Behaviour & safetyCCTV Fall & Incident Detection
Synthetic fall detection dataset: 1,000 CCTV images of fallen people plus sitting and kneeling look-alikes, labelled person / fallen_person in YOLO and COCO.
- Size
- 1,000 images
- Format
- YOLO + COCO
Models · falls
A person falls to the floor in a real home; the synthetic-only model alarms when they are down
Real footage from a home the model never saw: the score stays at 0.03–0.09 while the person sits and falls, then 0.98–0.99 once they are on the floor.
What the clip shows
The camera is fixed at roughly chest height in a small room with a plastic chair, and the person sits, slumps and falls sideways to the tiled floor. This is one of the 160 clips in GMDCSA-24, our out-of-domain test set: a different home, different people and a different camera from anything the model was trained or validated on.
The model is the synthetic-only D-FINE Small from our lying-on-the-ground study. It never saw a real frame during training. Its score is written on every frame; the box is drawn when the score passes 0.5. While the person sits and during the fall itself the score stays below 0.1. From the moment they are on the floor it is 0.98 or higher for the rest of the clip.
That is the state the dataset teaches: lying on the ground, not the motion of falling. On the full GMDCSA-24 set this model catches 82% of fall clips at a 5% clip false-alarm rate and 90% at 10%; the full table, the misses and the caveats are in the case study.
- Footage
- GMDCSA-24 (Alam et al.), MIT licence; real home, fixed camera at chest height
- Event
- Sitting on a chair, falling sideways to the floor, lying still
- Model
- D-FINE Small, 10.2 M parameters, trained on 1,000 synthetic images and no real frames
- Score
- 0.03–0.09 before the fall; 0.98–0.99 once on the floor (88 of 178 frames alarm)
- Clip length
- 6 seconds, 30 fps (re-encoded from 60 fps)
- Resolution
- 1280 x 720
Dataset
The dataset this clip belongs to
Behaviour & safetyMore clips
Other sample videos
CCTV handgun detection: a labelled night-time clip
A person walks along a lit shop front at night with a handgun in the right hand, followed frame by frame by a red weapon box.
CCTV knife detection: a kitchen knife on a night street
A person turns and walks past a shop door at night with a large kitchen knife held low in the right hand.
Drone detection of armed people among civilians
From a moving drone, three civilians walk a dirt track and a group of armed figures enters from the left.
Need this event on your own cameras?
Send 5–20 photos of your empty scenes and we render the event into them, as images or video clips with boxes, tracking IDs and event times.