Fall Detection Datasets: Public Options and What's Missing

What a fall detection dataset is

A fall detection dataset is a set of recordings of people falling and of normal activities of daily living (ADL), labelled so a model can learn to tell them apart.

The recordings come from one of three sensor families: RGB or depth cameras, wearable inertial sensors, or thermal and other ambient sensors. A camera-based product needs camera data, and ideally from the same viewpoint it will be installed at. That is where the public options run thin.

The public datasets compared

The table covers the vision datasets most cited in fall detection papers, plus two wearable datasets for contrast. Figures come from the papers and pages listed under Sources.

Notes on each

• UR Fall Detection pairs RGB and depth from two Kinects with accelerometer data. It is the only common set with a ceiling-mounted camera, which makes it a useful sanity check for overhead views, but 30 falls is too few to train on.
• Le2i was recorded in four different rooms precisely so that methods could be tested on a location they had not seen. Reviews disagree on its exact ADL count, so recount the files you download.
• UP-Fall is the most complete multimodal set: two cameras, five wearables, an EEG headset and infrared sensors, recorded at the same time. Its subjects are 17 healthy adults aged 18 to 24 falling onto a mattress, which is far from the population most products are built for.
• Multiple Cameras Fall records each scenario from eight synchronised viewpoints, which suits multi-view and 3D methods. Two of its scenarios contain only confounding events and no fall, which is useful for measuring false alarms.
• SisFall and FallAllD are wearable datasets. SisFall is notable for including participants over 62 for daily activities; FallAllD records from three body positions. Neither contains images.
• Thermal Simulated Fall was recorded with a FLIR ONE thermal camera, chosen partly because thermal images do not reveal identity. It has 35 fall segments in one room.

Image vs video

An image dataset teaches a model what a person on the floor looks like; a video dataset teaches it how the person got there.

All the camera datasets above are video, and most are small. Because each sequence contains one fall, the number of independent fall examples is in the tens or low hundreds, however many frames there are. Extracting frames inflates the count but not the variety: the same actor, room and lighting repeat.

A practical architecture splits the job. A detector trained on images finds people and classifies each as standing, sitting or fallen. A light temporal stage then checks that the person went from upright to fallen quickly and stayed down, which removes most false alarms from people lying down on purpose. The image stage needs variety (people, rooms, angles, lighting, clothing, poses on the floor); the temporal stage needs fewer, well-labelled clips.

CCTV height vs lab setups

CCTV height means a camera mounted near the ceiling looking down at a steep angle, which changes how a fallen person looks compared with a camera at chest height.

From the side, a fall is a tall shape becoming a wide one, and many early methods used exactly that aspect-ratio change. From above, a standing person is already foreshortened, and a person lying along the camera's line of sight can look short and compact. Most public cameras are at the side: UP-Fall's two cameras are at 1.82 m, Le2i uses a single camera per scene, and only one of UR Fall's two Kinects is on the ceiling.

Lab setups add other biases. Falls are performed by young volunteers onto mats, in a few rooms, in good light. Real falls in care settings happen at night, next to beds and chairs, partly hidden by furniture, and to people who move differently from a 20-year-old actor. Test your model on footage from the camera height and rooms you will deploy in before trusting a public benchmark score.

Privacy limits in care homes

Video of residents in care homes and patients in hospitals is health-related personal data, and under the GDPR that is special-category data.

That is why no large dataset of real falls on care-home CCTV is public. Residents may not be able to consent, operators cannot share footage outside the care relationship, and recordings of real falls are both rare and distressing. The privacy constraint also affects deployment: many sites prefer thermal or depth sensors, or on-device processing that never stores video. Whatever you train on, plan to validate on the sensor and site you will actually run on, with the site's agreement.

Our CCTV Fall & Incident Detection Dataset

The CCTV Fall & Incident Detection Dataset targets the image stage described above, from the viewpoint the public sets lack.

• 1,000 synthetic images from ceiling and wall-mounted CCTV positions, by day and at night.
• People lying on their back, side or front after a fall, in corridors, lobbies, shop aisles, plant rooms and outdoor areas.
• About a fifth hard negatives: people sitting, leaning, crouching, squatting or kneeling.
• Classes person and fallen_person, in YOLO and COCO format, with a fixed 900 / 100 train-val split and a dataset card.
• Synthetic, so no patient, resident or worker is filmed.

It is sold with a Research licence ($149) or a Commercial licence ($590); see pricing. Fall video is available on request: through custom data we make clips in your own fixed camera views with per-frame boxes, tracking IDs and event start and end times, quoted per project.

Frequently asked questions

Also on Simuletic

• Fall detection dataset write-up
• CCTV video anomaly detection datasets
• Human activity recognition datasets

Sources

1. UR Fall Detection Dataset page and licence (University of Rzeszów)
2. Martínez-Villaseñor et al., UP-Fall Detection Dataset: A Multimodal Approach, Sensors 2019
3. UP-Fall (HAR-UP) project site
4. Multiple Cameras Fall Dataset (Auvinet et al., DIRO, Université de Montréal)
5. Vision-based Human Fall Detection Systems using Deep Learning: A Review (Le2i, MCFD, TSF descriptions)
6. Sucerquia et al., SisFall: A Fall and Movement Dataset, Sensors 2017
7. FallAllD on IEEE DataPort (Saleh, Abbas, Le Bouquin Jeannès)
8. Vadivelu et al., Thermal imaging based elderly fall detection, ACCV 2016 Workshops