulary. It cannot eavesdrop on conversations, the scientist emphasized.
Berger's work with neural systems grows directly out of 30 years of research attempting to create a silicon system that can be transplanted into a living brain or other nervous tissue to restore function lost to disease or injury.
The current line of research that led to the gunshot recognition is being expanded in collaboration with computer scientist John Granacki at the USC Viterbi School's Information Sciences Institute.