Imagine the scene: A man approaches the front door of a home, crouches and examines the keyhole. Is this a resident or guest who has lost their house key? Or a burglar looking for an entry point? ‘Actlyzer’ is a new AI-powered security system that aims to tell the difference in such scenarios. The Fujitsu Laboratories and R&D Center behavioural analysis technology identifies suspicious activity by analysing complex combinations of human actions and movements — and does so with minimal training data.
Teaching deep learning technologies to recognize individual behaviours typically requires a large amount of video data for training systems. Moreover, the video data needs to be collected from scratch, so each new behaviour can be isolated and identified. It can take months to train a functional AI in this way. The Fujitsu researchers however found a shortcut by identifying basic behavior-signaling building blocks. Just as simple Lego bricks can together build complex entities, so can basic body movements and actions combine to represent complex human behaviors.
Fujitsu trained Actlyzer to recognize about 100 basic actions such as walking, nodding the head, extending the hand, etc. By stitching these together, the model developed the ability to recognize more complex human behaviors.
In the first step researchers trained the system to an average accuracy of 90 percent on identifying the 100 basic actions. The “more data, better accuracy” rule applies here, as the basic actions data used in training includes not only many variations on common movements, but also a number of specifically-selected subtle and sneaky actions, such as turning the head quickly to either side to check for witnesses.
By specifying combinations of basic actions and their order, place of occurrence, and target, Actlyzer learned to recognize different behaviors with only simple settings. Fujitsu researchers told Synced that recognition engines trained using conventional techniques cannot be so easily customized, making it necessary to train such systems on a large number of images. “The technology we have developed solves these problems.”
Its ability to identify suspicious behavior and convenient customisability have enabled Actlyzer to tap markets and industries for security applications and more. Fujitsu notes the tech works in real-time video analysis and could also be used for example for examining purchase behaviours of customers and response behaviours of clerks in stores, timing the working hours at manufacturing sites, or checking on work procedures in various environments.
There are increasing concerns these days regarding computer vision and other public surveillance technologies and their potential threats to privacy. Synced asked Fujitsu how it intends to safeguard this technology from malicious use, and the researchers responded: “Generally there are instances in which public opinion recognizes the need for crime investigation and crime prevention. If Actlyzer were being used within a company such as in a factory, however, it would be necessary to secure an agreement from the employees before deploying the technology in the workplace. For places like stores with ordinary people and customers, it should be clearly indicated that the technology is being used. Ultimately, specific measures on how to responsibly protect personal information and prevent misuse will be a matter of customer judgment.”
With Tokyo hosting the 2020 Summer Olympics this summer, the Japanese government has been cooperating with tech companies like Fujitsu on a vigorous AI development policy designed to maximize security while minimizing inconvenience and people flow problems at the games. A related tech is Japanese security company ALSOK’s emotional visualization system, which will monitor crowds via cameras mounted on squads of autonomous roaming robots. The AI-powered system measures psychological states based on body signals like the jitters, then assigns values accordingly and colour-codes the people in its environment on a display screen to indicate suspicious behaviours and possible threats.
Because deep learning is a compute-intensive technology, Fujitsu researchers suggest future work on Actlyzer should focus on how to make it faster and lighter.
Journalist: Fangyu Cai | Editor: Michael Sarazen