CCTV and surveillance datasets: public sources, licences and gaps (2026)
Seventeen public CCTV, surveillance and person datasets side by side: what each one contains, which camera it was shot from, how it is labelled, what its licence allows, and where the gaps are for a commercial product.
What counts as CCTV data
CCTV data is footage from fixed, usually elevated cameras that watch a scene continuously, as opposed to handheld, dashcam or film footage. The viewpoint matters more than the word "surveillance" in a dataset's title: a detector trained on people seen at eye level from a car learns different shapes, scales and occlusions than one trained on people seen from three metres up a wall.
Three properties separate real CCTV from other video. The camera does not move, so the background is learned and the foreground is small. The angle is steep, so heads and shoulders hide hands and objects. And the image is degraded: compression, low bit rates, infrared at night, motion blur and wide-angle distortion. A useful CCTV dataset has to reproduce all three. Several sets people search for under "CCTV dataset" do not, which is why the table below has a camera column and why AVA and Charades are listed as "not CCTV".
17 public datasets compared
Sizes and licences are taken from each dataset's own page or paper (sources at the end). Where a page states no licence, the table says so; no licence is not the same as permission.
| Task | Dataset | Size | Camera | Labels | Licence | Caveat |
|---|---|---|---|---|---|---|
| Anomaly detection | UCF-Crime | 1,900 untrimmed videos, ~128 h, 13 anomaly types | Real surveillance clips collected from YouTube and LiveLeak | Normal / anomaly and anomaly type per video | None stated on the project page | Web-sourced footage, 320×240; rights in the clips are unclear |
| Anomaly detection | ShanghaiTech Campus | 437 videos (330 train, 107 test), 13 scenes, 130 abnormal events | Fixed campus cameras | Normal-only training; abnormal events marked in test | None found; the original project page is offline | One university campus |
| Anomaly detection | CUHK Avenue | 16 train + 21 test clips, 30,652 frames, 47 abnormal events | One fixed camera on a campus avenue | Abnormal events marked with rectangles | None stated on the project page | Single scene; mostly staged events (running, throwing, wrong direction) |
| Anomaly detection (synthetic) | UBnormal | 543 clips, 236,902 frames, 29 virtual scenes | Rendered in Cinema4D on real background photos | Pixel-level anomaly annotations; open-set train/test split | CC BY-NC-ND 4.0 | Non-commercial and no derivatives |
| Violence detection | XD-Violence | 4,754 untrimmed videos, 217 h, 6 violence classes | Multi-scene, with audio | Weak (video-level) labels | None stated on the project page | Weak labels only; confirm terms with the authors |
| Fight detection | RWF-2000 | 2,000 clips of 5 s | Surveillance cameras, clips sourced from YouTube | Fight / non-fight per clip | No commercial use or redistribution without the lab's approval | Video files currently withheld on the site for privacy reasons |
| Events and objects | VIRAT Ground 2.0 | ~8.5 h HD video, 11 outdoor scenes | Stationary ground cameras, mostly mounted high on buildings | 12 annotated event types | Signed VIRAT protection agreement, per person | No sharing with anyone who has not signed |
| Multi-camera tracking / Re-ID | DukeMTMC | Withdrawn in 2019 | Campus cameras at Duke University | Person tracks and identities | Withdrawn by its creators | Do not use copies or derived sets such as DukeMTMC-ReID |
| Multi-object tracking | MOT17 | 14 sequences (7 train, 7 test), each with 3 detector sets | Pedestrian scenes, several cameras | Person boxes with track IDs | CC BY-NC-SA 3.0 | Some sequences carry extra third-party conditions |
| Multi-object tracking | MOT20 | 8 very crowded sequences (4 train, 4 test) | Includes an elevated night view of a stadium exit | Person boxes with track IDs | CC BY-NC-SA 3.0 | Crowds, not typical retail or corridor CCTV |
| Pedestrian detection | CityPersons | Person boxes on the 5,000 finely annotated Cityscapes images | Street-level driving footage | Pedestrian bounding boxes | Cityscapes terms: non-commercial use | Car viewpoint, not an overhead camera |
| Pedestrian detection | Caltech Pedestrian | ~10 h of 640×480 video, 250k frames, 350k boxes, ~2,300 people | Driving through urban traffic | Boxes, occlusion labels, temporal links | Check the Caltech page before use | Car viewpoint, low resolution by 2026 standards |
| Pedestrian attributes | PA-100K | 100,000 images, 598 scenes, 26 attributes | Outdoor surveillance cameras | Attribute labels (gender, age, bags, phone…) | CC BY 4.0 | Real, identifiable people: GDPR still applies to you |
| Person Re-ID | Market-1501 | 32,668 boxes of 1,501 identities, 6 cameras | Cameras outside a campus supermarket | Identity per box | README: research only, no distribution, no commercial use | Re-ID tracks real individuals across cameras |
| Shoplifting / anomaly clips | DCSASS (Kaggle) | 16,853 short clips (9,676 normal, 7,177 abnormal), 13 classes | Cut from UCF-Crime | Normal / abnormal per clip | CC BY-NC-SA 4.0 on Kaggle | Inherits UCF-Crime's sourcing questions |
| Action detection (not CCTV) | AVA Actions v2.2 | 430 videos, 15 min each annotated at 1 s | Feature films | Spatio-temporal atomic actions | CC BY 4.0 (dataset); source videos are on YouTube | Cinematic angles, nothing like a ceiling camera |
| Activity recognition (not CCTV) | Charades | 9,848 videos of ~30 s, 157 action classes | Crowd workers filming themselves at home | Action classes with temporal intervals | Non-commercial use licence | Handheld, eye-level home video |
How to read the table by task
- Anomaly detection: UCF-Crime is the only large set of real incidents, but labels are per video and the clips came from video-sharing sites. ShanghaiTech and Avenue are small, single-site and mostly staged. More detail in our CCTV video anomaly detection guide.
- Fights and violence: RWF-2000 is genuinely shot on surveillance cameras, but it bars commercial use without approval and its videos are currently withheld.
- CCTV person detection and tracking: MOT17/MOT20 and VIRAT Ground are the closest public sources; CityPersons and Caltech are driving datasets.
- Attributes and Re-ID: PA-100K is the rare permissive licence. Market-1501 is research-only and, like every Re-ID set, exists to follow real individuals between cameras.
- Action recognition: AVA and Charades are excellent action benchmarks, just not from security cameras. Our human activity recognition guide covers them.
Why most CCTV datasets are academic or non-commercial
A non-commercial dataset is one whose licence allows research and teaching but not use in a product you sell. Most CCTV sets end up there for three reasons.
- The footage was never theirs to licence. UCF-Crime's clips were gathered from YouTube and LiveLeak; RWF-2000's from YouTube. A university can share annotations, but rights in the underlying video stay with whoever filmed or uploaded it.
- The people in it did not consent to a product. Campus recordings were approved, if at all, by an ethics board for a specific study. Commercial reuse falls outside that approval.
- Copyleft and NC clauses travel. MOT17 and MOT20 are CC BY-NC-SA 3.0; DCSASS is CC BY-NC-SA 4.0; UBnormal is CC BY-NC-ND 4.0. Creative Commons' own guidance says the conservative reading of NC covers every stage, from copying during training to sharing the trained model.
Privacy, GDPR and withdrawn datasets
Under the GDPR, video of an identifiable person is personal data, and the European Data Protection Board's Guidelines 3/2019 apply that to video devices in general. A permissive copyright licence such as PA-100K's CC BY 4.0 does not change that: the licence covers copyright, not your legal basis for processing images of the people in it.
DukeMTMC shows how quickly a dataset can disappear. It was recorded on Duke University's campus and became a standard multi-camera tracking and Re-ID benchmark. In 2019, after reporting by Exposing.ai and the Financial Times, a Duke review found it was "neither collected nor made available to the public consistent with the terms of the study" approved by the IRB, and the creators took it down. Derived sets such as DukeMTMC-ReID kept circulating, and a Princeton study later found withdrawn datasets still in wide use. If a model in your product was trained on a withdrawn set, you inherit that history.
The EU AI Act adds a hard line: Article 5(1)(e) prohibits AI systems that build or expand facial recognition databases "through the untargeted scraping of facial images from the internet or CCTV footage". Detecting people or objects is not facial recognition, but any pipeline that harvests faces from CCTV is now out of bounds in the EU.
Synthetic CCTV data as the commercial route
Synthetic CCTV data is footage rendered from virtual cameras placed like real security cameras, with labels written by the renderer instead of by people. UBnormal showed academically that rendered scenes work for anomaly benchmarks; its NC-ND licence just keeps it out of products.
For a product team, the useful properties are practical. Nobody real appears in the frames, so there is no consent or GDPR question about the training set. The licence can be commercial because the vendor owns every pixel. Rare classes (a knife in a hand, a fall at night, a fight under IR) can be generated in volume, and hard negatives, the umbrellas and phones that cause false alarms, are generated on purpose rather than hoped for. What synthetic data does not remove is the need to validate on real footage from your own cameras before you ship; see our synthetic vs real data comparison.
From task to dataset
Each Simuletic dataset below is synthetic, rendered from CCTV viewpoints, and licensed for research or commercial use. Image datasets ship with YOLO and COCO labels and a fixed train/val split.
| Task | Public gap | Simuletic dataset |
|---|---|---|
| Weapon detection (handgun, rifle, shotgun) | No licensable public CCTV weapon set in the table above | CCTV Weapon Detection Dataset |
| Weapon look-alikes (umbrellas, chip bags) | Hard negatives are missing from public sets | Rifles vs Umbrellas, Handgun vs Bag of Chips |
| Knife detection | None in the table above | CCTV Knife Detection Dataset |
| Fall detection | No fall class in the CCTV sets above | CCTV Fall & Incident Detection Dataset |
| Fight and aggression | RWF-2000, XD-Violence: clip-level labels, no commercial terms | CCTV Fight & Aggression Detection Dataset (boxes + COCO-17 pose) |
| Smoke and fire on CCTV | Arson/explosion clips in UCF-Crime and XD-Violence only | CCTV Smoke & Fire Emergency Detection Dataset |
| Shoplifting | UCF-Crime and DCSASS clips: video-level labels, non-commercial or unclear | CCTV Shoplifting Detection Dataset (video, pose, VLM captions) |
| ATM robbery | Robbery clips in UCF-Crime only | CCTV ATM Robbery Detection Dataset |
| Pedestrian attributes | PA-100K is CC BY 4.0 but shows real people | Pedestrian Attribute Recognition Dataset (Pedestrian-1K) |
FAQ
What is a CCTV dataset?
A CCTV dataset is a set of images or video from fixed, usually elevated security cameras, labelled for a task such as person detection, tracking, anomaly detection or event recognition.
Is there a free CCTV footage dataset I can use commercially?
Rarely. Of the 17 public sets compared here, most are non-commercial (CC BY-NC-SA, CC BY-NC-ND, research-only) or state no licence at all; PA-100K is CC BY 4.0, and VIRAT ground video comes under a signed agreement. Even with a permissive licence, footage of real people is personal data under GDPR.
Which public dataset is best for CCTV person detection?
For people seen from elevated cameras, MOT17 and MOT20 (person boxes with track IDs, CC BY-NC-SA 3.0) and VIRAT Ground (high-mounted outdoor cameras) are closest; CityPersons and Caltech Pedestrian are filmed from a car, so the viewpoint differs from a ceiling or pole camera.
Why was DukeMTMC taken down?
Duke University found the campus recordings were not collected or released in line with the study's approved IRB terms, and the creators removed the dataset in 2019; derived copies such as DukeMTMC-ReID still circulate but should not be used.
How do companies get commercial CCTV training data?
They record and label their own footage with a legal basis under GDPR, buy a commercially licensed dataset, or use synthetic CCTV data, which contains no real people and can be licensed for products.
Sources
- UCF-Crime project page (CRCV, University of Central Florida)
- ShanghaiTech Campus figures (Papers with Code)
- CUHK Avenue dataset page
- UBnormal repository and licence
- UBnormal paper (CVPR 2022)
- XD-Violence project page
- RWF-2000 repository and licence
- VIRAT Video Dataset
- OD-VIRAT paper (VIRAT Ground camera description)
- DukeMTMC investigation (Exposing.ai)
- Princeton CITP on withdrawn datasets
- MOT17 (MOTChallenge)
- MOT20 (MOTChallenge)
- CityPersons paper
- Cityscapes dataset and terms
- Caltech Pedestrian benchmark paper
- PA-100K / HydraPlus-Net repository (CC BY 4.0)
- Market-1501 paper (ICCV 2015)
- DCSASS dataset on Kaggle
- AVA download page (Google)
- Charades paper (Hollywood in Homes)
- Charades data (Allen Institute for AI)
- EDPB Guidelines 3/2019 on video devices
- EU AI Act, Article 5 (prohibited practices)
Dataset sizes and licences checked on October 24, 2026. Licences change; confirm on the dataset's own page before use.
Related guides
Need CCTV data you can ship with?
Every catalogue dataset has a free sample and a commercial licence. If your cameras or classes are not covered, we build a custom set on your own camera views.