Behavior analysis AI is a technology that allows AI to analyze human behavior patterns over time based on skeletal structure and posture. For example, when a person falls, AI automatically detects that the person has fallen, or detects dangerous behavior that leads to a fall based on behavior patterns.
There is no limit to the number of actions that can be described as "behavior" in a single word. Asilla picks up security-related behaviors such as "abnormal behavior (falls, strokes, fights, and breakage)" and "suspicious behavior (staggering and uncomfortable behavior)" and is developing its own algorithm.
Asilla has been focusing on behavior recognition AI since 2017 and has been developing an AI security system since 2022. After the introduction of the system, we have been receiving an increasing number of requests from our customers, in addition to their initial needs, asking if the system can detect more detailed behavior, and recently we have been expanding the system to the retail domain in addition to the security domain.
In order to fulfill our customers' requests, we are currently conducting research and development focusing on even more detailed movements.
- What exactly do you mean by detailed actions?
Lets take shoplifting as an example. Shoplifting is the act of "actually taking something," but it is often preceded by a combination of various actions, such as going back and forth across the sales floor for an extended period of time or engaging in suspicious behavior such as scurrying around to check one's surroundings.
We believe that the range of detection will be further expanded if we can detect minute movements of people, such as a slight movement of the hand or scurrying, and assemble multiple actions, rather than obvious movements such as falling or entering.
微細な行動の検知を積み重ねて検知結果を導くためには、人間的視点が必要となります。前述の通り、例えひとつひとつの行動が検出できても、人間の行動を理解した上で複合的に行動を導���なければ、正しい検知を行うことはできません。
In order to detect such things as predictive behavior, it is first necessary to identify what kind of behavior precedes the target behavior, and then break it down into detailed behaviors that may occur. In addition, these behaviors need to be learned and defined by AI.
Let us illustrate this with a concrete example. For example, in the case of "shoplifting prediction," as described above, we need to identify what kind of behavior constitutes shoplifting in the first place. In this process, the AI will decipher specific examples such as "wandering around the sales floor more than necessary" or "paying more attention than necessary to the movements of store clerks" as signs of shoplifting behavior, and further define the behavior. It learns the minute movements of specific behaviors, and based on the complexities of these behaviors, it is able to "detect predictive signs of shoplifting.
In this way, when learning minute behaviors and detecting the signs of behavior that is the accumulation of such behaviors, there is always the question of "do people always behave in such a way? In this respect, a human science perspective is needed to define the behavior.
Learning -accurate AI requires "quality data" and "appropriate algorithms," and a human science perspective is also essential for both.
HSAR is conducting research and development with highly practical technologies for use in the real world in mind, and is building a proprietary behavioral database to maintain its high accuracy. The key point of this unique database is that it is constructed from a combination of "business perspective" and "human science perspective.
For example, training data in research is automatically generated by generative AI. AI cannot understand whether this data is "an inherently human behavior" or not. Therefore, it is necessary to make a correct judgment from the viewpoint of human science using psychology.
In the case of Asilla, the generative AI not only generates the original data, but also acquires actual human behavior data and uses it in a composite manner. In this way, the quality of the data itself is improved by reflecting actual human data, rather than processing AI alone, and the quality of the behavioral data generated by AI is also improved, thus promoting research from the bottom up.
AIの学習過程において非常に重要な役割を担うアルゴリズムの設計においても、HSARではヒューマンサイエンスの視点を取り入れています。
AIは万能ではないため、場合によっては人間の行動の違いを認識できず、正しい��ルゴリズムを作ることができない場合も考えられます。
たとえば「転倒」は数秒という短い時間に起こる行動に対し「ふらつき」は数秒に渡り起こる行動ですよね。人間の行動を熟知している人はこうした動きの特徴や違いを正しく理解することはできますが、AIは必ずしも正しく理解し判断するとは限りません。
また「右手と右肘は繋がっているため、規則性ある行動をする」など、人間の普遍的かつ身体的特徴を加味したアルゴリズム設計が行動認識AIでは非常に重要となります。
AIが正しく学習できてない場合は、ヒューマンサイエンスの視点からAIに誤りを指摘し、正しい理解を促進するよう改善を図っています。
このようにアルゴリズム設計に関しても、人間の行動を深く理解する人間が設計に関与しなければAIの精度を向上することは厳しく、行動心理学や犯罪心理学といったヒューマンサイエンス的な手法を組み合わせる必要があると考えます。
The ultimate goal of Asilla's products is "to prevent incidents and accidents before they happen and realize a safe and secure world.
In order to realize this world, it is essential to improve the accuracy of AI to achieve the realization of danger prediction.
The key to realizing "prediction of danger" is the accurate detection of even the smallest behaviors and the ability to identify potential risks through the combination of these behaviors. We are dedicated to advancing our research on behavior analysis to a level where AI can effectively detect even the slightest behaviors, make accurate determinations, and be applied across a wide range of real-world products, including Asilla.
Based on the philosophy of "Technology Driven Future," Asilla conducts research and development on a daily basis in order to realize a world in which all people can live safely and securely. In the field of behavior recognition AI, in particular, we have a strong desire to become a world leader, and our researchers and members are working together as a team to constantly evolve.
In the future, we will further promote the further evolution of our behavior recognition AI technology and strive to create an environment where all our engineers can pursue their own careers and always be excited about their work. If you are interested in AI related to human behavior, we look forward to working with you to create something together.
Inquiries about HSAR: pr@asilla.jp Contact: Wakasa
After completing his postgraduate studies at Tokyo Institute of Technology, he worked for JGC Corporation where he was involved in plant design IT work on overseas construction projects. He then joined Asilla Corporation, where he was in charge of proof-of-concept and product development projects related to behavior recognition AI, and became the company's executive officer and CTO in 2022. He continues to focus on product development using AI technology and research and development of new technologies.
