Common Lane Change Decision Making DL/RL models for highD dataset

Overview

I am trying to reproduce papers in Lane Change Decision Making which evaluate with highD dataset.

• Paper in arXiv for highD Dataset

Details

A Learning-Based Discretionary Lane-Change Decision-Making Model with Driving Style Awareness

• Paper in arXiv
• corresponding to dop(driving operational picture) model in this repo.
• test result, got overall accuracy of 88.9% in my training and testing process with 2 NVIDIA A40 GPUs. However, there exists a gap between my result and the offical result.

Lane-Change-Prediction-LSTM

• referenced and modified from nqyy's repo.
• Github repo
• corresponding to rnn model in this repo.

Reinforcement Learning methods (DQN / Double DQN / Dueling DQN / PPO)

• located in rl_models/.
• use the same 16-dimensional Feature-A state vector as the RNN model.
• a data-driven LaneChangeEnv wraps the pre-processed highD Normal-feature pickles so that any standard RL algorithm can be trained without an external simulator.

Usage

1. Please put the HighD dataset to */highd-dataset-v1.0/data/ corresponding to the value of DATASET_ROOT in ./configs/constant.py.

2. Modify the configs/config.py, The variable of FEATURE_CHOICE should be modified to CNN_FC, NORMAL for dop model and rnn model, respectively. Other Feature construction variables in configs/config.py also should be modified depends on the user's requirements.

3. Run python3 ./calculate/get_event_feature.py and python3 ./calculate/get_time_series_feature.py to process the dataset and the output data for dop model and rnn model will be stored into the folder of ./output in pickle format, respectively.

4. Run python3 dop_cnn_model/train.py and python3 rnn_model/rnn_model.py to train and test for dop model and rnn model, respectively.

5. RL methods — set FEATURE_CHOICE = "Normal" in configs/config.py and run python3 calculate/get_time_series_feature.py first to generate the Normal-feature pickles. Then launch any of the four RL agents:

   python3 rl_models/dqn/train.py          # DQN
   python3 rl_models/double_dqn/train.py   # Double DQN
   python3 rl_models/dueling_dqn/train.py  # Dueling DQN (+ Double DQN update)
   python3 rl_models/ppo/train.py          # PPO with GAE

   Best-checkpoint weights are saved to output/dqn_best.pth, output/double_dqn_best.pth, output/dueling_dqn_best.pth, and output/ppo_best.pth respectively.

   RL models overview:

   Model	Key idea
   DQN	Q-learning with experience replay and a target network
   Double DQN	Decouples action selection (online net) from action evaluation (target net) to reduce overestimation bias
   Dueling DQN	Splits Q into value stream V(s) and advantage stream A(s,a) for better generalisation across actions
   PPO	On-policy policy-gradient with clipped surrogate objective and GAE advantage estimation

6. Run python3 extract_conscend.py to generate a ConScenD-style scenario dataset from highD. The script exports:

   • scenario metadata JSON files in ./output/conscend/metadata/
   • OpenSCENARIO .xosc files in ./output/conscend/scenarios/
   • straight-road OpenDRIVE .xodr files in ./output/conscend/roads/

7. Optional arguments for the ConScenD workflow:

   • python3 extract_conscend.py --recordings 1 2 3 to process selected recordings only
   • python3 extract_conscend.py --dataset-root /path/to/highd-dataset-v1.0
   • python3 extract_conscend.py --output-root /path/to/output/conscend

ConScenD reproduction notes

The new extraction pipeline follows the paper title and available methodological details for The ConScenD Dataset: Concrete Scenarios from the highD Dataset According to ALKS Regulation UNECE R157 in OpenX. It adds a ConScenD-style scenario export workflow on top of the existing highD readers in this repository.

Implemented scenario logic:

• ALKS speed-domain filtering at 60 km/h equivalent
• lane-change detection from laneId transitions
• cut-in classification using adjacent following / alongside vehicles before the maneuver
• non-lane-change classification for free-driving and car-following scenarios
• OpenSCENARIO and OpenDRIVE export for each extracted scenario

This implementation is intended to be reproducible with the public highD dataset and repository code, while remaining explicit that some paper details were inferred from the regulation and related OpenX references.

Python version

python_version == 3.12.3

Requirements

Packages installation guide: pip3 install -r requirement.txt Anaconda was recommended here.