YOLO Security

By Lauren Spee for MAIS 202

YOLO Security is a home security application that harnesses the power of computer vision to keep the user safer and in the loop about the security of their home. It connects to the cameras around the user's house to keep an extra set of eyes on the perimeter. If the model sees a person, a car, or an animal, it will send the user a notification with an image of what it saw. This allows the homeowner to know in real-time what's happening outside their house, whether they're there or not.

Currently, the prototype is run on a web app, with a place where the user can upload a video of their own to see the model in action.

For the project showcase, the site is hosted here: https://7570-173-179-6-22.ngrok-free.app/ . Beware, due to computing constraints, it's very slow.

Information

The product uses the YOLOv5 model, a lightweight and fast object-detection model that can run in real-time. It was trained on the Pascal VOC 2007 and 2012 datasets to generate weights and optimize hyperparameters. The model is used to identify objects, classify them, and render bounding boxes on video frames as it's happening.

Performance

The final model has a mean average precision value of 71%

Organization

This repository holds the training notebook and video processing scripts, along with the code for the webapp.

Deliverables/ holds MAIS 202 project reports

webapps/ holds the code to run the flask webapp

heroku_attepmts/ contain my wip code to get this webapp running on heroku

YOLOv5_Training.ipynb the jupyter notebook used to train the model

process_footage.py the python script used to run the security footage through the model

Possible improvements

Given time, my plan is to get this webapp running on a "real" domain hosted on a stronger computer, and improve the weights for better performance.

How to run

Make sure you have the following installed on your system:

• Python 3.10 or later
• ffmpeg (for video processing)

Webapp

1. Clone this repository with git clone https://github.com/LaurenS419/Yolo-Security
2. Navigate to /webapp
3. Download the dependencies via pip install -r requirements.txt
4. Run the app with flask run
5. The app will be accessible at http://127.0.0.1:5000

Scripts

1. Training
   1. Run the training notebook either in Google Colab or on your laptop (must have an NVIDIA GPU and have CUDA and cuDNN installed)
      • if you're using your laptop, you must use a python virtual environment with a version that's compatible with the versions of CUDA and cuDNN that you have
   2. To get weights, go to yolov5/runs/train/expX/weights (replace X with your most recent training run number)
2. Running
   1. Import the weights file into the root directory of this project
   2. Download a video you want processed into the root directory, and make sure the paths in process_footage.py are right
   3. Run process_footage.py