Final Course Project

• Federal University of Roraima
• Bachelor's Degree in Computer Science
• Author: Natália Ribeiro de Almada
• Advisor: Professor Dr. Herbert Oliveira Rocha

Elderly Fall Detection Project

This project aims to use two existing models (YOLO-NAS and YOLOv8) to compare their performance, first with the original dataset provided with the models and then with adaptations for an author-created dataset, enabling performance evaluation.

Author-Created Dataset:

Warning

Legal Disclaimer: This dataset consists of frames extracted from videos and made available from publicly accessible YouTube videos, compiled exclusively for non-commercial academic research and the development of computer vision systems for elderly fall detection. By using it, you expressly agree that:

All visual material comes from public videos, but intellectual property remains with the original rights holders (creators, institutions, or individuals), according to YouTube Terms of Service and copyright law;

Frame extraction does not authorize commercial use, redistribution, or applications outside health research, elderly safety, or data science;

Commercial use or publication of the images requires explicit permission from the rights holders;

• The creator of this dataset is not responsible for:

Copyright or image rights violations committed by users,

Clinical accuracy or bias in the compiled data,

Consequences of misuse (unauthorized surveillance, algorithmic discrimination, etc.).

├── 156 videos                  ← Training/validation/test images
│   ├── 53 videos of elderly people labeled as "No Fall"/
│   └── 103 videos of elderly people labeled as "Fall"
│ └── 648 frames total          ← Split: training=70% / validation=20% / test=10%
10 videos were kept fully separate and never used in training, for final testing.
Classes: "fall" and "person"

Dataset Organized for v8

Dataset Organized for NAS

Videos Separated for Final Testing

YOLOv8 with Author-Created Dataset:

Project executed on Windows 11

P2_YOLOv8

│
├── dataset/                 ← Training/validation/test images
│   ├── images/
│   └── data.yaml
│
├── videos_testes_validacao/← Validation videos
├── videos_teste/           ← Test videos (never used in training at any stage)
├── runs/
│   └── detect/
│       └── predict/        ← Inference outputs
│           ├── queda/
│           └── naoQueda/
│
├── treinamento_yolo_com_relatorio.ipynb ← Notebook with full pipeline
└── requirements.txt         ← Project dependencies

🚀 Features

• YOLOv8 model training and validation with custom data.
• Evaluation with external videos.
• Automatic classification of test videos:
  • Videos with fall detection → videos_saida/queda/
  • Videos without fall detection → videos_saida/naoQueda/
• Automatic Telegram alert delivery with attached video.
• Chart generation (IoU, Loss, Accuracy, Histograms, etc.) saved as PNG.

---

🛠️ Installation

1. Clone the repository:

   git clone https://github.com/colocaroresto.git
   cd VisionFall

2. Use the notebook available at Notebook on Colab.

Execution

1. Create and activate a virtual environment (Windows):

   python -m venv venv
   .\venv\Scripts\activate
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Start the notebook:

   jupyter notebook
   

4. Run the cells.

YOLO-NAS with Author-Created Dataset:

Project executed in WSL2 on Windows 11

P2_YOLO_NAS

│
├── datasetNAS/ # Custom dataset
│ ├── train/
│ ├── valid/
│ └── test/
│
├── videos_teste/ # Videos for testing with the trained model
├── videos_saida/
│ ├── queda/ # Videos where falls were detected
│ └── naoQueda/ # Videos with no fall detection
│
├── checkpoints/ # Saved trained model weights
└── fall-detection-yolo-nas-train-predict.ipynb

🚀 Features

• YOLOv8 model training and validation with custom data.
• Evaluation with external videos.
• Automatic classification of test videos:
  • Videos with fall detection → videos_saida/queda/
  • Videos without fall detection → videos_saida/naoQueda/
• Automatic Telegram alert delivery with attached video.
• Chart generation (IoU, Loss, Accuracy, Histograms, etc.) saved as PNG.

---

✅ Prerequisites

WSL2 on Windows

Follow the official Microsoft guide.

You must install:

• Ubuntu
• WSL2 as default
• Install Miniconda in WSL, accept the terms, and allow it to be added to PATH.

🛠️ Installation

1. Create the Conda environment for the project:

conda create -n yolonas_env python=3.10 -y conda activate yolonas_env

3. Clone the repository:
   ```bash
   cd /mnt/c/Users/YOUR_USER/Desktop
git clone https://github.com/YOUR_USER/P2_YOLO_NAS.git
cd P2_YOLO_NAS

4. Install dependencies:

   pip install -r requirements.txt

📓 How to run the notebook 5. Start Jupyter in WSL2:

conda activate yolonas_env
jupyter notebook

It will show a URL such as:

http://localhost:8888/?token=... Copy and paste it into your Windows browser.

2. Open the notebook Open the file:

fall-detection-yolo-nas-train-predict.ipynb

2. Run the cells.

📦 Training with your own dataset Make sure your dataset is in YOLOv5/v8 format:

📹 Video Tests Place up to 10 elderly-fall videos (.mp4) in:

videos_teste/ -After testing, the system will create:

videos_saida/
├── queda/
├── naoQueda/

📲 Telegram Integration (optional for both models)

✅ Prerequisites

Have at least one Telegram account

1. Create an app at Telegram configure the following information in the code:

api_id = 'YOUR_API_ID'
api_hash = 'YOUR_API_HASH'
phone = '+55xxxxxxxxx'
dest_user_id = YOUR_ID
dest_access_hash = YOUR_ACCESS_HASH

YOLOv8 Reference Model

• Available on GitHub following the tutorial by: YouTube - Tech Watt

YOLO-NAS Reference Model

• Model available on Kaggle by @stpeteishii, or with small modifications required to run on WSL2 available on Colab
• Dataset available on Kaggle by @selwalyahmad
• COCO pre-trained weights available on Hugging Face

---

@misc{supergradients, doi = {10.5281/ZENODO.7789328}, url = {https://zenodo.org/record/7789328}, author = {Aharon, Shay and {Louis-Dupont} and {Ofri Masad} and Yurkova, Kate and {Lotem Fridman} and {Lkdci} and Khvedchenya, Eugene and Rubin, Ran and Bagrov, Natan and Tymchenko, Borys and Keren, Tomer and Zhilko, Alexander and {Eran-Deci}}, title = {Super-Gradients}, publisher = {GitHub}, journal = {GitHub repository}, year = {2021}, }

@software{yolov8_ultralytics, author = {Glenn Jocher and Ayush Chaurasia and Jing Qiu}, title = {Ultralytics YOLOv8}, version = {8.0.0}, year = {2023}, url = {https://github.com/ultralytics/ultralytics}, orcid = {0000-0001-5950-6979, 0000-0002-7603-6750, 0000-0003-3783-7069}, license = {AGPL-3.0} }