Published 24 May 2021 by Aktuell Säkerhet. Reporter: Linda Kante
With AI technology, the Swedish tech company Irisity is developing smart cameras that can alert independently when dangers are identified. “It’s smart to let machines do what they’re good at,” says Marcus Bäcklund at Irisity.
A major problem with video surveillance today is the amount of information collected as well as the time it takes for operators to manually monitor screens. But with Irisity’s algorithm-based AI solution, the system acts itself.
“The system analyzes and filters video data in real time and sends alarms to the alarm center only when a detection is actually made. Thousands of hours of video can thus be filtered down to just a few minutes, which drastically increases efficiency and improves the work environment in the surveillance operations”, says Marcus Bäcklund, CEO at Irisity.
Using AI and computer vision, each image in the camera stream is analyzed with high accuracy and precision to detect unauthorized activity in defined zones. The system performs a pixel analysis of the image and an object classification to determine if a movement refers to a human, an animal, or an object. When unauthorized or suspicious activity is detected, such as intrusion, fire, fall, abandoned objects or violent behavior, the system alerts the security center directly, which can ensure that the correct action is taken.
In order to not violate integrity, masking is often applied. This is done either through complete anonymization, which is used in sensitive installations such as schools, or by masking out selected areas. By doing so only relevant parts of the scene with a camera surveillance permit are recorded or monitored.
“As a company, we work to create smart technology to protect people and assets – not to intrude on personal integrity. More specifically, Irisity does not develop or apply any facial recognition or other technique to identify individuals’ identities. We also do not work to categorize or systematize recorded video data based on personal characteristics such as appearance, ethnicity or clothing“, says Bäcklund.
How does the technology stand up against cyber threats and criminals who want to hack the system?
“We have very rigorous procedures around how we systematically teach and train the system, which can then itself recognize scenarios, people and behaviors. A key in this is to have sufficiently relevant and annotated training data, both situations that the system should alert on and situations that the system should not alert on. Of course, it is also important in the training of the system to specifically take into account situations that can deceive other camera-based systems, such as rain, snow, poor installations or low-resolution cameras for example.”
“Throughout the chain of events, humans will of course always be a key part of the operations. Both at the alarm center to make a nuanced assessment of the situation as a whole, and on-site to evict intruders or enforce physical security. But it’s smart to let machines do what they’re good at. We now have a new acceptance of camera-driven solutions and many agree that the camera is in many cases a superior sensor that, in addition to the alarm, provides context to the situation via a video sequence. Technically, the development of hardware is moving fast and we have completely different “muscles” to intelligently process video data and make decisions based on it. Going forward, we believe that more and more intelligent algorithms will be able to run directly in the camera or on specialized hardware near the camera, concludes Marcus Bäcklund.