Michigan Traffic Lab, University of Michigan
This project contains the source code and data for the paper "Learning naturalistic driving environment with statistical realism".
Learning-Naturalistic-Driving-Environment/
|__ basemap: visualize the simulation environment
|__ behavior_net: behavior modeling network
|__ configs: configuration files
|__ data: source data for inference and training
|__ geo_engine: coordinate transformation
|__ result_analysis: analyze experiment results and plot figures
|__ road_matching: region of interest (ROI) matching for metrics analysis
|__ sim_evaluation_metric: calculate simulation metrics
|__ simulation_modeling: simulation environment
|__ trajectory_pool: manage vehicle trajectory during simulation
|__ vehicle: vehicle class
|__ requirements.txt: required packages
|__ run_inference.py: main function for running simulations
|__ run_training_behavior_net.py: main function for training behavior modeling network
If you find this work useful in your research, please consider cite:
@article{yan2023learning,
title={Learning naturalistic driving environment with statistical realism},
author={Yan, Xintao and Zou, Zhengxia and Feng, Shuo and Zhu, Haojie and Sun, Haowei and Liu, Henry X},
journal={Nature Communications},
volume={14},
number={1},
pages={2037},
year={2023},
publisher={Nature Publishing Group UK London}
}
git clone https://github.com/michigan-traffic-lab/Learning-Naturalistic-Driving-Environment.gitYou are recommended to create a new Conda environment to install the project
conda create -n NNDE python=3.9To install the Python packages required for this repository, execute the command provided below:
conda activate NNDE
pip install -r requirements.txtFor Anaconda users (recommended), you can use conda-develop to add the main folder path to the specific conda virtual environment as follows:
conda activate NNDE
conda install conda-build
conda develop .Alternatively, set the main folder path (your-absolute-main-folder-path) to PYTHONPATH in system variables.
The data for inference (e.g., model checkpoints), training (e.g., the trajectory data for training), and analysis
(e.g., ground truth data for metrics analysis)
can be downloaded here(852MB).
Unzip the file and put them into the data folder, the file structure should look like:
Learning-Naturalistic-Driving-Environment/data/
|__ inference
|______ AA_rdbt
|___________ behavior_model_ckpt
|___________ safety_mapper_ckpt
|___________ ROIs-map
|___________ ...
|______ rounD
|__ paper-inference-results
|______ AA_rdbt_paper_results
|______ rounD_paper_results
|__ statistical-realism-ground-truth
|______ AA_rdbt_ground_truth
|______ rounD_ground_truth
|__ training
|______ behavior_net
|___________ AA_rdbt
The inference folder contains the model checkpoints and the other files needed for inference.
The paper-inference-results folder contains the inference results for the paper.
The statistical-realism-ground-truth folder contains the ground truth data for metrics analysis.
The training folder contains the trajectory data for training the behavior modeling network.
If you do not want to re-train the model and only want to do the simulation, you will only need the inference and
statistical-realism-ground-truth data.
In this repository, we provide the Ann Arbor dataset that we used to train NeuralNDE in the training folder.
The dataset contains the vehicle trajectory data perceived by the roadside perception system
deployed at the two-lane roundabout at the intersection of State St. and W. Ellsworth Rd. in Ann Arbor, Michigan.
- The data was collected from 10 am to 5 pm on May 2nd, 2022.
- The data sample rate is 2.5Hz.
The dataset contains multiple pickle files.
|-2021-05-02 10-03-48-671927.pickle
|-2021-05-02 10-03-49-076785.pickle
|-2021-05-02 10-03-49-475787.pickle
...
Each pickle file contains the vehicle trajectory data of one frame. Here is an example of the information in a pickle file
[
{
"id": "1",
"location":
"x": 86.76476951608201,
"y": -25.098855800926685,
"z": None
"speed_heading": 163.5021444664423
...
},
{
"id": "3",
"location":
"x": 69.60102021666243,
"y": -24.766377145424485,
"z": None
"speed_heading": 198.9397031364029
...
},
...
]
where id denotes the unique vehicle id, x, y denote the vehicle position (local x, y coordinates, unit in meter),
and speed_heading denotes the vehicle heading (starting from east and anticlockwise, unit in degree).
The rounD dataset is available here and can be processed into the same format for training.
Config the configuration file (e.g., simulation wall time, number of simulation episodes, etc.)
in the configs folder
and run the following to initialize the simulator:
No visualization:
python run_inference.py --experiment-name your-experiment-name --folder-idx 1 --config ./configs/rounD_inference.ymlWith visualization:
python run_inference.py --experiment-name your-experiment-name --folder-idx 1 --config ./configs/rounD_inference.yml --viz-flagNote that in the default configuration file, we set the use_gpu flag to False to use CPU for inference. Change it
into True if you want to use GPU for inference.
We simulate approximately 15,000 hours of simulation to record data and generate simulated statistics to validate the statistical realism of the NeuralNDE, where all data are used for calculating crash-related metrics (e.g., crash rate, type, severity) and 100 hours of data are used for other metrics.
The default time resolution is 0.4s, set interpolate_flag to True in the configuration file to interpolate the
trajectory to a finer resolution.
Config the configuration file and run the following commands to train behavior modeling network.
The AA roundabout training configuration file is used as an example. The result will be saved
in ./results/training/behavior_net folder if you are using the default --save-result-path.
python run_training_behavior_net.py --config ./configs/AA_rdbt_behavior_net_training.yml --experiment-name AA_rdbt_behavior_net_trainingChange the model path behavior_model_ckpt_dir in the configuration file if you want to use
newly trained behavior modeling network for inference.
Important note: Note that training can continue from previous checkpoints. When the --experiment-name and --save-result-path are the same, the code will check whether this experiment has been conducted and continue from the last_ckpt.pt under this experiment.
Tensorboard can be used to visualize the training process (loss, etc).
Run the following code in a terminal if you are
using the default --save-result-path during training, otherwise, change
the --logdir path to your defined save-result-path. Then, open a browser and go to http://localhost:6006.
tensorboard --logdir=./results/training/behavior_netConfig the settings (e.g., simulation results path) in result_analysis/plot_all_results.py
and run the following to analyze simulation results and plot figures.
cd result_analysis
python plot_all_results.pyNote that you need config accordingly when running the inferences to save data for the analysis.
For example, in order to generate instantaneous distributions, the
gen_realistic_metric_flag and instant_speed flags need be set as True in your inference yml file.
In order to generate crash type and severity results, you need to save crash data
(i.e., set save_collision_data_flag as True).
Contributions are what make the open source community such an amazing place to be learn, inspire, and create. Any contributions you make are greatly appreciated.
- Fork the Project
- Create your Feature Branch (
git checkout -b feature/AmazingFeature) - Commit your Changes (
git commit -m 'Add some AmazingFeature') - Push to the Branch (
git push origin feature/AmazingFeature) - Open a Pull Request
This project is licensed under the [PolyForm Noncommercial License 1.0.0]. Please refer to LICENSE for more details.
This work is supported by the U.S. Department of Transportation Region 5 University Transportation Center: Center for Connected and Automated Transportation (CCAT) of the University of Michigan, and National Science Foundation.
Xintao Yan (xintaoy@umich.edu)
Zhengxia Zou (zzhengxi@umich.edu)
Henry Liu (henryliu@umich.edu)