fall-detection7

Fall Detection Using YOLOv8-Pose

Project Overview

This project detects human falls in pre-recorded videos/CCTV cameras and produces an annotated output video that clearly highlights the fallen person. It is designed to be simple to run in Google Colab, reading your video from Google Drive and saving the annotated output back to Drive.

What This Project Does

• Uses a pose-estimation model (YOLOv8-Pose) to identify key body joints.
• Applies a lightweight, explainable rule to infer when a person is likely to have fallen (based on posture and shape over consecutive frames).
• Draws a thick red bounding box and pose skeleton around the fallen person.
• Saves the annotated video and optional event information to your Google Drive.

Where It’s Useful

• Elderly care and assisted living monitoring.
• Hospitals, clinics, and rehabilitation wards.
• Industrial and workplace safety monitoring (warehouses, construction).
• Public buildings and office corridors. Please ensure you follow local privacy regulations and obtain consent where required.

Features

• Clear, bold on-screen fall alerts.
• Explainable rule-based approach (no additional training required).
• Google Colab workflow using Google Drive as input/output storage.
• Adjustable sensitivity through simple thresholds (angle, box ratio, and persistence).

Requirements

• A Google account with access to Google Colab and Google Drive.
• A video file stored in Google Drive.
• A modern browser. A GPU runtime in Colab is recommended for faster processing but not required.

How to Use (High Level, No Code)

1. Open Google Colab in your browser.
2. Mount your Google Drive within the Colab notebook.
3. Point the notebook to your input video in Drive and choose an output path in Drive.
4. Run the fall-detection cell to process the video.
5. When processing finishes, open the saved annotated MP4 from your chosen Drive folder.
6. If the results feel too sensitive or not sensitive enough, adjust the thresholds (e.g., angle, height-to-width ratio, and number of consecutive frames) in the notebook and run again.

Output

• An annotated MP4 saved to your Google Drive showing:
  • Red box and skeleton over the fallen person(s).
  • A bold red “FALL DOWN DETECTED” banner when a fall is confirmed.
• Optional logs and snapshots can also be saved, depending on the configuration in the notebook.

Tuning Guidance

• Increase the required number of consecutive frames to reduce false positives.
• Lower the maximum torso-angle threshold or height-to-width ratio to make detection stricter.
• Use a more accurate pose model if you have a GPU available.
• Ensure adequate lighting and camera angle for best results (slight side view generally helps).

Results:

Not Fall

<img width="572" height="889" alt="image" src="https://github.com/user-attachments/assets/bd5e51bc-d041-4ab8-80ef-cefe752391b9" />

Fallen Down

<img width="581" height="802" alt="image" src="https://github.com/user-attachments/assets/e40400dd-59c4-414f-a8ec-20ce3c71a26f" />

Privacy & Ethics

• Obtain consent where necessary.
• Minimize retention of video and snapshots.
• Protect stored media with appropriate access controls.
• Follow all applicable laws and organizational policies.

Who Created This Project

Created by: Fatima Noor Contact: fatimanoorr39@gmail.com

License

This repository is provided under the MIT License (or your preferred license). Please review and adjust the license text and file to match your needs.