There are plenty of systems out there to identify the direction of a shooter after he has opened fire. But it might be too late by then. That's why Darpa is developing a system to spot a rocket propelled grenade* before *it's fired. It's a major challenge, but the solution may lie with a swarm of software agents.
The Rocket Propelled Grenade Pre-launch Detection and Cueing Program aims to deliver "an omni directional, visual, and vehicle-mounted surveillance system for threat detection using cognitive swarm-recognition technology to rapidly detect and identify the locations of attackers with RPGs before they are launched." Fitting a set of video cameras giving 360-degree coverage is easy enough, but the hard part is the software.
Machines are notoriously bad at identifying things. Recognizing a chair or an apple is one of those everyday skills that humans take for granted, but it is incredibly difficult to replicate. (Darpa's Grand Challenge driverless car contests were, in many ways, just fancy ways of getting machines to see the world more clearly.) Objects do not come in standard shapes and sizes, and they may be partly concealed or at an unusual angle. So Darpa is simplifying the approach by just looking for one fairly standard object, an RPG launcher. This is a cheap, widely available weapon used by insurgents everywhere in Afghanistan, Iraq and many other places. There are of course many versions, but the Russian RPG-7 and its many clones are the most common, and the ones to look for.
Even looking for one specific object takes a vast amount of processing power, however. The normal technique is to have an analysis window scanning across the image, looking for a match. This is not fast enough –- it's no use having a system telling you that it spotted an RPG 30 seconds ago. Even if you have a large number of search windows scanning at the same time, they're too slow.
The search can be sped up using a technique called Particle Swarm Optimization.
First developed in 1995, it's based on the flocking behavior of birds and insects. Instead of having the search windows scanning in fixed paths or at random, they react to each other and work together like swarming insects.
Imagine the search is being carried out by a large number of software agents -– like the hordes of Agent Smithsin the *Matrix *series -- who all start out looking in different directions and reporting what they see:
Smith #1 : Just empty road here
Smith #2 : There's some trees ... might be something, I need to look more
Smith #3 : Nothing here at all
Smith#4: Just a bare field here...
As they know that there is a more promising area to search, Smiths
#1, #3 and #4 now start scanning the same general area as Smith #2. By exchanging information rather than continuing to search in less promising areas, the Smiths can focus their efforts where they are most likely to be fruitful and complete the search more quickly. If there's nothing to find, they will continue scanning the entire area including the less likely areas. But if there is something there, they're likely to find it much more quickly.
HRL Laboratories, a company that works with Darpa in the field of distributed computing, has shown that in a sample application of spotting a pedestrian in a picture, the swarm recognition technique finds the pedestrian 70 times faster.
The other advantage of the swarm approach is in reducing false alarms. Because most of the system's attention is rapidly focused on objects of potential interest, there is much less chance of a false positive getting through.
"This combination of accuracy and speed is superior to any published results known to us," HRL notes in an article about the technique.
Darpa's
RPG-spotter is intended to have an accuracy of 95 percent and be able to cope with up to five simultaneous threats, with a minimum of false alarms.
The project has a budget of $3 million for this fiscal year, which will be spent on developing and maturing the detection and classification algorithms.
If it works, it could be a real life saver, keeping a 24/7 lookout for threats in all directions. And spotting RPGs may only be the start.
If the software pans out, later versions can look for other types of threat as well -– both hardware and human. As the algorithms improve and processors get faster, picking out a known terrorist face from a crowd in an instant might become a real possibility.
[Photo: Warner Bros.]
