Computer vision · Object detection
What is in the framebecomes information.
Object detection is the computer vision technique that finds the objects you are looking for in an image and marks them with a position and a confidence score. We train or adapt models for the objects that matter to you, from a pallet or a machine to a particular type of packaging, and the result becomes information that people or other systems can use.
A generic model does not know what matters to you
Off-the-shelf models recognise common categories, such as people or cars, but rarely tell apart your types of packaging, machinery or parts. Without an adapted model, images remain just images, and someone has to look at them to find out what they contain.
- Objects specific to your industry are missing from generic models or are confused with one another
- Whether an object is present is checked visually, by a person, at random moments
- Images can only be searched by date and camera, not by what they show
- Internal applications, the ERP or the WMS receive no information from what the cameras capture
Who it is for
- Companies that need to know whether a particular object is present in or missing from an area
- Operations teams that want to turn images into data for other applications
- Businesses that have tried generic models and found they do not recognise their specific objects
- IT teams building their own application that need a visual detection component
What the detection model can do
- 01
Recognises the classes you define
The model learns the categories we agree together, for example types of pallets, containers or machinery, and tells them apart.
- 02
Locates objects in the frame
Each object found gets a bounding box, so you know not only what it is but also where it is.
- 03
Checks presence or absence
Can flag when an expected object is missing from an area or when one appears that should not be there.
- 04
Attaches a confidence score
Each result comes with a confidence level, used to separate clear cases from those that need a human look.
- 05
Sends structured results
Detections are sent as data (class, position, time, camera) via API or messages to your applications.
- 06
Indexes images by content
Analysed images can later be filtered by the objects detected, not just by time and camera.
How it works
How an image becomes a usable detection
The steps marked in yellow are approved by a person.
- Input
- AI
- Action
- Human in the loop
- Result
- 01
Input
Image or frame
A compatible camera or a set of photos provides the images to analyse.
- 02
AI
Detection and classification
The model looks for objects from the defined classes and estimates the position of each.
- 03
Action
Confidence filtering
Detections below the set threshold are put aside for review.
- 04
Human in the loop
Review of uncertain cases
A team member confirms or corrects uncertain detections.
- 05
Action
Sending to systems
Validated results reach the chosen application or database.
- 06
Result
Structured data
You have a record of detected objects that is searchable and measurable.
What data a detection model needs
- Real images or video footage from your site, captured in normal working conditions
- Labelled examples for each class, including difficult situations: partly covered objects, objects at a distance or against the light
- A clear list of object classes and the rules for telling them apart
- Access to camera streams (for example via RTSP or ONVIF, where the equipment supports them) or to an image archive
- Processing resources suited to the volume: edge device, local server or cloud
Possible integrations
- IP cameras with an RTSP stream
- ONVIF-compatible cameras
- NVRs and VMS systems
- NVIDIA Jetson edge devices
- Local servers with GPUs
- Cloud storage, for example Amazon S3 or Azure Blob Storage
- REST API and webhooks
- MQTT
- PostgreSQL databases
- IP cameras with an RTSP stream
- ONVIF-compatible cameras
- NVRs and VMS systems
- NVIDIA Jetson edge devices
- Local servers with GPUs
- Cloud storage, for example Amazon S3 or Azure Blob Storage
- REST API and webhooks
- MQTT
- PostgreSQL databases
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What stays with people
- Defining the object classes and the criteria for an object belonging to one or another
- Reviewing low-confidence detections before they are used in decisions
- Validating the labels in the training set, because labelling errors carry over into the model
- Approving any action that a detection triggers automatically in other systems
Limits and operating conditions
- Detection works best on clearly visible objects; those that are very small in the frame, partly covered or very similar to each other are harder to detect
- The model only recognises the classes it was trained on; a new object needs new examples and retraining
- Differences in lighting, angle or resolution compared with the training data can reduce detection quality
- Not every existing camera provides an accessible stream or a clear enough image; compatibility is checked for each device
How we evaluate object detection accuracy
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Detection precision
- The share of correct detections out of all those flagged, checked manually on real images from the site
- Missed objects
- Objects present in images but not detected, counted by the team on an independently labelled sample
- Confusion between classes
- The confusion matrix on the test set, to see which classes get mixed up with each other
- Stability across conditions
- Results compared separately on images taken by day, by night, against the light or from different angles
How we build a detection model, in 4 steps
- 01
We define the classes and the purpose
We agree which objects matter, how they differ from each other and what happens with the result of a detection.
- 02
We collect and label images
We gather representative images from your site and label them together with the people who know the objects.
- 03
We train and test
We adapt a model, test it on images it has not seen and discuss openly where it makes mistakes.
- 04
We integrate and monitor
We connect the results to your systems and monitor quality over time, retraining when conditions change.
Use cases
The scenarios below are conceptual examples, designed to illustrate realistic applications. They are not implementations delivered by WEB-AZ.
- Construction
Tracking machinery on site
A camera at the site gate recognises the type of machinery coming in and going out, and the record goes into a table checked by the site manager.
Conceptual example - Distribution
Packaging type on the sorting belt
The model tells apart boxes, sacks and multipacks on the belt and sends the type to the sorting system; uncertain cases are stopped for an operator.
Conceptual example - Insurance
Pre-sorting damage photos
Elements such as the windscreen, headlights or wheels are detected in photos sent by customers, so claim files reach the assessor who evaluates them in order.
Conceptual example
Useful guides
Before you decide, read
Frequently asked questions
What is AI object detection?
It is a computer vision technique in which a model finds objects from defined classes in an image or video frame, for example types of pallets or machinery, and gives the position and a confidence score for each. It is the basic building block for counting, visual inspection and video alerts.
How does an adapted model differ from an off-the-shelf one?
Generic models recognise common categories. When your objects have specific shapes, colours or markings, a model adapted with images from your site can make the difference. We first check whether an existing model is enough before proposing to train a new one.
How many labelled images are needed?
There is no fixed number. It depends on how many classes there are, how similar they are and how varied the filming conditions are. We start with a small set, measure the results and expand collection where the model makes mistakes.
What happens when the model is not sure?
Each detection has a confidence score. Below the set threshold, the result is not used automatically but goes to a person who confirms or corrects it. The corrections can then be used for retraining.
Who has access to the images used for training?
Only the people involved in labelling and development, with access set out in the contract. Where images show people, we blur faces or crop only the relevant area, and the retention period is agreed before collection.
Does it work with the cameras I already have?
It can, and we check in practice. We review resolution, angle, streaming protocol and image quality in real conditions before confirming compatibility.
Can’t find the answer? Write to us and we’ll reply directly.
Do you have objects your cameras should recognise?
Send us a few real images and the list of objects of interest. We will tell you whether the task suits detection and what data would be needed.
Related solutions
- People and object countingAnonymous counting of crossings, zone occupancy or the items in a batch, without identifying anyone.
- Industrial visual inspectionCameras on the production line flag parts with visible defects, and the operator confirms every rejection.
- Edge AI for computer visionVideo analysis on devices installed on site: fast response, less network load and images that never leave the building.