Realtime-DWLS

Detection - Warning - Logging - Scraping

This project was developed as a result of my bachelor thesis:
«Detection and warning of game in vicinity of roads with deep learning»

The goal was to develop a deep learning model and program to detect
and warn against game in vicinity of public roads.

An application utilizing PyTorch with YOLOv5 and cv2 to process video input in
and perform detections in realtime.

The application uses a PyTorch and a custom trained YOLOv5s model to accurately detect
game based on video input. Video data for GUI is processed with cv2.

A simple MQTT messaging client is used to send information about warnings to remote
units, which makes the application able to run in two configurations where you can
process the input locally or remotely.

This application can also be used as a tool to gather new training data for future
iterations of your machine learning model. Images are saved in .png to avoid
artifacts from .jpg

Authors

• @jwlei

Badges

Demo

YOLOv5 Results

Modell	Precision	Recall	[email protected]	[email protected]:.95
YOLOv5 Nano	0.893	0.837	0.869	0.544
YOLOv5 Small	0.897	0.865	0.898	0.564
YOLOv5 Medium	0.963	0.910	0.945	0.691
YOLOv5 Large	0.966	0.915	0.964	0.726
YOLOv5 X-Large	0.970	0.895	0.947	0.687

Features

• Video input from remote URL, Local camera or local media file
• Can scrape images of detections at user defined interval
• Local or remote processing and warning independently
• Can push new YOLOv5 model from remote by supplying a URL in a message
• Downloads remote model on first startup
• Customizable configuration
  • Resize video output
  • Run with setup or straight from configuration
  • Set default video/model and remote model source
  • Save images on detection, with user defined interval
  • User defined confidence threshold at which detections are made
  • Run headlessly or with GUI

Installation

Clone the project

git clone https://github.com/jwlei/real-time-object-detection-YOLOv5-cv2

Go to the project directory

cd Realtime-dwls

Optional: Create and activate a virtual environment

python -m venv name_environment

name_environment\Scripts\activate.bat

Install dependencies

pip install -r requirements.txt

Install PyTorch for your version, supplied is for Python 3.10:

• CUDA

pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/cu113

• Non-CUDA

pip install torch torchvision

Run Locally

Run the main application

python main.py

External MQTT Subscriber with a simple warning GUI

python external_mqtt/ext_mqtt_subscriber.py

External MQTT Publisher which can supply an URL for a model for the main application to download

python external_mqtt/ext_mqtt_publisher.py http://url.to/yourmodel.pt

Links to trained models

The URL supplied can be used to directly download the model to the program

YOLOv5-Nano (https://dl.dropboxusercontent.com/s/cbvc681akdp9rc1/yolov5N.pt)

YOLOv5-Small (https://dl.dropboxusercontent.com/s/nxobi6gciwsaygb/yolov5S.pt)

YOLOv5-Medium (https://dl.dropboxusercontent.com/s/3y47tbcz6e33a40/yolov5M.pt)

YOLOv5-Large (https://dl.dropboxusercontent.com/s/a1t8w7tetq4naov/yolov5L.pt)

YOLOv5-XLarge (https://dl.dropboxusercontent.com/s/d4ouyyqj4ji49a3/yolov5XL.pt)

License

MIT