Lane detection ML training on Euro Truck Simulator 2 (ETS2).

A small article is also published here: https://lotdcotw.medium.com/autobahn-autopilot-euro-truck-simulator-2-4183d2b3f948

This repository/tutorial is heavily fed by the following sources;

1. tvtLANE Dataset
   Please check docs/tvtLaneDataset.pdf
2. @quinnzou
   https://github.com/qinnzou/Robust-Lane-Detection
3. Sentdex
   1. https://pythonprogramming.net/game-frames-open-cv-python-plays-gta-v/
   2. https://www.youtube.com/user/sentdex

1. Roadmap
2. Requirements
3. Configuration
4. Basic
5. Citation
6. Social Media

1. Lane Detection
   1.1. Use a pre-trained model from real-life for lane detection. (OK)
   1.2. Create a dataset from ETS2 (OK)
   1.3. Train a model with ETS2 data (TODO)
   
   
2. Steering
   1.1. Train steering data (TODO)
   
   
3. Object Detection
   1.1. Environment awareness (TODO)
   1.2. Braking (TODO)

1. Python >= 3.6 and PIP3
2. CUDA >= 8
3. Torch with CUDA support
   https://pytorch.org/
4. Euro Truck Simulator 2
   This is optional. You can try this with other games.

Run the following command to install all required libraries:

pip3 install -r requirements.txt

1. Application Configuration
   Please refer to config.py for dataset settings, screen positioning, paths and more.
   Application arguments are located in this file.
2. Euro Truck Simulator 2
   Windowed mode (1280x720 is used in this project), front camera (button 6), auto wheel positioning, sensitivity adjustments. All for your consideration.
3. Model
   If you are going to use a pretrained model, download it and move it under ./model and validate pretrained_path in config.py. Please refer to ./model/pretrained.txt

python3 dataset_creator.py

A sequence of images from ETS2 will be created under ./data/testset/ets2_...
All images will be scaled, cropped and ready for the model.
Do not forget to remove unnecessary captures while you are switching windows.
Basically, check the images inside first and last 2 created folders.

After the sequence of images are ready to be trained, first create an index file.

python3 dataset_indexer.py {FOLDER}

If your GPU is not powerful enough, please refer to ./model/pretrained.txt
Before running the command below, make sure you have entered correct folder names like explained in config.py

python3 train.py

When your model is ready as trained or pretrained and GPU power, test your model with test.py
If you are just want to test with a pretrained model, go to the next step.

python3 test.py

1. Run ETS2 and position it to top left screen and validate your screen positioning configuration.
2. Start the game with camera option 6 (front camera, no truck visuals)
3. Run the following command:

python3 live.py --mode=0 --continuous=False

A the moment, just 1 mode is available.
Not a solution but for practice and experiement, HoughLinesP method and small geometry calculations only for testing purposes.

Understanding the lines in model's matrix is work in progress. Please refer to the article.

Follow first and second setup in 'Single Capture Mode' and run the following command:

python3 live.py --mode=0 --continuous=True

Check the video below out; with mode=0 and continuous=True:

For mode=0, you can train your geometry skills by using lane vectors.

After mode 2 is ready, the update will take its place here also.

1. Q. Zou, H. Jiang, Q. Dai, Y. Yue, L. Chen, Q. Wang
   Robust Lane Detection from Continuous Driving Scenes Using Deep Neural Networks, IEEE Transactions on VehicularTechnology, 2019.
2. TuSimple lane-detection dataset. http://benchmark.tusimple.ai/#/t/1/datase

Linkedin: lotdcotw
Youtube: lotdcotw
Twitter: lotdcotw
Instagram: lotdcotw
Facebook: lotdcotw