AI Video Analysis: Requirements and Process in Operations

What are the requirements for AI video analysis?

The most important requirement for AI video analysis: cameras that capture the relevant area well. Your existing ones are often enough, and then no new hardware is needed.

To start, simply select a specific process and check the following conditions:

• Angle and light: The camera must see the event clearly. Counting tasks often require a direct top-down view of the counting zone. Lighting and product variations are part of the pre-start check.
• Clear event definition: The system needs a clearly observable event. Examples include "person without helmet in Zone A" or "missing part on the line".
• Approved recordings: The project team needs two to three approved recordings from ongoing operations for the test. We use these to check the camera angle and lighting.
• Access and contact person: You need access to the camera system to capture the video stream. You also need a dedicated contact person in the company.
• Works council: We coordinate workplace camera analysis with the works council in Germany before starting. Purpose, data, and access must be defined in advance.

Modifications to cameras or machines are unnecessary. The video analysis simply captures and evaluates the existing image.

What events does AI video analysis evaluate?

These examples from our project experience show what is measurable in operations and what limits apply:

Project experience: Occupational safety in the plant. The system evaluates 18 existing plant cameras for a processing facility. It detects people without helmets, people in danger zones, and open fires. Every event is reported to the plant systems with an image. A second AI checks the reports again and filters out false alarms. The system achieved a detection quality of 94.5% (mAP50 metric) for helmet detection. 93% of reported helmet cases were correct. 91% of existing helmets were found. These values were measured on test images from one location (110 images with 221 helmets, February 2025) that were not used for training. Values may vary in other environments.

Important for safety: The danger zone warning is an assistance system. It is not a safety function according to ISO 13849. Fire detection is a visual early warning supplement. It does not replace a fire alarm system according to EN 54 or VdS.

In operations: Counting portions via video. The AI counts prepared portions in a marked zone using the existing kitchen camera for a food stall. The system achieved a daily total match of around 95% in testing (35 counted, operator reported 33 to 37). It reached 79% exact matches per 30-minute segment after optimisation. Results arrive every 30 minutes via messenger bot. These values were measured at one location and compared with the operator's figures. A correctly aligned camera above the counting zone is required for this accuracy.

In operations since 2026: Parcel shops and collection points. The system evaluates cameras across multiple locations for a collection point operator. It records opening times, handovers and their duration, courier deliveries, and queues. The system delivers an hourly report per location via messenger bot. AI costs are under €2 per camera per month in operation. This is because the AI only evaluates one image every two minutes. We evaluate processes rather than individual employees for use in Germany.

How does a pilot project work in practice?

A pilot project tests reliability in everyday operations. It takes around four weeks and follows this process:

• Review view and task: We review approved recordings or the deployment location. We clarify perspective, image quality, event definition, and data handling.
• Test under real conditions: Setup takes place on your existing camera. The test considers typical operating states and difficult situations. Detected and missed events are compared with a manual evaluation.
• Plan response and operation: You receive a pilot report. This report shows detected events, false alarms, and missed cases based on the tested recordings. This helps assess the practical benefits.

The report also includes a proposal for reporting channels, responsibilities, required technology, and further testing scope. The report will clearly state if the pilot does not prove viable for your operations.

What does implementation cost?

Costs depend on the use case and the number of cameras. The pilot price is offset against the operating costs of the first twelve months. The following guide prices apply to the smallest sensible scope.

Operating costs decrease through volume pricing from 10 cameras (for example, to €45 per camera for occupational safety). They drop further from 25 cameras (to €35 per camera). New cameras, network modifications, or safety-related applications are excluded from the pilot price and require separate review.

What about data protection and co-determination?

Technical installations that can monitor employee behaviour or performance are subject to works council co-determination under Section 87(1) No. 6 of the German Works Constitution Act (BetrVG). This applies as soon as a camera could technically do so, even if that is not its purpose. We involve the works council and data protection officers before starting. We provide the technical details for your data protection impact assessment.

If we process personal data on your behalf, we sign a Data Processing Agreement (DPA) under Article 28 of the GDPR before the first data transfer. We specify which data is processed, who has access, and when it is deleted in this agreement and the Technical and Organisational Measures (TOM). We process personal data of EU customers in EU data centres. We can operate the system with a German provider or on your own servers upon request.

Read the details on security and data protection for an overview of our measures.

The next step: Free initial check

We review your process before you spend any money. The simplest first step is showing us the process in a short 30 to 60-minute video call. You can provide a short mobile video or a few photos beforehand. Alternatively, send us an image from a camera and we will check what the AI detects.

You will receive a sketch with the approach, feasibility, limits, and a fixed price for the next step after about two weeks. We also check whether a funding programme is suitable for your project. You then decide without obligation whether you want to start. Deciding against implementation is also a valid outcome. Book your free initial check here.

Sources

• BetrVG, Co-determination rights
• General Data Protection Regulation (EU) 2016/679, EUR-Lex

Written with AI assistance, checked against the information on our pages and approved by DDC’s editorial review. Not legal advice for individual cases.