AI and Video Analytics in HSE: From Idea to In-House Launch

Artificial Intelligence in HSE: From Idea to a Working System

The implementation of machine vision and artificial intelligence systems in HSE is often perceived as an expensive and complex process available only to large corporations. However, modern technologies allow such projects to be implemented by internal specialists. Ivan Maksimov, an HSE specialist at Qummy, shares his practical experience in creating and launching a video analytics system in food production using open-source solutions.

Prerequisites for Implementation: Why Automated Control Was Needed

Food production has strict sanitary standards and safety requirements. The speaker analyzes the main problems that pushed the company towards automating control:

• Violation of sanitary clothing rules: workers regularly forgot to wear gloves, hairnets, or change shoes when moving between clean and dirty zones.
• Unauthorized actions with equipment: attempts at self-repair or interfering with the operation of active mechanisms without proper qualifications.
• High time costs for managers: line managers had to spend a significant part of their working time visually monitoring compliance with the rules.

Choosing the Technology: MoonDream Open-Source Model

Instead of purchasing expensive commercial solutions, the company chose the path of in-house development based on the open-source language model MoonDream. The presentation details the functionality of this system:

• Real-time video stream analysis: the model is capable of analyzing images frame-by-frame and identifying specified patterns (lack of PPE, smoke, a person falling).
• Training flexibility: the system is trained on real photos and videos from the production facility, allowing it to be adapted to the specific conditions of a particular enterprise.
• Cost-effectiveness: using an open-source product on the company's own server capacities eliminates the costs of licenses and cloud storage.

Implementation Stages: From Pre-training to Commercial Operation

The implementation process took several months and included sequential steps:

• Pre-training on artificial data: creating basic models for recognizing fire, smoke, and the presence of masks, gloves, and headwear.
• Training on a real video stream: adapting the model to the conditions of specific production, correcting errors and false alarms.
• Setting up the notification system: developing a Telegram bot to promptly inform shop managers about identified violations, indicating the date, time, and location.

Implementation Results and Work with Personnel

During the pilot operation period, the system identified over 2,700 non-conformities, while the share of false alarms was reduced to 5%. The speaker shows by example how automation affected the safety culture:

• Increased self-discipline: the number of violations of sanitary clothing rules decreased significantly.
• Reduction in micro-injuries: a decrease in the number of micro-injuries was recorded compared to the same period last year.
• Positive motivation: instead of punishing for violations, the company introduced a reward system for workers who do not deviate from safety rules.

What you will learn from this webinar:

• How to deploy a video analytics system based on free open-source solutions?
• What training stages must a neural network go through to work correctly in production?
• How to set up prompt notification of managers about violations via Telegram?
• How to overcome personnel resistance when implementing total control systems?
• What are the technical requirements for cameras for effective video analytics operation?