T-NEX
Menu
Camera AI for companies

Use camera AI to detect what matters in your operations.

Detect objects, measure occupancy or assess visual quality features: we develop camera AI for your task. Camera, model and environment are considered together. A pilot under your conditions shows which detection capabilities can be used in practice.

Models selected for the taskLocal, hybrid or cloud operationQuality assessed in a pilot
T-NEX · Detection · Camera 03 · HallEdge inference
CAM-03 · 1/125 · non-biometricPerson 0.98PPE: helmet ✓Safety zone · clearModel mixOn-premiseno identification
Illustrative rendering: schematic detection frames, not a real camera image, no person identification.
Camera
Agree optics, lighting and installation
Model
Select the detection approach
Pilot
Assess detection quality and false alarms
Operation
Define data flow and response
T-NEX camera AI · event review

Turn a detection into an actionable event.

From the detection task to a reviewed notification: events are linked to their source, assessed and passed on for follow-up.

Define the task

Define the task

The image source, analysis area and target event are defined. This view records occupancy of an agreed zone.

A clear detection task for the pilot.

Detection task

SettingSelection
Image sourceCamera A · test environment
Analysis areaZone 1
EventZone occupied
Follow-upHuman event review

Review events

The overview links detected events to their source and zone. False alarms and unclear cases are part of evaluating detection quality.

Detections are reviewed in their operational context.

Event list

TimeSource / zoneEventReview
10:14Camera A / 1OccupancyConfirmed
10:18Camera A / 1OccupancyUnclear · review
10:22Camera A / 1OccupancyFalse alarm

Assign follow-up

Relevant events get an agreed responsible role and next step. An automatic machine response would require separate design and validation.

A defined workflow connects detection and response.

Event 10:18 · follow-up

  1. 1
    Event recorded

    Source and zone assigned

    Complete
  2. 2
    Review situation

    Responsible role evaluates the case

    Open
  3. 3
    Record follow-up

    Record the decision in the agreed workflow

    Waiting
Definition

Analyse camera images for specific tasks

AI camera analytics automatically evaluates camera images to detect specific things in them: Whether and where a person is, whether protective equipment is being worn, whether an object is present or an area has been entered. Technically these are computer vision tasks: classification (what is in the image), object detection (what and where, as a bounding box), segmentation (pixel-accurate area) and tracking (the same object across frames, with a track ID, not with an identity). Crucially, the system detects attributes and states, not the identity of individual people.

Presence, object, PPE, zone: detected, not identified.

Presence ≠ identity
Use cases

Where does camera-based AI detection create value?

You will recognise your case here, from workplace safety in manufacturing to shelf monitoring in retail. For each case we choose the right detection task and the right model.

Protective equipment

Detect visible features such as a helmet or safety vest and provide indications for further review.

Zones and access

Detect entry into defined image areas. Safety functions and machine responses require separate design and validation.

Counting and occupancy

Count objects or people and assess occupancy. Data minimisation is planned for the specific use.

POS anomalies

Flag visual anomalies for human review; a detection is not evidence of misconduct.

Shelf availability

Make empty spaces and defined deviations visible for operational handling.

Visual quality

Check for missing parts, visible defects or deviations in products.

Further use cases on request: logistics/warehousing (cargo damage, forklift-pedestrian proximity), construction site safety, perimeter/access monitoring, agriculture (counting, condition monitoring). ⚠︎ Sensitive environments (e.g. healthcare) and anything involving personal data are assessed separately.

Technology

Choose the right model for the detection task

Pretrained detectors, adapted models and vision-language models support different tasks. Selection and combination depend on image data, required accuracy and permitted response time.

Three approaches and when each one fits
ApproachBest forImage data requiredLatency / operation
Pretrained detectorcommon object classes and a starting point for the pilotsample data for assessmentdepends on model and hardware
Adapted modelcustom objects and site-specific featuressuitable labelled training and test datameasure on the intended system
Vision-language modelscene-related questions and supplementary interpretationrepresentative test examplesassess cost and response time separately
Detector: Recognise defined object classes and provide the results for the next workflow step.Detection
Additional interpretation: A vision-language model can address further questions about selected images; its usefulness and response time are evaluated in the pilot.Context
Operating setup: Camera, optics, lighting and computing resources are assessed as one system.Practice
Honestly: no miracle model

The pilot assesses representative situations, including difficult lighting and false alarms. Results determine the achievable scope and required human review.

Camera & hardware

Camera selection and complete kit: hardware, model and software fit together

Camera, optics, lighting, installation and compute capacity are specified together. Existing cameras and video management systems are assessed for the required interfaces and fitness for the use case.

Installation

Consider field of view, distance, weather and mechanical protection.

Environment

Assess temperature, dust, vibration and power supply.

Lighting and sensor

Select lighting, infrared or thermal imaging for the detection task.

Image and compute

Plan resolution, frame rate and processing requirements together.

Edge or central

Choose local devices or central compute capacity based on the requirements.

Integration

Check connections to existing cameras and video management systems in the actual setup.

Operations

On-premise or cloud: where do the images live?

Analysis can run locally, in a hybrid setup or in an agreed cloud environment. Data flow, access needs, response time and operating costs guide the design. Local operation alone does not make processing automatically compliant with data protection law.

On-premise vs. cloud, an honest comparison
CriterionOn-premise / edgeCloud
Data flowlocal processing according to architecturetransfer to agreed services
Response timemeasure on siteaccount for network and service latency
Operating costshardware, maintenance, energy and supportusage, transfer and support
Model choicelimited by local resourcesdetermined by service availability and terms
Data protectionassess legal basis and safeguardsalso assess providers and data transfers
Data protection and scope

Assess the specific camera AI deployment

Detection without identification can still process personal data. Purpose, legal basis, field of view, storage and access are defined for the deployment. AI Act classification also depends on intended purpose; being non-biometric is not a blanket exemption.

Checks for the use case
CategoryWhat it meansClassification
Detection goalObjects, conditions or occupancyDefine scope and error tolerance in the pilot
People in the imageIdentifiability and possible impactsAssess data protection and minimisation
WorkplaceRecording employeesAccount for employee privacy and participation rights
Safety functionMachine response or protective measureSeparate design and validation required
Biometrics and sensitive analysisIdentification or inference of personal traitsOutside the standard scope described here
Purpose and field of view: Analyse only areas required for the taskPlanning
Storage and access: Define deletion rules, permissions and data flowsOperation
Impact assessment: Assess whether one is required and supply technical documentationDocumentation
Our line

The standard scope focuses on objects and conditions. Biometric identification, emotion monitoring and covert continuous surveillance are outside this offering.

Before use

Before deployment, assess the lawfulness of the specific use, any required impact assessment and the classification of the intended purpose.

Honest limits

Define the right scope of use

Trust also comes from clear boundaries. We do not sell camera AI that is legally or ethically problematic, and we say where technology or the law draws the line.

No emotion monitoring

Inferring employees’ emotions is outside this offering.

No covert continuous surveillance

Purposes, capture areas and access are defined in advance.

No standard identity recognition

The described scope detects objects and conditions.

Assess quality in the real environment

Detection quality and false alarms are assessed using representative situations.

Decision aid

Selection checklist: commissioning AI camera analytics

A vendor-neutral checklist: What matters in a credible camera AI partner, whoever you choose.

  1. 01Task: What should be detected, and which response follows?
  2. 02Data: Which people or confidential information appear in the image?
  3. 03Technology: Do camera, lighting, installation and compute fit the task?
  4. 04Pilot: Which situations, false alarms and edge cases will be tested?
  5. 05Operation: Who maintains the system and reviews its results?
FAQ

Frequently asked questions

What can camera AI detect?

Depending on the model and environment, it can detect objects, occupancy, protective equipment, zone events or visual quality features. The pilot assesses the specific scope.

Can we use existing cameras?

We assess image quality, installation, interfaces and compute capacity. Suitable additions are specified where necessary.

Can it run locally?

Yes. Local, hybrid and cloud options are planned for the use case, including data flow, maintenance and access.

Is non-biometric analysis automatically compliant with data protection law?

No. It may still process personal data. Purpose, legal basis and safeguards must suit the specific deployment.

Can AI replace a safety control?

General image detection is not a promised safety function. Machine responses and protective functions require separate design and validation.

What image data does a camera AI pilot need?

Representative examples from the intended setting, including different lighting, distances, occlusions and relevant exceptions. If a model needs adaptation, training and evaluation data are planned separately. This allows quality to be assessed on situations that were not used only for adaptation.

How should we assess detection quality?

Define the events to detect and which false alarms or missed cases would cause problems in the workflow. The pilot evaluates those outcomes alongside response time and the human review required. A high model score for a single image does not replace this assessment.

Does camera AI replace a video management system?

The analysis detects features and events. Camera administration, recording and playback are video management tasks. For an existing setup, we review how images and detection results can be transferred and where employees will handle a detected case.

Your use case: from pilot to operation

Tell us what should be detected and where. We scope a pilot and choose the right model and camera, on-premise or cloud, non-biometric by default.

Discuss your use caseAn offering of T-NEX GmbH · Legal notice below

Sources & further reading

  1. DSGVO / GDPR, Art. 5, 6, 13, 35Regulation (EU) 2016/679eur-lex.europa.eu
  2. EU AI Act: bestimmungsgemäßer Zweck / intended purposeRegulation (EU) 2024/1689eur-lex.europa.eu
  3. Mitbestimmung / employee participation§ 87 BetrVGgesetze-im-internet.de
  4. T-NEX guideWorks council and AI: planning implementation togetherT-NEX

Management: Andreas Unruh and Christoph Gembruch.

Published by T-NEX GmbH.

Useful next steps