VTGNet

VTGNet: A Vision-based Trajectory Generation Network for Autonomous Vehicles in Urban Environments

Requirements

conda env create -f environment.yaml

This command will create a conda environment named vtgnet

Test the model

conda activate vtgnet
python test.py

The result will be saved in folder test_results/ as the following image:

VTG-Driving dataset

Download our dataset here and extract it into folder VTG-Driving-Dataset/

Alternate download link with code: 4xn8

VTG-Driving-Dataset
├─ clear_day/
├─ clear_sunset/
├─ foggy_day/
├─ rainy_day/
├─ rainy_sunset/
├─ dataset_left.csv
├─ dataset_right.csv
├─ dataset_straight.csv

This dataset is collected in CARLA simulator. The driving information (location, speed, control actions, etc.) is recorded in efo_info.csv for each episode. The setups are listed as following:

• Dynamic traffic with pedestrians and other vehicles
• Collected in Town01 at desired speed of 40 km/h
• Five weather conditions
• 100 driving episodes for each weather
• 16.6 hours, 288.7 km
• With behavoirs recovering from periodic off-center and off-orientation mistakes.

The dataset_{left, right, straight}.csv files keeps the training information extracted from the dataset.

Train the model on our dataset

python train.py --direction 2  --load_weights True --batch_size 15

--direction {0,1,2} 0: keep straight; 1: turn right; 2: turn left

--load_weights {True,False} load pre-trained weights on Robotcar or not

Citation

@article{Cai2020VTGNetAV,
  title={VTGNet: A Vision-based Trajectory Generation Network for Autonomous Vehicles in Urban Environments},
  author={Peide Cai and Yuxiang Sun and H. Wang and M. Liu},
  journal={IEEE Transactions on Intelligent Vehicles},
  year={2020},
  doi={10.1109/TIV.2020.3033878}
}