Asilla's Custom VLM Boosts Security Detection Accuracy by 80%
Asilla's move from motion detection to contextual understanding with a custom VLM marks a practical shift in AI security, leveraging proprietary data and edge computing to address real-world facility needs.
Reporting from 1 sources: ASCII.jp.
Asilla has integrated its proprietary Vision-Language Model AsillaVision into its AI Security asilla system, claiming a detection accuracy improvement of up to 80% in pilot deployments. The lightweight 4B-parameter model runs on edge devices, combining with existing action recognition AI to distinguish between falls, fights, and other events. Full-scale deployment at large Japanese facilities began in June 2026.
Asilla, founded in 2015, has been developing action recognition AI for security. Its AI Security asilla system, launched in 2022, is now deployed in over 200 facilities and has accumulated more than 8 million surveillance video data points. The company has now integrated a custom Vision-Language Model, AsillaVision, into the system. The 4B-parameter model runs on edge devices, combining with existing pose-estimation AI to understand context-distinguishing a fall from a person leaning against a wall. In pilot tests, detection accuracy for falls and fights improved by up to 80%. Full-scale deployment at large facilities began in June 2026, with plans to expand detection categories and add predictive features.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.