When AI Sees Back Pain: Ergonomics Analysis from Video
24/09/2026
A Tense Meeting Amidst Numbers and Pain
It was a Tuesday afternoon in April, in the third-floor conference room of a large food processing plant in northern Hanoi. The atmosphere was heavy. The Occupational Safety and Health (OSH) department head was tapping a pen on the table, his tone slightly sharp as he confronted the Production Director. The issue was not minor scratches or major accidents, but the rising number of chronic back pain cases in the packaging area. OSH argued that they needed to change the workflow, add supportive chairs, and implement more frequent breaks. The Production Director shook his head, dismissing it as a matter of individual awareness, claiming that altering the production line would reduce productivity and double operational costs. Both sides had valid points, but both lacked one crucial element: visual evidence.
We were invited to that meeting as independent consultants. Looking around, I saw young workers hunched over trays of goods for hours on end. That was when I proposed an alternative: instead of debating based on intuition or paper reports, let the computer observe their work. We would use Computer Vision to analyze ergonomics directly from existing security surveillance videos. No need to install expensive sensors or have workers wear bulky devices. All we needed was image data and an AI model trained to understand the human body. This was the first step in transforming static frames into dynamic alerts about injury risks.
What Everyone Says vs. Factory Floor Reality
People often say that workplace safety is the responsibility of the worker. While technically true, reality in factories across Vietnam, Thailand, and even multinational distribution chains like Masan or TTN, which we have partnered with, tells a different story. Workers typically comply with regulations when supervisors are present, but when no one is watching, they choose the most comfortable posture for themselves, even if it harms their spine in the long run. They hunch down because standing straight tires their necks. They twist their bodies because their arms cannot reach. This is the body’s survival instinct, but it is the enemy of long-term health.
The key point that older systems often overlook is continuity. Traditional ergonomic assessments are usually seasonal, conducted once a year. However, poor posture does not cause pain immediately; it accumulates over shifts and weeks. Computer Vision addresses this by providing continuous observation. We piloted this in an instant noodle packaging area. Instead of just monitoring hand speed, our AI model focused on the lumbar tilt angle, lower back curvature, and neck posture. The output was not an absolute number, but a real-time relative risk index. This allowed the OSH team to intervene at the right time with the right person, rather than waiting for an employee to request sick leave due to chronic back pain.
How It Works and What We Measure
The deployment process is not as complex as many think, but it requires meticulous data preparation. We do not use generic data from the internet. We collect hundreds of hours of video from on-site security cameras, after which AIVISION’s technical team, along with the factory’s ergonomics experts, labels the postures: safe, neutral, and high-risk. This process takes about two weeks, but it forms the foundation for the AI to learn the specific “language” of that factory. Since each facility has different spatial dimensions, camera placements, and workflows, the model must be fine-tuned individually. We successfully applied this method for a partner in the lubricant industry in the South, where bottle-capping operations require significant wrist force. The AI model learned that not only is the arm angle important, but even minor vibrations caused by uneven force application are indicators of recurring injury risk.
Once the model is operational, what we measure is not “the number of injured people” (as that is the result of many factors), but “risk exposure time.” We discovered that about one-third of work shifts contain short but high-intensity periods of poor posture, often occurring at the end of a shift or when minor technical errors occur on the line. These periods were previously completely overlooked in periodic reports. With this data, leadership can make fact-based decisions: instead of demanding workers work faster, they can adjust workstation heights or rearrange material placement. The results are qualitative but clear: fewer complaints, a more comfortable work environment, and a sense of being cared for by management. We do not promise to eliminate injuries entirely, but we commit to significantly reducing unnecessary risks.
Lessons and the Outcome of That Meeting
What would change if I could go back to that first meeting? I would state clearly that technology does not replace humans; it helps humans see what the naked eye misses. I would also emphasize that investing in AI ergonomics is not an expense, but an insurance policy for the team’s health. Initial resistance from the production side is understandable, but it dissipates when they see that productivity does not drop, and may even increase slightly because workers are no longer interrupted by dull pain. Companies like Gene Solutions and other major corporations are also realizing that personnel data is not just in payroll records, but in how employees stand, sit, and move.
Returning to the opening scenario, the outcome of that meeting was not a decisive victory for OSH or Production, but a smart compromise. They agreed to install two additional dedicated cameras with better angles, and we deployed the video analysis system within one month. Six months later, the OSH head returned to that conference room, no longer tapping his pen on the table. He presented a concise report showing a 20% reduction in sick leave due to back pain in the pilot area. The Production Director nodded, said little, but had already started applying a similar analysis model to the bottle packaging area. This story does not end with a solved problem; it begins with a new culture: one where workplace safety is not just a mandate, but something that is seen, measured, and valued. That is the true value of applying Computer Vision to ergonomics: transforming silent pain into actionable data.
This series comes out of projects that actually shipped. More on the AIVISION blog, details on face recognition and the rest of our solutions, or reach out to us.