What is a CCTV video anomaly detection dataset?

A CCTV video anomaly detection dataset is a collection of surveillance-style video clips labeled with normal behavior versus abnormal behavior — falls, fights, weapon draws, loitering, abandoned objects. Models trained on these clips learn to flag events in real-time surveillance feeds.

What behaviors are usually included?

The common anomaly classes are: falls and medical collapses, physical fights and aggression, weapon presentation, shoplifting and concealment, ATM robbery, loitering, abandoned bags, unauthorized entry, and smoke or fire onset. Serious datasets label each with a temporal window, not just a clip-level tag.

Why synthetic CCTV video instead of real footage?

Real surveillance footage of anomalies is rare, legally restricted, and expensive to label. Synthetic CCTV lets you generate any anomaly at any camera angle, height, resolution, and lighting condition, with per-frame skeletons and bounding boxes attached automatically. It also removes GDPR and biometric-training exposure entirely.

What annotation formats are used?

Simuletic ships surveillance video datasets with YOLO bounding boxes per frame, optional 17-keypoint COCO skeletons for pose, and clip-level action labels. Datasets can be delivered as image sequences, MP4 clips, or both.

CCTV Video Anomaly Detection Datasets

What CCTV anomaly detection actually is

A CCTV video anomaly detection model watches a surveillance feed and flags anything that isn't ordinary: a fall, a fight, a weapon coming out, a person loitering, a bag left behind. The dataset is what teaches the model where the line between normal and abnormal sits.

The clip below is an example of exactly the kind of anomaly a CCTV model has to catch — a sudden collapse from cardiac arrest. This one is 100% synthetic, generated in our pipeline with per-frame skeleton and bounding-box annotations.

Example anomaly: sudden collapse. Synthetic clip with per-frame labels — no real CCTV, no PII, no consent overhead.

The anomaly classes that matter

Real deployments almost always care about the same short list of behaviors:

Falls and medical collapses — eldercare, hospitals, transit platforms.
Fights and aggression — bars, transit, schools, custody.
Weapon draws — retail, banking, transport hubs.
Shoplifting and concealment — retail loss prevention.
ATM robbery and hold-up — banking and self-service.
Loitering and abandoned objects — transit, public venues.
Smoke and fire onset — indoor CCTV, warehouses, wildfire perimeter.

Why real CCTV data is a dead end

Building an anomaly dataset from real surveillance footage means three things nobody wants: waiting for events to happen on your cameras, paying a labeler by the frame, and defending the whole thing against GDPR and the EU AI Act's biometric-training rules. It's why every open surveillance dataset is small, biased toward a handful of camera setups, and years out of date on modern hardware.

Synthetic CCTV inverts every one of those problems. You render the exact behavior at the exact camera angle, height, resolution, and lighting your deployment uses. Labels are attached automatically because the pipeline knows the character's skeleton and the action class.

Related datasets you can download today

Every dataset below ships as synthetic CCTV video and images, YOLO-annotated, GDPR-safe.

Fall & Incident Detection Dataset

Aggressive Pose Detection Dataset

Shoplifting Detection Dataset

ATM Robbery Detection Dataset

CCTV Weapon Detection Dataset

Smoke & Fire Detection Dataset

Also on Simuletic

Need a specific anomaly class we don't have yet?

Send us the behavior, camera type, and label format. We generate custom synthetic CCTV video with per-frame annotations, usually in days.

Talk to us about a custom dataset Browse all datasets Video datasets guide