AI and Machine Vision in HSE: From Idea to Launch

From Idea to a Working System: Machine Vision in Food Production

The implementation of artificial intelligence and machine vision technologies is becoming a natural stage in the development of industrial safety systems. This is especially relevant for the food industry, where HACCP requirements dictate strict rules for hygiene and personnel presence in clean zones. Ivan Maksimov, a representative of the FoodTech company Qummy, shares practical experience in creating and implementing a machine vision system based on Open Source solutions.

The speaker breaks down the journey from the inception of the idea to the launch of a working tool that helps monitor compliance with sanitary standards and HSE rules without resorting to expensive off-the-shelf solutions.

Choosing a Language Model and the Training Process

The foundation for the system was the open language model MoonDream. Choosing an Open Source product allowed the company to save significantly at the start and flexibly configure the system to its needs. The presentation details the model training process:

• Initial training on synthetic data: the model was taught to recognize smoke, the absence of masks, headwear, and gloves, as well as non-standard situations (for example, a person lying on the floor). This allowed laying down basic recognition patterns.
• Fine-tuning on a real video stream: the system was connected to cameras in the factory kitchen. The model began analyzing frames in real time, identifying violations (for example, entering the production zone without sanitary clothing). This ensured high accuracy in the specific conditions of the enterprise.
• Integration with alert systems: for prompt response, sending notifications to a Telegram bot for responsible managers and technologists was set up. The time from detecting a violation to receiving a signal is no more than 15 seconds.

The Psychology of Implementation: From Punishments to Rewards

Technical implementation is only part of the success. The speaker shows, using his company as an example, how important it is to properly organize work with the team when implementing control systems. The main principle is the rejection of financial penalties based on machine vision data.

• The system as an assistant: employees are explained that machine vision is analogous to passive safety systems in a car (ABS, airbags). It is designed for protection, not for fines.
• Positive motivation: instead of "walls of shame," the company uses a reward system. Employees who do not commit violations during the month receive small bonuses or gifts. This reduces staff resistance and forms a culture of conscious safety.
• Preliminary preparation: before launching the system, training sessions and discussions were held, which helped relieve anxiety and even spark interest among employees.

Project Economics and Development Prospects

Creating an in-house system requires investment, but in the long run, it turns out to be more profitable than buying ready-made boxed solutions. The main costs are the programmer's salary (about 6 months of work) and the purchase of server equipment (around 2-3 million rubles). However, in-house development allows scaling the system for free, adding new features (such as facial recognition), and integrating it with the company's internal IT products (1C, Bitrix).

What you will learn from this webinar:

• How to choose and train an Open Source machine vision model for HSE needs?
• How to organize prompt notification of managers about violations using Telegram bots?
• How to overcome staff resistance when implementing total video control systems?
• What are the real costs of developing an in-house machine vision system compared to buying off-the-shelf solutions?
• How does machine vision help reduce the level of micro-injuries in production?