Duckiebot Object Detection with YOLO v3

This project focuses on the training and implementation of YOLO v3 in a Duckiebot for real-time object detection. The goal is to identify and classify objects such as Duckiebots, Ducks, Road signs, and Stop signs.

Darknet

Darknet is an open source neural network framework written in C and CUDA. It is fast, easy to install, and supports CPU and GPU computation.

Getting Started

Cloning the Darknet Repository

Clone the Darknet repository from https://github.com/pjreddie/darknet.git by running the following command:

git clone https://github.com/pjreddie/darknet.git

Creating the Dataset

To train the model, create the dataset by collecting images of Ducktown and taking your own images. Use a Laplacian filter to determine image blur. The dataset should consist of high-quality images and should be stored in the /datasets directory. It can be done running the detect_blurry_img.py code.

Labeling Images

Label the images using the labelImg software. For installation instructions, refer to the labelImg repository. Define the classes in a text file named predefined_classes.txt located in /labelImg/data. Here are the predefined classes:

bot
duckie
stop_sign
road_sign

To label the images, open the dataset directory in the labelImg software.

Dataset Split

Divide the labeled dataset into three sets: 85% for training, 7.5% for testing, and 7.5% for validation. The dataset split files train.txt and valid.txt, which include the path to each training and validation image, are included in this repository inside /datasets. This can be done running the create_datasets.py code.

YOLO Configuration Files

Prepare the YOLO configuration files. Create the duckie.data file and place it inside /darknet/cfg. Here is the content of the duckie.data file:

classes=4
train=datasets/train.txt
valid=datasets/valid.txt
names=data/duckie.names
backup=duckie_backup

Create the duckie.names file and place it inside /darknet/data. The file should contain the class names listed above.

Duplicate the yolov3-tiny.cfg file and name it yolov3-tiny-duckie.cfg. Make the following changes inside the yolov3-tiny-duckie.cfg file: filters=27 (lines 127 and 171) classes=4 (lines 135 and 177)

Training the Model

To train the model and get the weights, run the Colab notebook.

Testing the Model

To test the model with the weights generated after training it with the provided dataset, run the Colab notebook.