Page 24 - Security Today, JulyAugust 2023

“ ... modern AI-based video cameras are enabling operations and security departments to shift their stance from being purely reactive to past events, to proactively addressing situations in real time and preventing them from escalating.” combing through hours of footage. If desired, analytics can also trigger alarms or notifications that can be sent to a security team. To sum it up, in its most basic form, machine and deep learn- ing algorithms (aka AI) detect and describe objects, while analyt- ics analyze what the objects are doing and report on them. DEEP LEARNING AND EXTRACTING DATA FROM VISUAL IMAGERY When we talk about AI in the physical security industry, we are usu- ally talking about a subset of the larger AI field. Machine learning and deep learning are both types of artificial intelligence that allow computers to learn without being explicitly programmed. Machine learning is a broad term that encompasses any type of AI that learns from data, while deep learning is a subset of machine learning that uses artificial neural networks to learn from data. Artificial neural networks are inspired by the human brain, and they can be used to accomplish a wide variety of tasks, including image recognition, natural language processing (à la ChatGPT) and speech recognition. One of the key advantages of deep learning is that it can learn from, and process through, large amounts of data. Video, com- pared to something like a spreadsheet of numbers, represents a lot of data. In the field of AI, video is often referred to as un- structured data (images, audio) versus structured data (numeri- cal, categorical) that is more organized and easier to analyze. Kryuchka Yaroslav/Shutterstock.com Although unstructured data is much harder for deep learn- ing algorithms to process, the algorithms can do amazingly well at image recognition when presented with enough training data. This is because artificial neural networks can learn complex pat- terns in data that would be difficult for traditional machine learn- ing algorithms to learn. It is for this reason that deep learning is particularly adept at identifying objects in images. Today, surveillance video has be- come a major contributor of unstructured “big” data. By using deep learning, we can make sense of the mountains of video that are being captured and have it help us find the proverbial needle in the haystack. DOING MORE WITH LESS Deep learning has enabled security teams all over the world to be able to do a better job with the limited resources they have. Em- ploying a surveillance camera with object recognition means that new and unique forensic searches can be done in a tiny fraction of the time it would take a team of people to do manually. Likewise, because searches can be saved, we can be on the lookout for known vehicle types that are suspected of crimes in the area, for example. When that red van loiters outside the jew- elry store at 4am, we want security teams to get an alert. If there is a person wandering around the loading docks after hours, se- curity staff want to know about it. Deep learning is what enables security teams to evolve to a more proactive stance to potential threats, versus only reacting to past events. Having a traditional “pixel-based” motion detection camera in the above scenarios would trigger an alert every time a shadow goes by, or a car headlight sweeps across the cameras field of view causing non-stop false alarms. Deep learning and object detec- tion makes motion-based analytics truly useful. GOING BEYOND SECURITY While the advantages of deep learning to the physical security 24 JULY/AUGUST 2023 | SECURITY TODAY DEEP LEARNING