jHaselberger/SADC-Situation-Awareness-for-Driver-Centric-Driving-Style-Adaptation

Dataset Card for Dataset SADC

There is evidence that the driving style of an autonomous vehicle is important to increase the acceptance and trust of the passengers. The driving situation has been found to have a significant influence on human driving behavior. However, current driving style models only partially incorporate driving environment information, limiting the alignment between an agent and the given situation.

Therefore, we propose a dataset for situation-aware driving style modeling.

Dataset Details

Dataset Description

The dataset is composed as follows: the pretrain set DP is split into a training subset DP,T with 242 887 samples, and a validation subset DP,V with 61 400 samples. Similarly, the validation set DV is split into a training subset DV,T and a validation subset DV,V with 138 572 and 34 767 samples. Each subset consists of 1280 × 960 images, driving behavior indicators like the distance to the lane center, vehicle signals like velocity or accelerations, as well as traffic conditions and road type labels.

• Curated by: Johann Haselberger
• License: CC-BY-4.0

Dataset Sources

We collected over 16 hours of driving data from single test driver as pretrain data. For the driving style adaptation, we collected driving behavior data from five different subjects driving on the same route for one hour, denoted as validation data.

Usage

Download Script

For an easy usage of our dataset, we provide a download script with our repo: https://github.com/jHaselberger/SADC-Situation-Awareness-for-Driver-Centric-Driving-Style-Adaptation/blob/master/utils/download_dataset.py.

python download_dataset.py --target_dir ../data --split pretrain_train

List Available Split Names

from datasets import load_dataset, get_dataset_split_names

split_names = get_dataset_split_names("jHaselberger/SADC-Situation-Awareness-for-Driver-Centric-Driving-Style-Adaptation")
print(f"Available split names: {split_names}")

Inspect some Samples

from datasets import load_dataset, get_dataset_split_names
from matplotlib import pyplot as plt
import pandas as pd

dataset = load_dataset("jHaselberger/SADC-Situation-Awareness-for-Driver-Centric-Driving-Style-Adaptation", split="val_val", streaming=True)

samples = dataset.take(50)
df = pd.DataFrame.from_dict([s for s in samples])
print(df.head())

Visualize some Time-Series

fig, ax1 = plt.subplots()
ax2 = ax1.twinx()
ax1.plot(df["frame_nr"],df["v_kmph"],"ko-",label="velocity")
ax2.plot(df["frame_nr"],df["steering_torque_N"],"ro-",label="steering torque")

ax1.set_xlabel('Frame')
ax1.set_ylabel('Velocity in km/h', color='k')
ax2.set_ylabel('Steering Torque in N', color='r')
plt.show()

Visualize the Camera Image

plt.imshow(df["frame"].iloc[-1])
plt.axis('off')
plt.show()

Dataset Structure

Dataset Splits

Split	Number of Samples	Description
Used for the Experiments in the Paper
pretrain	304287	The full pretrain dataset.
pretrain_train	242887	Subset of pretrain used for training.
pretrain_val	61400	Subset of pretrain used for validation.
val_train	138572	Subset of validation used for training.
val_val	34767	Subset of validation used for validation.
Additional Data
pretrain_unfiltered	1180252	The full unfiltered pretrain dataset.
val_unfiltered	686328	The full unfiltered validation dataset.

Files

• The folder driving_data contains the vehicle signals. Downloading these files is optional and is only required if you do not want to download the entire image data set.
• The folder image_lists contains the image lists used for training of the featrue encoders and NN-based behavior predictors. Downloading these files is optional.

Personal and Sensitive Information

To blur vehicle license plates and human faces in the camera frames, we utilize EgoBlur https://github.com/facebookresearch/EgoBlur.
Furthermore, all subject-related data, including the socio-demographics, are anonymized.

Bias, Risks, and Limitations

Considering the limitations of our dataset, real-world tests should be conducted with care in a safe environment. To publish the data concerning privacy policies, we utilized a state-of-the-art anonymization framework to blur human faces and vehicle license plates to mitigate privacy concerns.

Citation [optional]

BibTeX:

@misc{haselberger2024situation,
      title={Situation Awareness for Driver-Centric Driving Style Adaptation}, 
      author={Johann Haselberger and Bonifaz Stuhr and Bernhard Schick and Steffen Müller},
      year={2024},
      eprint={2403.19595},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

APA:

Johann Haselberger, Bonifaz Stuhr, Bernhard Schick, & Steffen Müller. (2024). Situation Awareness for Driver-Centric Driving Style Adaptation.