Models
    Free for researchSimuletic CCTV state models · lying on the ground or floor

    Fall Detection Model

    A small detector trained only on synthetic CCTV images and measured on real footage it never saw. It catches 95.6% of people lying on the floor at 1% false alarms and 94.5% of falls in a home it never saw. The 2-second alarm caught every staged fall with no false alarms. Free to download for research and evaluation.

    Released 15 November 2026: weights, code and a notebook are public. Free for non-commercial research and evaluation. Commercial use and the Pro model need a licence.

    See it run

    On real footage it never saw

    The released Nano model on real test footage it never saw (GMDCSA-24, MIT; CAUCAFall, CC BY 4.0). Two falls raise the 2-second alarm; an everyday pick-up and a long, deep bend do not. Clips chosen by a fixed rule.
    Synthetic training images: people lying on the floor (red boxes) and look-alikes such as bending, kneeling and sitting (blue boxes), day and night infrared
    Synthetic training data: no real people, no real footage. Red = person on the floor, blue = upright look-alike.

    Get the model

    Free for research. Licensed for products.

    Commercial use

    Pro model and commercial licence

    • Trained on 10,000+ images: more night infrared, camera heights and look-alikes
    • Edge builds
    • Commercial licence and support
    • Optional fine-tuning on your own camera views
    Contact us

    Products, pilots and paid services need a commercial licence.

    Headline results

    Measured on real footage the model never saw

    The released model (round 3) in two sizes, trained only on synthetic images and scored on the same real footage. Means over training seeds, never the best seed.

    Nano (recommended)

    3.7 M params · recommended for most uses; small enough for edge boxes
    95.6%
    Lying frames caught at 1% false alarms

    Test 1, range 94.8–96.1%

    0.995
    Frame AP

    Test 1

    94.5%
    Falls caught in a different home

    Test 2, at 10% clip false alarms

    100%
    Falls caught by the 2-second alarm, no false alarms

    Test 1, all 24 falls, every seed

    Mean of 3 training seeds. The published weights are seed 0, chosen by a rule fixed before the results.

    Small

    10.2 M params · slightly higher recall in the second home
    79.8%
    Lying frames caught at 1% false alarms

    Test 1, range 74.2–85.5%

    0.975
    Frame AP

    Test 1

    96.8%
    Falls caught in a different home

    Test 2, at 10% clip false alarms

    60%
    Falls caught by the 2-second alarm, no false alarms

    Test 1; one long, deep bend still alarms

    Mean of 2 training seeds. Small still takes one long, deep bend (held for more than 4 seconds) for a fall, so prefer Nano where look-alikes matter.

    Specification

    What it is

    Task
    Detects a person lying on the floor after a fall, with a 2-second stays-down alarm, and stays quiet for bending, kneeling, sitting, standing and walking.
    Architecture
    Nano (recommended): D-FINE Nano (DETR-style detector), fine-tuned from the public COCO checkpoint. Small: D-FINE Small, fine-tuned from the public COCO checkpoint with the same recipe.
    Parameters
    Nano (recommended) 3.7 M · Small 10.2 M
    Licence
    Weights: CC BY-NC 4.0 (non-commercial research and evaluation). Code: Apache-2.0. Commercial licence on request.
    Trained on
    CCTV Fall & Incident Detection Dataset plus targeted synthetic top-ups: about 1,580 synthetic images, no real frames.
    Training recipe
    Fine-tuned from public COCO checkpoints for 60 epochs at 640 px, with the pretrained feature layers at a 10× lower learning rate than the head. Trained on about 1,580 synthetic images and no real frames: the 1,000-image CCTV Fall & Incident Detection Dataset plus targeted top-ups (people lying end-on or partly hidden, and look-alikes such as deep bends, kneeling, crawling and sitting on the floor, about half of them night infrared).
    Speed
    Nano (recommended): 16 ms per 640 px frame on an Apple M4 Pro GPU, 94 ms on its CPU (PyTorch, including pre- and post-processing). Small: 22 ms per 640 px frame on an Apple M4 Pro GPU, 132 ms on its CPU.
    Edge and camera chips
    PyTorch and ONNX for both sizes; runs on GPUs and Jetson-class edge boxes. Builds for camera chips are part of the Pro model.
    Release
    Free for research and evaluation: weights (Nano and Small, PyTorch and ONNX) on Hugging Face and Kaggle, inference code, the 2-second alarm and the scoring scripts on GitHub, and a Kaggle notebook on real footage. Commercial use needs a licence; the Pro model is trained on request.

    Real test sets

    Tested on

    CAUCAFall, subjects 6–10 (Test 1)

    1,923 frames, 615 lying; one staged room, never trained on

    CC BY 4.0 →

    GMDCSA-24 (Test 2)

    160 clips from a different home; the out-of-domain test

    MIT →

    Full table

    Every model and data combination we scored

    Round 1 (October 2026): every architecture and data combination, one training run per row unless seeds are listed. It is where the headline finding comes from: real footage from one home does not carry to the next. The released round-3 model is above.
    ModelTrain dataParamsT1 APT1 @5% FAT1 @1% FAT1 end-on @5%T2 clip APT2 falls @5%T2 falls @10%
    D-FINE Smallsynthetic10.2 M0.98597.4%83.1%90.1%0.98482.3%89.9%
    D-FINE Nano (seed 0)synthetic3.7 M0.97292.7%74.8%64.0%0.95359.5%87.3%
    D-FINE Nano (seed 1)synthetic3.7 M0.93790.2%51.7%54.0%0.96767.1%83.5%
    D-FINE Nano (seed 2)synthetic3.7 M0.97191.9%83.6%61.3%0.96165.8%83.5%
    D-FINE Nanoreal only3.7 M1.000100%100%100%0.85022.8%24.1%
    D-FINE Nanomixed3.7 M1.000100%99.5%100%0.95857.0%89.9%
    Two-stage (MobileNetV4 crop classifier)synthetic6.3 M0.89577.6%71.2%6.3%0.97884.8%88.6%
    Two-stagereal only6.3 M0.88177.1%39.2%73.9%0.91449.4%54.4%
    Two-stagemixed6.3 M0.94891.2%80.6%73.9%0.97386.1%92.4%
    RF-DETR Nanosynthetic30.2 M0.89888.3%16.1%61.3%0.95546.8%93.7%
    YOLO26nAGPL referencesynthetic2.5 M0.44919.4%19.4%1.8%0.7557.6%13.9%
    YOLO26n, feature layers frozenAGPL referencesynthetic2.5 M0.81373.0%41.8%40.5%0.88426.6%26.6%
    YOLO26nAGPL referencereal only2.5 M0.86670.9%13.7%82.9%0.74610.1%15.2%
    YOLO26nAGPL referencemixed2.5 M0.94087.8%39.4%82.9%0.80120.2%31.6%
    YOLO26sAGPL referencesynthetic10.0 M0.38012.5%11.1%0.0%0.6682.5%11.4%

    Test 1: CAUCAFall subjects 6–10 (1,923 frames, 615 lying), one staged room; in-domain for the real-only and mixed models, which trained on subjects 1–4 of the same set. Test 2: GMDCSA-24, 160 clips from a different home, the fair cross-domain test for every model. Bed and sofa clips are left out. YOLO26 is Ultralytics' AGPL-3.0 model, measured as a reference and never distributed.

    Honest limits

    Caveats and what we do not claim

    Caveats

    • Both real tests are staged, small and indoor, with few actors.
    • Small still alarms on one long, deep bend in Test 1; Nano does not.
    • Speed was measured on an Apple M4 Pro, not on edge hardware.

    Not claimed until measured

    • Public spaces, real night-time infrared CCTV, long distance or crowds: not measured yet.
    • The falling motion itself: the model reports a person lying on the ground or floor, so the alarm comes when someone is down, not mid-fall.
    • Lying in bed or on a sofa: out of scope; a care-home bed product needs its own negatives.
    • Production readiness or medical use: these are research previews measured on two staged indoor sets.

    Using it in a product?

    The free weights are for research and evaluation. For products, pilots or paid services, get a commercial licence or the Pro model: more training data, edge builds, support and fine-tuning on your camera views.