Asilla, Inc. (Headquarters: Machida City, Tokyo; Representative Director and CEO: Go Onoue; hereinafter "Asilla"), which champions the vision of "Creating a safe and comfortable world through the power of technology," has integrated its independently developed Vision-Language Model "AsillaVision" into "AI Security asilla," the AI security system it develops and provides. In an early adoption in spring 2026, the company confirmed an improvement in detection accuracy of up to approximately 80%*1 compared with conventional methods, and from June it will begin a full-scale rollout to large-scale facilities in Japan.
"AsillaVision" is a domestically developed VLM created by Asilla specifically for security camera footage. By capturing the situation itself from video, it works toward realizing a new stage in which security cameras advance from merely detecting movement to "understanding situations."
*Comparison value against conventional methods, based on operational data at early-adoption sites and the company's own verification (verification period: March–May 2026)
Since its founding in 2015, Asilla has specialized in the research and development of behavior-recognition AI that understands human behavior from video, developing proprietary models that classify behaviors such as falls, fights, intrusion, and loitering with high accuracy. In 2022, the company began providing "AI Security asilla," an AI security system that leverages this technology, and it is now deployed at more than 200 facilities in Japan and overseas. The security camera footage data accumulated through operations at deployment sites nationwide exceeded a cumulative total of 8 million cases as of May 2026*. At the same time, the company has built up insights that cannot be obtained from data alone — namely, what is truly required to improve accuracy in real-world environments.
What Asilla focused on as a means to further evolve these technological assets and on-site insights was the VLM (Vision-Language Model, hereinafter "VLM"). VLMs, which understand video and language in an integrated manner, have advanced rapidly in recent years, but applying a general-purpose VLM directly to security settings is not optimized for the differing environments of each facility or for identifying the events that must be addressed, and it falls short of the required accuracy and responsiveness.
Believing that its own technological assets and on-site insights are what can close this gap, Asilla began developing "AsillaVision" — a domestically developed VLM specialized for security camera footage, built on its accumulated proprietary data — at the end of 2025. After proof-of-concept testing in actual facility security environments, the company implemented "AsillaVision" in "AI Security asilla" and conducted an early adoption at select facilities from spring 2026. Having confirmed its effectiveness in real-world environments, the company will begin a full-scale rollout to large-scale facilities in Japan from June.
*Data is collected and used after obtaining the consent of deployment facilities and applying anonymization processing.
"AsillaVision" is a domestically developed, security-specialized VLM that Asilla developed from its own proprietary data. With a lightweight design of 4B (4 billion) parameters, it runs in an edge environment within the facility and can operate without transmitting video outside the facility. It is a new security-specialized model that combines accuracy adapted to real-world environments with an architecture that can be lightly embedded into products.
In "AI Security asilla," "AsillaVision" is not used on its own; instead, it is implemented in combination with the posture-estimation-based behavior-recognition AI that Asilla has refined over the years. The behavior-recognition AI captures the movement of "a person falling down," and the VLM then reads the context of the video. For example, when a person is lying on the floor, the system can distinguish whether they have fallen, are resting against a wall, or are lying down after a scuffle — and by discerning the situation to this level, it can more accurately capture the events that truly require a response. Through this combination, the company succeeded in improving detection accuracy by up to approximately 80%*1.
As a first phase, the targeted detection covers falls and fights, and it will be expanded sequentially going forward. As for deployment facilities, the rollout will center on commercial facilities, office buildings, and others already using "AI Security asilla," and will be extended sequentially to new facilities as well.
The training data is collected with the consent of deployment facilities and, after a privacy impact assessment (PIA) by a third-party organization, information that could identify individuals is removed.
*Comparison value against conventional methods, based on operational data at early-adoption sites and the company's own verification (target: falls and fights; verification period: March–May 2026)
Asilla has set forth the concept of "moving from AI that detects behavior to AI that understands situations," and this full-scale rollout marks the first step toward realizing it.
This time, the company improved detection accuracy through the combination of behavior recognition and a VLM. Going forward, in addition to expanding the events that can be detected, Asilla will support the transition from a reactive to a preventive style of operation through the provision of functions such as extracting signs of danger, automatically generating video logs, and assisting administrators' decision-making.
As the development of general-purpose VLMs advances on a global scale, what we have pursued is the realization of a model that is truly usable on the front lines of security. Implementing our in-house developed VLM "AsillaVision" in "AI Security asilla" and beginning operation at multiple large-scale facilities is, we believe, an important step in connecting the technology cultivated through R&D to real-world, on-site value.
We will feed the on-site insights gained from this product implementation back into next-generation models, contributing to solving the social challenge of labor shortages and to strengthening the competitiveness of AI foundation models originating in Japan.
In product development, what we value is not the novelty of the technology itself, but whether it can continue to be used in the field. We believe that even a new technology like a VLM only has meaning when it leads to genuine value at our customers' facilities.
This "AsillaVision" was first deployed in the environments of customers already using "AI Security asilla," and we have refined it while verifying its effectiveness in actual operation. Being able to reflect on-site feedback one by one was, more than anything, the reason we were able to take the step into a full-scale rollout. We will continue to evolve it into a product that customers can entrust with peace of mind — one that is grown together with customers through use.
Asilla, a Developer of Proprietary Algorithms, Aims to Revolutionize AI Security Accuracy by Fusing Behavior-Recognition Technology with a VLM
https://jp.asilla.com/post/vlm-202603
Video-Analytics AI Company Asilla Develops Industry-Specialized Proprietary VLM "AsillaVision" Using Proprietary Data of Over 7 Million Security Camera Footage Cases
https://jp.asilla.com/post/rd-asillavision-20260408
Asilla Selected for "GENIAC," a National Project Conducted by the Ministry of Economy, Trade and Industry and NEDO — Promoting the Development and Social Implementation of the 4B/27B Foundation Models of Its Industry-Specialized Proprietary VLM "AsillaVision"
https://jp.asilla.com/post/news-geniac-20260604
"AI Security asilla" is a system in which AI analyzes footage from existing security cameras 24 hours a day, 365 days a year, instantly detecting abnormal behaviors such as violence, falls, and intrusion, as well as behaviors requiring attention such as wandering, crowding, and signs of poor physical condition. Amid an increasingly serious shortage of security personnel, it captures anomalies that are easily overlooked by human monitoring and immediately notifies security staff and administrators. Because existing cameras can be used as-is, no capital investment is required, making it a next-generation security solution that maintains a high level of safety even with limited staff.
Representative: Go Onoue, Representative Director and CEO
Location: 1-4-2 Nakamachi, Machida City, Tokyo
Business: Development and provision of various products and solutions based on behavior recognition AI
Official website: https://jp.asilla.com/
.png)