Simuletic/fall-detection-dataset

Fall Detection Dataset (CCTV, YOLO pose + boxes): free 80-image sample

This fall detection dataset is a free 80-image sample of synthetic CCTV images of people lying on the ground after a fall, with YOLO pose labels (bounding box plus 17 body keypoints) for laying and standing, for training and testing fall detection and pose models on security-camera footage. It is a subset of Simuletic's full free 111-image sample, CCTV_Incident_Dataset_Fall_Lying_Down_Detection, which has the same images and labels plus 31 more. The full paid 1,000-image dataset uses person and fallen_person boxes in YOLO and COCO, with about 20% hard negatives (sitting, crouching, squatting, kneeling) and a 900 / 100 train-val split; a model trained on it alone catches 97% of lying frames on real footage (see below). Made by Simuletic, a Swedish company that has built synthetic CCTV datasets for computer vision since 2023.

Get the full dataset · Free sample licence: CC BY 4.0.

Facts

	Free sample (this repo)	Full dataset
Images	80 PNG (subset of the 111-image sample)	1,000 (900 train / 100 val)
Classes	laying (fallen), standing	person, fallen_person
Boxes	80 laying, 6 standing, each with 17 keypoints	every person boxed; fallen people as their own class
Hard negatives	none	about 20%: sitting, leaning, crouching, squatting, kneeling
Resolution	1,024 × 1,024 px	CCTV frames; day and night
Camera view	elevated CCTV angles, many wide-angle; indoor and outdoor	ceiling and wall mounts; corridors, lobbies, shop aisles, plant rooms, outdoors
Labels	YOLO pose: box + 17 COCO keypoints	YOLO + COCO boxes, dataset card
Licence	CC BY 4.0	one-time: research $149 / commercial $590 per 1,000 images

What's in the images

People lying on the ground after a fall, seen from elevated CCTV cameras: in shop aisles, on streets and pavements, and in other indoor and outdoor places. Many views are wide-angle. Every image has one person lying down (laying); 6 images also show a person standing or walking in the background (standing). Each person has a bounding box and a 17-point body skeleton, so you can train either a detector or a pose model that tells lying from standing.

Classes and boxes

Class	YOLO id	Boxes in sample
laying	0	80
standing	1	6

Label format and folder layout

laying_dataset/
  dataset.yaml             # class names and keypoint shape
  example_image.png        # preview image (not labelled)
  images/                  # 80 PNG images (laying01.png, ...)
  labels/                  # one YOLO pose .txt per image

Each label line is one person: class_id x_center y_center width height, then 17 keypoints as x y v (56 values in total). Coordinates are normalised to 0–1. The keypoints follow the COCO order (nose, eyes, ears, shoulders, elbows, wrists, hips, knees, ankles); v is 2 for a labelled keypoint and 0 for one that is not labelled. Shortened example:

0 0.579 0.658 0.346 0.398  0.451 0.477 2  0.452 0.469 2  ...  0.491 0.813 2

Quick start (Ultralytics YOLO)

Create a data.yaml next to the laying_dataset/ folder:

path: laying_dataset
train: images
val: images   # small sample: use the same images, or split them yourself
kpt_shape: [17, 3]
flip_idx: [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]
names:
  0: laying
  1: standing

The included dataset.yaml has path: /; use the folder path as above instead.

Then train:

from ultralytics import YOLO

model = YOLO("yolo11n-pose.pt")
model.train(data="data.yaml", epochs=30, imgsz=640)

Use cases

• Fall detection in care homes, hospitals and assisted living
• Slip-and-fall detection in shops, stations and workplaces
• Spotting people lying on the ground (medical emergencies) on public-space cameras
• Pose-based posture classification: lying versus standing
• Pre-training before fine-tuning on your own camera footage

Measured on real footage

We trained D-FINE Small and Nano on the full 1,000-image version of this dataset and nothing else, then scored them on public real-footage test sets the models never saw. The synthetic-only Small model catches 97% of people lying on the floor at 5% false alarms on real CAUCAFall footage (CC BY 4.0) and 83–87% of falls in a second home (GMDCSA-24, MIT), where a model trained on real footage from one room catches 24%. Full table, method and misses: case study · fall detection model · all models.

Full dataset and prices

The full Fall Detection Dataset has 1,000 images, YOLO + COCO, classes person, fallen_person. Every licence is a one-time payment, perpetual, for one organisation; you get a download link within five working days (card or invoice). Research covers students and non-commercial work and is available for 1,000 images only; packs of 5,000 images and more are sold with the commercial licence, which lets you train models and ship them in products.

	Research	Commercial
1,000 images	$149	$590
5,000-image pack	not offered	$2,400
10,000-image pack	not offered	$4,500
20,000-image pack	not offered	$8,500

Packs are the same dataset at larger size, with the same labels and split, under the commercial licence. 50,000+ images, video datasets and enterprise agreements (from $10,000) are quoted. Other scenes, camera angles or classes: custom data or contact@simuletic.com.

How the images were made

The images are fully synthetic (computer-generated) by Simuletic: no real people, places or incidents are depicted, so there are no privacy or consent issues.

Limitations. Synthetic images do not cover every camera, lens, compression artefact or scene you will meet in deployment. Validate on your own footage, and for production combine synthetic and real data. This sample uses pose labels with the classes laying and standing and has no hard negatives; the full dataset uses person and fallen_person boxes and adds hard negatives such as people sitting, crouching or kneeling.

More synthetic datasets from Simuletic

• CCTV Fall & Incident Detection Dataset
• CCTV Fight & Aggression Detection Dataset
• CCTV Pedestrian-1K: Person Attribute Dataset
• CCTV Smoke & Fire Emergency Detection Dataset
• CCTV Knife Detection Dataset
• All datasets: simuletic.com/datasets · Hugging Face · Kaggle · GitHub

Licence

This free sample is licensed under CC BY 4.0: you may share and adapt it for any purpose, including commercial use, as long as you credit Simuletic. The full dataset is sold under a separate one-time research or commercial licence; it is not redistributed.

Citation

@misc{simuletic_fall_split_2026,
  author       = {{Simuletic}},
  title        = {Fall Detection Dataset},
  year         = {2026},
  publisher    = {Simuletic},
  howpublished = {\url{https://huggingface.co/datasets/Simuletic/fall-detection-dataset}},
  note         = {Free sample. Full dataset: \url{https://simuletic.com/datasets/cctv-fall-incident-detection-dataset}}
}

Questions or feedback: contact@simuletic.com