Training data for AI security cameras: a buyer's checklist

Training data for an AI security camera is the set of labelled images or clips a detector learns from, and it decides what the camera will and will not alarm on. Most failures in the field (missed knives, umbrellas flagged as rifles, a sitting person reported as a fall) trace back to the data, not the model architecture. Use the checklist below whether you buy a dataset, label your own footage or generate synthetic data.

1. What the detector must see

The class list is the set of things the camera must find; the hard negatives are the things it must learn to ignore. Both need to be in the training set on purpose.

• Write down every alarm class and the exact rule for when a box is drawn (a holstered gun? a knife on a kitchen counter?).
• Include a person class. It anchors the scene and lets you link an object to whoever is holding it.
• List the look-alikes for each class: phones, wallets and bottles for knives; umbrellas and tools for rifles; crouching, kneeling and sitting for falls; hugs and dancing for fights; steam, fog and reflections for smoke.
• Ask the vendor what share of images are hard negatives. For reference, our catalogue sets carry about 16% (smoke and fire), 20% (weapons, knives, falls) and 28% (fights).
• Check small objects. A knife seen from an overhead camera covers a tiny part of the frame; the dataset must include objects at that scale.

2. Label formats

A label format is the file layout your training code reads the boxes, classes and keypoints from; getting it wrong costs days of conversion.

• YOLO: one .txt per image, one row per object as class x_center y_center width height, normalised to 0–1, classes zero-indexed.
• COCO: one JSON with images, annotations and categories; boxes are [x, y, width, height] in pixels from the top-left corner.
• Pose: for falls and fights, ask for COCO-17 keypoints (x, y, visibility). The YOLO-pose format appends keypoints to each box row.
• Video: for temporal models, ask for per-frame boxes, track IDs and event start/end times, not just a label per clip.
• A fixed train/val split, so results stay comparable across vendors and retrains.

3. Camera geometry

Camera geometry is where the camera sits and what its optics do to the image: mounting height, tilt, lens and sensor mode.

• Mounting height and tilt match yours: ceiling, wall, pole or ATM fascia. A model trained at eye level degrades at a steep down angle.
• Lens: wide-angle and fisheye distortion at the frame edges, where many incidents happen.
• IR and night: monochrome infrared frames, headlight glare, low-light noise.
• Degradation: compression artefacts, motion blur, low bit rate. Clean studio images teach the wrong texture.
• Scenes: the environments you sell into (retail, transit, banking, care homes, warehouses), not only streets.

4. Licence terms to check

A dataset licence is the contract that says what you may do with the data and with the models trained on it. Read it before you train, not before you ship.

• Commercial use: is it explicitly allowed? CC BY-NC, CC BY-NC-SA and "research only" are not.
• Derived weights: can you ship and redistribute the trained model, on devices and in the cloud?
• Raw data: what you may not do with the files themselves (resell, republish, share with partners).
• Seats and scope: per user, per project or per organisation? Perpetual or annual?
• Provenance: who owns the footage, and did the people in it agree? For real CCTV, the GDPR applies to identifiable people regardless of the copyright licence.
• The model licence too: Ultralytics YOLO is AGPL-3.0 unless you buy an Enterprise licence. See YOLO licence vs dataset licence.

5. Validation

Validation is measuring the trained model on data it has never seen, from the cameras it will run on, before customers do it for you.

• Keep a held-out set of real footage from your own sites, never used for training or tuning.
• Report precision and recall per class, and mAP if you compare architectures.
• Measure the false-positive rate on long stretches of normal footage. A camera that alarms on umbrellas is switched off, however good its recall.
• Break results down by condition: day/night, near/far, occluded/visible.
• Run a dedicated hard-negative test before launch. Our False-Positive Stress Test runs your detector against our hard-negative and edge-case set and returns every failure, ranked.

6. EU AI Act notes

The EU AI Act classifies some AI systems as high-risk, and high-risk systems must meet data-governance rules for their training, validation and test sets.

• Is your camera high-risk? Annex III lists, among others, remote biometric identification, biometric categorisation by sensitive or protected attributes, emotion recognition, and safety components of critical infrastructure. A weapon, fire or fall detector is not listed as such; how it is used decides. Take legal advice on your specific deployment.
• New dates. The AI Omnibus, Regulation (EU) 2026/1744, published on 24 July 2026, moved the Annex III high-risk obligations from 2 August 2026 to 2 December 2027, and high-risk AI in products under EU product-safety law to 2 August 2028.
• What Article 10 asks of data. Training, validation and test sets must be "relevant, sufficiently representative, and to the best extent possible, free of errors and complete", examined for bias, and documented: origin, collection, labelling and cleaning.
• Already in force. Article 5 prohibitions apply now, including building facial recognition databases through untargeted scraping of facial images from CCTV footage.

A dataset with a dataset card (classes, counts, generation or collection method, split) makes the Article 10 documentation much easier, whatever its source.

7. Build, buy or generate

There are three ways to get the data; most teams end up combining at least two.

8. Budget

These are Simuletic's public prices, in USD, excluding VAT, one-time. They give you a floor for planning; your own collection and labelling costs come on top.

Full details on pricing, the dataset catalogue and custom data. To see which public datasets exist for each task, and what their licences allow, read CCTV and surveillance datasets compared.

FAQ

Sources

• Ultralytics: object detection dataset format
• Ultralytics: pose estimation dataset format
• AutoGluon: COCO annotation format
• EU AI Act, Annex III (high-risk areas)
• EU AI Act, Article 10 (data and data governance)
• EU AI Act, Article 5 (prohibited practices)
• Mondaq: EU Digital Omnibus on AI enters into force (Regulation (EU) 2026/1744)
• Gibson Dunn: EU AI Act Omnibus agreement, postponed high-risk deadlines
• Cloud Security Alliance: high-risk deadline pushed to December 2027
• EDPB Guidelines 3/2019 on video devices
• Ultralytics licensing (AGPL-3.0 and Enterprise)

Checked on October 24, 2026. This guide is not legal advice.