Fujitsu Laboratories Ltd. and Fujitsu Research and Development Center Co., Ltd. have innovated an AI technology for video-based behavioral analysis. Dubbed "Actlyzer", the tech can recognize a variety of subtle and complex human activities without relying on large amounts of training data.
Deep learning technologies conventionally demand large amounts of video data for training systems to recognize individual behaviors, and video data must be collected from scratch in order to add each new behavior. This time-consuming process means that it can often take several months to introduce functional AI into the field.
Taking advantage of the fact that human behaviors generally consist of a combination of basic movements and actions, (e.g. walking, nodding, extending the hand) Fujitsu has created a technology that makes it possible to recognize more complex human behaviors, including when an individual appears to be acting suspiciously or seems inclined to make a purchase, by training the AI to recognize about 100 basic actions in advance.
Potential use cases for this new technology include automatic detection of individuals engaged in suspicious-seeming activity, product interest surveys based on recognition of purchase behavior by customers, and training applications through comparing the skills of experienced and novice workers at factories, to name a few. Actlyzer demonstrates Fujitsu's continued commitment to delivering exciting new AI technologies with an emphasis on their human centric potential, offering rapid deployment to business frontlines to contribute to enhanced security, streamlined operations, as well as better workplaces for employees.
Moving forward, Fujitsu will initially offer Actlyzer to customers in Japan during fiscal 2019, and subsequently aims to commercialize the technology for international markets as a service as part of its "Fujitsu Human Centric AI Zinrai" portfolio.
Development Background and Issues
Recent advances in AI and deep learning techniques have made image-based behavioral recognition technologies a reality. Still, acquiring sufficient image or video data to train AI systems-especially for the recognition of behaviors consisting of multiple, discrete actions or movements-has presented a significant hurdle to the real-world use of technologies that can recognize complex or subtle behaviors.
About the Newly Developed Technology
To address this challenge, Fujitsu Laboratories and FRDC have developed their Actlyzer behavioral analysis technology, which can recognize complex behaviors composed of multiple actions or movements without the need for extensive video learning. The primary features of this technology are as follows.
- Accurately recognize basic movements and actions that make up complex behaviors
Actlyzer can uniquely define about 100 basic actions that constitute complex behaviors and can recognize all these basic behaviors through deep learning. Approximately 100 types of basic actions can be recognized with an average accuracy of 90% or more through training the system using a large amount of video data in advance. In addition to walking and running, this system can accurately recognize more subtle basic movements such as turning the head to the right, turning the head to the left, tilting the face upward and tilting the face downward.
- Easily recognize complex human behaviors from combinations of basic actions and movements
The technology can recognize complex behaviors by specifying combinations of basic actions, order, place of occurrence, and target of action. Actlyzer recognizes various actions with simple settings, and the recognition accuracy can be adjusted immediately by changing parameters.
For example, it is possible to identify potentially suspicious behavior by specifying combinations of basic actions or movements, such as sitting in front of a door, looking at a keyhole, or reaching for a keyhole. In addition, recognition accuracy can be further refined by specifying additional conditions such as turning the head left or right to look around, and by specifying the duration of each action.
Using 21 kinds of video data taken indoors and outdoors, Fujitsu conducted an experiment to identify 8 kinds of suspicious behavior (checking out a house, wielding a weapon, etc.). Because it took only one day to create rules for the combination of the basic actions to detect these eight types of suspicious behavior, the customer can determine the applicability of this technology in the field with just a one-day verification experiment.
Aside from its security applications, Actlyzer offers the potential to streamline operations by recognizing human behavior from video data in a variety of industries. This includes the analysis of purchase behavior of customers at stores, checking the response behavior of clerks, measuring working hours at manufacturing sites, and confirming work procedures.
Fujitsu will initially offer Actlyzer to customers in Japan during fiscal 2019. At the global level, Fujitsu aims to commercialize the technology as a service as part of its "Fujitsu Human Centric AI Zinrai" portfolio.