Russian Scientists Develop Lightweight Neural Network for Industrial Safety
Engineers in St. Petersburg have built an AI model that spots unsafe behavior in real time, already cutting safety inspections threefold at a major factory.
Researchers at ITMO University in St. Petersburg have introduced ActionFormer, a neural network designed to automatically detect dangerous actions in industrial settings. The system was piloted at a large plant in the Perm region, where it has already shown strong results.
The model can simultaneously recognize up to ten types of incidents, including smoking, eating, phone distractions, unauthorized equipment movement, and attempts to tamper with surveillance cameras. Its accuracy reaches 80 percent—significantly higher than that of comparable open-source foreign systems.
What makes the system stand out is its efficiency. With only 3.7 million parameters, ActionFormer uses a precision-focused architecture that analyzes just key points and object masks. That makes the AI lightweight, less demanding on computing resources, and easier to deploy.
The network was trained on a unique dataset of more than 180,000 frames, part of which was collected in-house to improve recognition of highly specific industrial scenarios.
According to its developers, the system not only enhances workplace safety but also reduces costs associated with inspection procedures.
Because of its flexible design, ActionFormer can be adapted for use in wearable cameras or in public surveillance systems.
