Back to Datasets

    CCTV Fall & Incident Detection Dataset

    Sample video clip from the dataset — synthetic retail CCTV with person + pose detection.

    CCTV Fall & Incident Detection Dataset sample 1

    This fall and incident detection dataset is built for training pose-estimation and object-detection models that monitor vulnerable populations in hospitals, care homes, and public spaces. Each image is annotated with both YOLO bounding boxes and 17-keypoint skeletons following the COCO standard, providing rich posture information. The dataset distinguishes between 'standing' and 'fallen' states, covering scenarios like slips on wet floors, staircase falls, and collapses in corridors. Perspectives are captured from overhead and angled CCTV cameras at 2–4 meter height. The full package includes 1,200 images and evaluation videos to benchmark real-time fall detection accuracy. Compatible with YOLOv8-Pose and YOLO11-Pose architectures.

    1,200 images

    Full Package

    113

    Open Source Samples

    YOLO + COCO Keypoints

    Annotation Format

    100%

    Privacy Compliant

    Dataset Features

    Dual annotations: bounding boxes + 17-keypoint skeletons
    COCO standard keypoint format
    Standing vs fallen state classification
    Overhead and angled CCTV perspectives
    YOLOv8-Pose and YOLO11-Pose compatible
    113 open-source sample images

    Intended Use Cases

    Elderly care monitoringHospital patient safetyPublic space incident detectionWorkplace safety compliance

    Free Sample vs. Commercial Package

    Free Open-Source Sample

    • 113 annotated images
    • Format: YOLO + COCO Keypoints
    • Hosted on Kaggle
    • Licence: See the hosting platform's terms of use

    Commercial Package

    The Creative Commons licence above applies only to the free sample, not to the full commercial package.

    Limitations & Recommended Validation

    This dataset is 100% synthetic. While it is designed to closely match real-world sensor and camera conditions, synthetic imagery can still differ from live footage in ways that affect model accuracy (a "domain gap"). Validate a trained model against real-world footage from your specific deployment environment before production use.

    Not intended as a sole basis for biometric identification, legal evidence, or safety-critical decisions without independent human review and real-world testing.