YOLOFall

YOLOFall is a small project that detects people, extracts pose keypoints, tracks individuals across frames, and applies a simple time-based heuristic to detect falls.

Key features

• Person detection and pose: uses an ultralytics YOLOv8 pose model (yolov8*-pose.pt) to get bounding boxes and keypoints.
• Multi-object tracking: DeepSORT (deep_sort_realtime) keeps persistent IDs for people so per-person histories can be tracked over time.
• Simple fall detection: a time-windowed heuristic marks a person as FALLEN if they go from standing to laying within a short time window.
• Reattachment: a small cache attempts to reattach recently lost track IDs to new detections using IoU + recency so short occlusions don't break a person's history.

Folder structure

yolofall/
├── data/
│   ├── fall_*.mp4      // sample videos for testing
│   └── ...
├── output/             // saved output videos
│   ├── ...
├── requirements.txt
└── main.py

Usage

1. Create a Python environment and install dependencies (example):

   python -m venv .venv
   .\.venv\Scripts\Activate.ps1
   pip install -r requirements.txt

2. Run the script on a video or webcam:

   python main.py --video "data/fall_5.mp4" --model x --save
   # or for webcam
   python main.py --webcam --model s

Note: The videos available in the data/ folder are extracted from this video.

Analysing results

If --save is used, output videos will be saved to the output/ folder with bounding boxes, keypoints, track IDs, and fall status overlaid.

The colors indicate fall status:

• Green: standing position
• Yellow: intermediate position
• Orange: laying position
• Red: fallen position

The model is not perfect and may produce false negatives. The data/fall_5.mp4 video works well with the x model. Feel free to try your webcam, it works better, even with smaller models.