The term artificial intelligence or AI, for short, which essentially involves computers being able to think and make decisions like a human, is becoming much more commonplace in the security industry. Over the past several years, a handful of companies have introduced AI solutions to the market as a way to provide end-users with more intelligent video analysis capabilities. AI also enables robotic technologies that have been recently introduced to the security industry to navigate their surroundings and make decisions as to what types of events require an alarm to be activated.
On Wednesday, BrainChip, a relative newcomer to the security space, unveiled a new AI-powered software dubbed "BrainChip Studio," which they believe will transform the way law enforcement and intelligence agencies use video surveillance to detect and classify faces. Not only does the software enable authorities to search large quantities of recorded video faster and more efficiently than most traditional facial recognition solutions, it also doesn’t suffer some of the typical limitations.
For example, rather than needing high-definition video to make a positive identification, BrainChip Studio can use low resolution footage, requiring only a 24x24 pixel image, to detect and classify faces. The software can also use video recorded in poor lighting conditions.
With the continuing proliferation of surveillance cameras worldwide and the desire to hold onto video data for longer periods of time increasing, the forensic capability to identify and track persons of interest in footage is simply outside the realm of human ability and will require AI solutions to be able to analyze efficiently. According to market research firm IHS Markit, 127 million surveillance cameras and another 400,000 body cameras are expected to ship this year alone, furthering the need comprehensive video analysis solutions.
In a recent trial, BrainChip’s software was able to, in real-time, detect, extract and classify more than half a million facial images from three-and-a-half hours of video recorded by eight different cameras. In another trial, it was able to process 36 hours of recorded video in less than two hours and extracted over 150,000 facial images.
“We take in live or recorded video, detect the faces and then for each face we detect we create a spiking neural network model,” says Robert Beachler, BrainChip’s senior vice president of marketing and business development, explaining how the software works. “Then we take the next image, detect the faces and compare from the previous model. If it’s a match, we know that this is the same person in the field of view and we can track somebody through the field of view of an entire camera and then we basically group those images into tracks. In any given camera, you will have a person walk through it and they become a track and there may be 20 images of their face that we have captured and we do that for all faces in the field of view of the camera.
“We keep this cache of faces going so that we can track them as they go through the field of view,” He continues. “Once we have collected that information, we can compile the images into a database with these spiking models. We then take the suspect image of the person we’re trying to find, run that through our spiking neural network to create a model and do a comparison to the other images we have previously acquired. Once we get a hit on one face we think is a high probability of being the suspect, we look at all the images in the track of that person. All of that information is put into a ranking that the suspect you’re looking for is this person and we provide that as a classification of the most likely matches.”
How a Spiked Neural Network Works
Unlike most of the other AI solutions that have been launched in the industry that use “convolutional neural networks” to process data, BrainChip Studio leverages “spiked neural network” technology which, according to Beachler, is a way of computing that more closely resembles how the human brain actually works by mimicking the functions of neurons and synapses.
“A neuron consists of synapses which connect to other neurons and then the neuron itself is an integration function of all of the different inputs and decides then if it should fire and pass on information to the neurons downstream. This passing of the information is called spikes and it is really a bioelectrical process saying, ‘here’s a signal of a certain intensity that goes through a synapse,’” Beachler explains. “The way that networks get trained is by reinforcing or inhibiting the connections of the synapses and changing the threshold function at which time the neuron fires.
“Contrast that to the way computers work: Computers do math, whether it integers or floating point, but the brain doesn’t think that way,” he adds. “We’re very good at recognizing patterns, we’re not very good at doing math just because of the way our brains are setup. We use these spiking trains as threshold logic to determine the functionality that we program into our minds.”
Thus, when you take those same characteristics of the human mind and apply it to software in the form of a spiking neural network, a computer becomes better equipped at recognizing patterns in video images.
“With that functionality, we get some interesting characteristics of this type of spiking neural network,” Beachler explains. “One is it is very good at finding patterns in noisy environments. If you get a repeating pattern, it reinforces those connections and it gets amplified while the noise gets inhibited. Also, it only requires a single image, so unlike convolutional neural networks where you have large datasets of pre-labeled information, the human mind can see something once and remember it forever.”
The Benefits of Spiked vs. Convolutional Neural Networks
As opposed to convolutional neural networks that require days or sometimes even weeks on a GPU server farm to train on the datasets that the user wants them to recognize, a spiked neural network only has to be given a single image from which to learn and identify patterns. “We can instantaneously train in the field to find patterns,” Beachler adds.
Also, because spiked neural networks do not have to do the math-intensive computations that convolutional neural networks do, they require much less computational power.
Beachler is quick to point out that the company is not simply trying to create another type of biometric facial recognition solution but rather something that can be leveraged by video surveillance end-users in everyday scenarios.
“What we’re trying to do is come up with a product that can be used in the real world in existing, non-sterile video surveillance environments,” he says. “We train our network in the deployment environment. We extract all of the faces in real-time from the camera or recorded video. We use the existing infrastructure, so it’s not an add-on to the camera or a new DVR or anything like that.”
Not only does Breachler expect BrainChip to significantly reduce the time and manpower needed to search surveillance images for suspects and other persons of interest, but he also believes it will allow authorities to do things they’ve never done before.
“In one real world example, there was a suspect who (authorities) believed had committed a crime, they had some footage of the crime itself but they wanted to find out who this person had been talking to, if he had any accomplices and where else he might have been prior to the crime. We trained our network on a unique pattern on this person’s t-shirt, took in all of the subway video feeds and were able to identify three other locations where this person had been,” he says. “We can train (the software) to find anything and this is something that couldn’t have been done previously unless you wanted to throw a bunch of people at it and have them stare at video to find this unique t-shirt pattern.”
About the Author:
Joel Griffin is the Editor-in-Chief of SecurityInfoWatch.com and a veteran security journalist. You can reach him at [email protected].
