Computer Vision in Manufacturing: 5 Uses That Aren't Quality Inspection
Most plants use cameras only for defect detection. The same cameras and edge compute can support equipment health checks, PPE and zone monitoring, process drift detection, and verification steps. Here is how to pilot them.
Most plants already have cameras on their lines, and most of those cameras do one job: finding defects. That is a valuable job. It is also a narrow use of hardware that is already powered, networked, and pointed at the equipment.
The missing piece is usually not more cameras. It is inference models trained for other tasks, compute close enough to run them in real time, and integration that routes each detection to the person who should act on it. This article covers five non-inspection uses, where each one fits, and how to run a pilot that proves or disproves the value on your line.
Start with an inventory of what you already have
Before buying anything, walk the plant and list every camera: location, resolution, frame rate, field of view, and who owns it. For each camera, note what else is in frame. A camera watching a conveyor for label defects may also see a bearing housing, a guarded zone entry, or the fill level on a hopper.
The compute side has changed most. Embedded edge modules now run multi-camera inference on the plant floor; NVIDIA, for example, rates its Jetson Orin NX at up to 100 TOPS of AI performance and the AGX Orin at up to 275 TOPS [1]. That makes it practical to run several models against the same camera feeds without sending video to the cloud.
| Use case | What the model looks for | Who acts on it |
|---|---|---|
| Equipment health | Visible leaks, belt tracking, guard damage, thermal hot spots | Maintenance |
| PPE and zone monitoring | Missing PPE at defined points, people inside guarded zones | Supervisors and EHS |
| Process drift | Changes in fill level, web position, flame shape, product placement | Operators and process engineers |
| Verification | A step performed, a component present, a label applied | Operators and quality |
| Counting and staging | Pallets, totes, or WIP at defined points | Logistics and planning |
Five uses beyond inspection
1. Equipment health you can see
Some degradation is visible before it is catastrophic: oil or coolant leaks, belts drifting off track, loose guards, abnormal vibration you can see on a structure, and hot spots on a thermal camera. Vision is good at noticing that something changed compared with a reference image.
Treat vision as a complement to condition monitoring, not a replacement. Vibration and current analysis detect internal faults such as bearing defects and rotor problems that no camera can see. A useful pattern is to let vision flag the visible symptom and attach the latest sensor trend to the alert, so the planner sees both. Warning time is where the value comes from: NIST associated predictive maintenance with 15% less downtime in its study of U.S. manufacturers [2].
2. PPE and zone monitoring
Cameras at zone entries can check for required PPE and detect people inside guarded areas while equipment is running. That can help supervisors spot patterns, such as a particular entry point or shift, and target coaching.
It does not change the legal obligation. OSHA's general PPE standard requires employers to assess the workplace for hazards that necessitate PPE, select it, and have each affected employee use it, and requires protective equipment to be provided, used, and maintained wherever hazards make it necessary [3]. Vision supports that program; the hazard assessment, training, and engineering controls still come first.
Decide the response before you deploy the camera
Monitoring people raises trust and privacy questions that equipment monitoring does not. Agree with supervisors, EHS, and employee representatives what is detected, who sees it, how long images are kept, and whether the response is coaching, an alert, or an automatic interlock. Real-time interlocks belong in the safety system design, not in an analytics add-on.
3. Process drift you can see but are not measuring
Many process problems are visible long before a quality check catches them: a fill level creeping down, a web wandering toward an edge, a flame changing color, parts sitting off-center on a fixture. Manual spot checks sample a few items an hour. A camera can look at every item and flag a trend as it starts.
The key is to compare against the process's normal envelope rather than a single image. Train on a representative range of good production, alert on sustained deviation, and send the alert to the operator who can correct it, with the image attached.
4. Verification steps
Vision can confirm that a step happened: a component is present and oriented correctly, a label is applied, a fastener is in place, a lockout device is on before a door opens. These checks are often simpler to train than defect detection because the target is a clear yes or no. They also produce a record that can be attached to the batch or the work order.
5. Counting and staging
Overhead cameras can count pallets, totes, or work in progress at defined staging points and update planning systems more often than manual counts. Before relying on it, compare camera counts with physical counts in one area for several cycles and confirm the error rate is acceptable for the decisions you will make from it.
Edge or cloud: where to run inference
Run inference at the edge when the response is time sensitive, when video volume would overwhelm the network, or when images should not leave the site. Use the cloud for model training, fleet-wide comparison, and reporting, where a delay of seconds or minutes is acceptable.
Security matters either way. NIST SP 800-82r3, the federal Guide to Operational Technology Security, provides guidance on securing OT while addressing its unique performance, reliability, and safety requirements [4]. Keep cameras and edge devices on segmented plant networks, push results outward through defined paths, and avoid exposing camera streams directly to the internet.
From pixel to work order
A detection only creates value when it reaches someone who can act. A practical pipeline has five stages:
1. Capture. Existing cameras stream to an edge device on the plant network. 2. Infer. Models run on the edge device and produce detections with confidence scores. 3. Classify. Each detection is labelled as equipment, safety, process, or verification. 4. Route. Business rules map severity and type to a destination: a CMMS work order request, an EHS log, or an operator display. 5. Learn. The person who responds marks the detection as correct, wrong, or not actionable, and those labels feed retraining.
Set confidence thresholds per use case and start conservative. A safety detection that is wrong half the time trains people to ignore it. Our guide to CMMS integration covers turning alerts into work orders without flooding planners.
Treat model performance as something to measure, not assume. The NIST AI Risk Management Framework is intended to help organizations incorporate trustworthiness considerations into the design, development, use, and evaluation of AI systems [5]. In practice, that means a documented test set from your own line, measured accuracy before go-live, and a review of false positives and misses every month.
A 90-day pilot
Weeks 1 and 2: audit and select. Inventory cameras and pick one use case with a clear owner and a clear baseline, such as a recurring leak point, a guarded zone, or a fill-level problem.
Weeks 3 to 6: shadow mode. Run the model without sending alerts. Compare its detections with what people observe and with your existing records.
Weeks 7 to 10: tune and connect. Adjust thresholds, then route detections to the owner through the CMMS, EHS system, or operator display.
Weeks 11 and 12: measure. Compare the result with the baseline: detections confirmed, false positives, time to response, and the outcome you care about, such as leaks caught, zone entries avoided, or scrap reduced. Decide from the numbers whether to expand.
For the broader picture of where vision fits in the plant, see our guide to computer vision in manufacturing and our computer vision platform page.
Frequently asked questions
Can existing quality inspection cameras be used for other tasks?
Often, yes, if the equipment or area of interest is in frame and the resolution and frame rate suit the task. New models and edge compute are usually the main additions.
Can computer vision replace vibration monitoring?
No. Vision sees visible symptoms such as leaks, belt drift, and hot spots. Vibration and current analysis detect internal faults a camera cannot see. Use them together.
Does PPE detection satisfy OSHA requirements?
No. Employers must still assess hazards, select PPE, and have employees use it [3]. Vision can support compliance by showing where and when gaps occur.
Should computer vision run at the edge or in the cloud?
At the edge for time-sensitive responses and high video volume; in the cloud for training, fleet comparison, and reporting.
References
[1] NVIDIA, "Jetson Orin." https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/
[2] NIST, "Research Suggests Significant Benefits to Investing in Advanced Machinery Maintenance," 2020. https://www.nist.gov/news-events/news/2020/06/research-suggests-significant-benefits-investing-advanced-machinery
[3] OSHA, "29 CFR 1910.132, Personal Protective Equipment: General requirements." https://www.osha.gov/laws-regs/regulations/standardnumber/1910/1910.132
[4] NIST, "SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security," September 2023. https://csrc.nist.gov/pubs/sp/800/82/r3/final
[5] NIST, "AI Risk Management Framework." https://www.nist.gov/itl/ai-risk-management-framework
Related articles
Ready to put this into practice?
See how Monitory helps manufacturing teams implement these strategies.