On a production floor, many important signals never appear in an ERP system, spreadsheet or report. They can be seen through a camera, heard in the way a machine operates or measured by a sensor.
Movement in a working area. A part stopping on a conveyor. A product placed in the wrong position. An employee entering an area that should be closed. A missing component on a workstation. Unusual machine behaviour.
These are situations a person may notice, but not always at the right moment. Computer Vision and IoT can turn observation into a working system: one that detects defined events, reacts to them and sends the right information to the right person.
The goal is not to make a camera “watch everything”. The goal is to make the system understand specific situations that matter for production, quality and employee safety.
Computer Vision detects events, not just objects
The simplest way to imagine Computer Vision is object recognition. In practice, its capabilities are much broader. A system can analyse not only what appears in a frame, but also where it is, how it moves and how the situation changes over time.
Depending on the process, Computer Vision can support:
- detecting the presence of a person or vehicle,
- tracking object movement,
- detecting entry into a defined zone,
- checking whether an object appears in the right position,
- detecting missing components,
- checking the order of actions,
- identifying visual defects,
- detecting collisions or unsafe proximity,
- monitoring process stoppages,
- identifying changes from a normal scene.
This changes the questions the system can answer. It does not have to ask only: “What is visible?”. More useful questions may be:
- Did something appear where it should not?
- Is it moving in the wrong direction?
- Has it stayed in this zone for too long?
- Is the process behaving differently than usual?
- Did the expected action follow the previous one?
- Is a required component missing from the workstation?
Analysing relationships, movement and time is what makes Computer Vision useful in real processes rather than just a digital object catalogue.
Detecting movement and presence
Motion detection is one of the basic uses of vision systems, but its value depends on what the system does with the detected movement.
A simple change in the image is often not enough. A shadow, lighting change or moving piece of plastic may trigger an alert that has no operational meaning. The system should therefore distinguish the type of object, its direction, the area it occupies and how long the event lasts.
It may detect:
- a person appearing near a machine,
- a vehicle passing a defined point,
- an object entering a controlled zone,
- a product moving along a conveyor,
- a part stopping for too long,
- an object moving in the wrong direction,
- the absence of expected movement during a defined period.
This allows the system to do more than record an event. It can send a notification, save the event, update an operator panel or communicate with another part of the infrastructure.
In some processes, the speed of movement matters more than movement itself. An object may still be moving, but too slowly, too quickly or irregularly. That change may indicate a mechanical problem, blockage or deviation from the normal process.
Restricted zones and employee safety
One of the most practical applications of Computer Vision is monitoring areas that nobody should enter under specific conditions.
This may include:
- a robot operating area,
- the zone around a press or machine,
- a part of the floor closed during changeover,
- an area reserved for authorised personnel,
- a maintenance zone,
- a vehicle or forklift lane,
- a high-risk area.
The system can detect a person entering such a zone while also considering when and under what conditions the event occurred. Entry while a machine is stopped may have a different meaning than entry during active operation. This is why Computer Vision is often most useful when combined with data from machines, controllers, sensors or production management systems.
Reducing false alerts is equally important. If an operator receives dozens of irrelevant notifications every day, they will quickly stop taking the system seriously. A good solution should consider duration, direction, zone, schedule and process context.
Computer Vision does not replace safety procedures or physical protection systems. It can provide an additional layer of observation and fast notification, especially when one person cannot continuously monitor many areas at the same time.
Detecting process deviations
Not every production issue appears as an obvious product defect. Sometimes the problem is the order of actions, the position of a component or a situation that differs from the expected process.
A vision system can support the detection of:
- missing components,
- incorrectly positioned elements,
- incorrect assembly,
- foreign objects,
- changes in product position,
- visual inconsistencies,
- skipped process steps,
- repeated workstation errors.
The objective is not always to reject a product automatically. In many cases, it is better to flag the case for additional review. The system can stop a process, mark an item or notify an operator who makes the final decision.
This combines automation with human control. The model does not have to be an infallible judge. It can be a fast filter that identifies situations requiring attention.
Detecting anomalies
Not every situation can be described through a closed list of rules. In many processes, the important task is detecting behaviour that differs from the normal pattern.
An anomaly may be:
- an unusual movement path,
- a sudden stop,
- an object staying in one place too long,
- a change in movement density,
- unusual machine behaviour,
- a visual difference in a product,
- an object appearing at the wrong time,
- a combination of several small deviations.
Anomaly detection requires a definition of what “normal” means. This is not always one perfect image. More often, it is a range of acceptable behaviours that may vary depending on the shift, product, workstation or production stage.
The system should therefore support threshold tuning and historical event analysis. Otherwise, every deviation may be treated as a failure, and the production floor starts to resemble a fire alarm triggered by a toaster.
A well-designed anomaly detection system does more than produce alerts. It helps identify when and under which conditions deviations happen most often.
IoT provides context
A camera image can be useful, but it often does not contain enough information to make the right decision. A person in a restricted zone may mean something different depending on the machine state. A stopped product may be normal during a changeover but concerning during standard operation.
IoT connects visual information with other data sources:
- machine state,
- temperature,
- speed,
- sensor signals,
- batch identifiers,
- production schedule,
- maintenance status,
- production system data.
This makes it possible to create a more precise rule:
“Send an alert if a person is detected in the robot zone, the robot is active and no maintenance procedure is in progress.”
That is far more useful than simply saying “a person was detected”.
IoT can also preserve the full context of an event. Later, the team can review what the camera detected, what the device state was and what reaction followed.
Edge AI when time matters
In safety and production monitoring systems, latency can matter. When image analysis happens only in the cloud, data needs to be transferred, processed and sent back.
Edge AI allows part of the analysis to happen locally, directly on a device or on a local computer close to cameras and machines. This can help the system react faster to movement, restricted-zone entry or process changes.
Local processing can also reduce the amount of video sent outside the production environment. The backend may receive an event, metadata or an analysis result instead of the complete video stream.
This does not mean every analysis should happen locally. The decision depends on latency, infrastructure, connectivity, privacy and operating cost requirements. A common architecture distributes responsibility:
- the camera and edge layer detect the event,
- the local system reacts immediately,
- the backend stores the event and history,
- the application gives the operator context,
- the cloud supports reporting or further analysis.
From an alert to a useful response
An alert alone does not solve a problem. The system should define what happens after an event is detected.
Possible actions include:
- notifying an operator,
- storing the event in history,
- marking an image or video segment,
- stopping the process when the infrastructure supports it,
- sending information to a higher-level system,
- creating a task for a responsible person,
- escalating when nobody confirms the event,
- collecting data for later analysis.
Every alert should have context. The operator needs to know where the event occurred, when it started, what it involved and whether someone has already responded.
Otherwise, the result is another dashboard glowing red and waiting for someone with the required level of patience.
Start with scenarios
The best Computer Vision systems do not begin with the question: “Which model should we choose?”. They begin with a scenario.
First define:
- which event should be detected,
- why it matters,
- what the normal process looks like,
- which deviations are important,
- what reaction should follow,
- who should receive the information,
- how success will be measured.
Only then should the team choose cameras, processing methods, models, edge devices, backend services and interfaces.
Testing in real conditions is equally important. Different lighting, dust, an occluded object, a new product or a changed workstation configuration can affect the result. The system should be developed using real cases, not only ideal recordings prepared for a presentation.
Computer Vision as part of the production system
Computer Vision and IoT can support quality, safety and production management, but their value does not come from using a camera, sensor or AI model by itself.
The value appears when the system:
- detects specific events,
- understands their context,
- reduces false alerts,
- reacts at the right time,
- stores history,
- supports human decisions,
- can be monitored and improved.
A well-designed solution does not try to replace the entire process with one model. It connects camera data, devices, backend services and user applications into one coherent system.
Then the production floor image stops being just a recording. It becomes a source of information that can help detect deviations earlier, respond to hazards and manage the process more effectively.
Have a system that
simply has to work?
Tell us what should be built, who will use it and which parts are the most critical. We reply within one working day — usually straight away with the questions that pin down the biggest risks and the next step.