lukicdarkoo / cv-car-control

Quick Start

Dependencies

Here is the list of dependcies:

• Webots
• OpenCV for Python

In case you use Ubuntu or similar, run:

sudo snap install webots
pip3 install opencv-contrib-python

Project Files

Clone the project as:

git clone --depth=1 https://github.com/lukicdarkoo/cv-car-control.git

Make sure you have git lfs installed if you want to get pre-trained model as well.

Simulation

Standalone fist detector (without simulation) is located in ./controllers/steering_wheel_controller/wheel_detector.py and you can run it as:

python3 ./controllers/steering_wheel_controller/wheel_detector.py

However, if you want to have a fun and run the simulation, execute the following command:

webots ./worlds/city.wbt

(or simply open ./worlds/city.wbt in Webots GUI)

Training

If you want to improve the model provided within this project, here is an explanation on how to do it.

Make Dataset

Install labelImg and open it:

pip3 install labelImg
labelImg

In labelImg make sure to click YOLO and annotate all the images. Once all the images annotated move them, together with the corresponding *.txt files, to dataset/train and dataset/valid. It is very important to move 10-20 percent of the images to dataset/valid directory and the rest, move to dataset/train. YOLO also needs a list of images for validation and training written to *.txt files. You can make the files by invoking the following commands from the project's root:

cd ./dataset
find train -name '*.jpg' > train.txt
find valid -name '*.jpg' > valid.txt

Train Locally

If you have GPU and CUDA available then you may prefer to train the YOLO locally.

First, install darknet with GPU and CUDA support:

git clone https://github.com/pjreddie/darknet
cd darknet
sed -i 's/CUDNN=0/CUDNN=1/g' darknet/Makefile
sed -i 's/GPU=0/GPU=1/g' darknet/Makefile
make -j8

To accelerate training it is usually a good idea to start from pre-trained model. Download the pre-trained model in the project directory:

wget https://pjreddie.com/media/files/darknet53.conv.74

Finally, you can initiate the training as:

cd ./dataset
$HOME/darknet/darknet detector train ./custom.data ../yolov3-tiny_train.cfg ../darknet53.conv.74 

and your models will be available in ./dataset/backup.

Train on Google Colab

In case you don't have GPU it probably a good idea to use Google Colab (click on the badge above). After the Google Colab is ready, upload yolov3-tiny_train.cfg and ./dataset folder compressed as dataset.tar.xz.

You can create dataset.tar.xz as tar -cf ./dataset ./dataset.tar.xz

Run all cells and after approx. 1 hour of training the model will be ready waiting in ./dataset/backup.