Analyzes even the emotions of the person captured on film and identifies suspicious persons! Behavior Recognition AI Accelerates the Future of Security Camera Society

By 2021, the number of security cameras installed in Japan will exceed 5 million. And the number is still increasing at a rate of about 5% per year. We often see the moment of a crime captured by cameras. Security camera footage has become indispensable for crime investigations.

Major streets and train stations in central Tokyo are now covered so well that it is difficult to act without being caught on camera. However, the reality is that very little can be done to monitor them with the human eye 24 hours a day, 365 days a year. Rather than detecting incidents in real time, they are more often used as evidence after an incident has occurred.

Recently, however, a new security system has emerged that can instantly detect abnormalities in images captured by cameras using AI, according to Yuka Suzuki, public relations manager of Ajira Corporation, which developed and provides the AI security system “Ajira” (all comments are her own).

AZIRA” uses behavior recognition AI to detect abnormal behavior of people on the screen in real time. The behavior recognition AI is a combination of AI that estimates the posture of a person on an image and AI that estimates behavior based on changes in posture over time. It may be easier to visualize if you imagine a parapara manga. Human posture is extracted from an image of a person’s skeleton based on the connection of joint points. Ajira” has developed a superior AI algorithm for posture estimation, so it can detect individual people and estimate their actions even in an environment where multiple people are gathered together.

If the behavior recognized by the AI corresponds to an abnormal behavior that has been learned in advance, such as violent behavior, an alert is immediately sounded and not only the monitor but also the security guards on site are notified by e-mail. Video footage of the detected abnormal behavior can also be replayed on the spot. Detection is not limited to abnormal behavior. The system is also useful in detecting nuisance behavior such as damage to property and break-ins, as well as detecting falls and recognizing white cane and wheelchair users.

But AI security systems can do more than just detect in real time. It can also predict possible future dangers.

This is “Discomfort Detection”. The system collects data on the behavior of people in the camera’s angle of view as normal behavior and performs autonomous learning. Behavior that deviates significantly from the normal behavior established by this learning can be detected as “discomfort.

The system is able to identify “people who are likely to do something” that is out of place in the field. However, since this system only uses video images to make judgments, it still falls short of the five human senses, which include the eyes, facial expressions, and voices of people. Also, since it is only a “sense of discomfort,” not everything can be directly linked to a major incident. However, it seems that there are some aspects of AI’s “eyes” that humans are still unable to match. It is said to be able to estimate the age and gender of a person from his or her movements.

The age and sex of a person can be estimated from their posture and the way they move, as well as their height. The spine is actually a very important part of a person’s body, and even if we don’t tell it what it does, it seems to be able to determine the line of a person’s back as a ‘backbone. It can also tell if a person is the same person from different cameras based on their posture and the way they walk. Even if they are dressed differently, the way they walk can be used to determine some of their characteristics. It is not 100% yet, but it can identify individuals with a level of accuracy close to that of fingerprints.

This means that no matter how much a person hides his or her face when doing something wrong, he or she will be recognized. Furthermore, this AI can detect things that are not visible to the naked eye.

From a person’s behavior, we can assume whether this person is happy, sad, or angry, based on posture, gait and trajectory, and speed. For example, if the person is a man in his 40s, we can assume that he is walking faster than average and that his stride is narrower than average. We are developing technology to estimate a person’s emotions based on the characteristics of his or her movements.

The ability of AI to detect not only gender, age, and other attributes, but also emotions, just by looking at a person’s behavior on a camera is astonishing. In addition, a tracking function that shares information not only from a single camera, but also from multiple cameras is also under development. In other words, it can search for a specific person as long as he or she is captured by one of the cameras, which may be quite useful in the search for lost or missing persons.

Many people are concerned that a society in which security cameras abound is an “invasion of privacy” and a “surveillance society. However, in the case of Kenya, which imported a system from China, the world’s leading country in terms of security cameras, in 2003, many people in the city of Kenya said, “The city has become safer,” and “I am not worried because the government is managing the system properly. How about Japan?