Page 30 - Security Today, March 2021

sure the pixels across that line. Unfortu- nately, that type of analysis can’t take into account whether what is crossing the line is a person, a dog, a cat, or even just a spi- der on the camera lens. All it knows is that pixels are changing.
Advancements in analytics have allowed us to take into account much more than the movement of individual pixels. Much of today’s AI is based on implementing and comparing different models, such as a da- tabase of images. The better the data set, the more effective the AI. This has enabled a wide range of impressive improvements, such as the ability to identify a shadow cast at a certain time of day, rather than mistaking it for a new object in the area. This has helped dramatically reduce alarm fatigue, which remains a serious issue in the security space. Over time, these incremental improvements result in significant achieve- ments that provide practical solutions with great accuracy.
TRAINING AI TECHNOLOGY IS KEY
If an AI is only as good as its data set, then improving those data sets is of paramount importance. While you might expect critical industries like healthcare or security to have the most cutting-edge technology, the truth is that most analytics programs are used in non-critical situations first in order to refine them for more mission-critical uses.
Retail. If a shelf is stocked with the same products all the time, the AI can be trained to recognize those products by comparing against other images. When that product is depleted, it can alert the store that it needs to be restocked or reordered.
Manufacturing. Routinely deployed, a camera, a camera watching the assembly line at a cookie factory, for instance, might be trained to identify when a cookie failed to receive its cream filling. It is something important to that particular company, but it isn’t safety critical. This provides a clearly defined benefit to the user, but also provides long-term benefits by fine-tuning the technology. Those same technologies, now more refined, can then be applied to security and other critical applications.
Using AI in these environments helps build the necessary models, adjust to dif- ferent situations, and identify what is im- portant and what is not. This is also why many new AI solutions are essentially “next step” improvements on what al- ready exists. It is easier for a company to improve upon an existing solution if they have thousands of images (and years of experience) working on that same technol- ogy. Access to this type of training data
can help power incremental improvements to the technology, underscoring why AI adoption requires a more thoughtful and measured approach.
To put it another way, think of how long it takes a doctor to learn to identify signs of one specific disease. An ophthalmologist might look at tens of thousands of images of eyes showcasing the exact indicators to look for. AI, on the other hand, is trained on data sets that aren’t nearly as specialized. After all, developers can’t just go to a local retailer and collect random video that fits their needs. They have to rely on more abstract videos of cars, shadows and things crossing a line in space. While some have experimented with creating synthetic training data, there is significant risk in- volved: what if one element is wrong? One small mistake can ruin an entire training program. As a result, training AI in non- critical applications remains the most effective way to refine the technology.
AI’S IMPACT ON SECURITY
AND OPERATIONAL EFFICIENCY Once technology has been proven in non- critical areas, it can be deployed in more critical applications. As previously men- tioned, one of the biggest challenges in the security space is alarm fatigue, or the issue of too many false alarms. If sensors cannot tell the difference between a tres- passer entering a building and a shadow moving with the sun, security teams will be deluged with false alarms, potentially drowning out actual security incidents amid all the noise. Forcing security staff to sift through dozens of alerts is time con-
suming and inefficient, and increases the likelihood that a genuine threat will slip through the cracks.
Today’s analytics are not only better able to differentiate between true security incidents and false alarms, but can be pro- grammed to trigger when a given event oc- curs. That event might be a person in an area they aren’t supposed to be, too many people standing in line at a register or even someone not wearing a mask.
If detected, an alert can be issued to the appropriate staff and a designated response can be triggered, such as a light turning on, an audio message playing, or even a human interaction. From a security standpoint, this works as a highly effective deterrent: shine a light on a bad guy in an area they’re not supposed to be, and they feel busted. An audio alert that lets them know they’ve been seen will, more often than not, lead them to beat a hasty retreat, foiling whatever plans they may have had.
There is also a business intelligence aspect to artificial intelligence. Today’s ad- vanced analytics can combine and analyze data in new and innovative ways, leading to insights that can greatly increase opera- tional efficiency.
A weekly data report for a retail store might include data correlating sales num- bers to a recent marketing promotion, or an analysis on which direction customers are entering the store from, or which di- rection they are heading once they enter. It can even focus on how customers are dressed, identify where choke points are within the store, or track how many cus- tomers enter and exit the store without
30
MARCH 2021 | SECURITY TODAY
ARTIFICIAL INTELLIGENCE
Phonlamai Photo/Shutterstock.com