TL;DR
This paper introduces LiRA-CD, a comprehensive, open-source dataset for road condition modeling. It combines in-vehicle sensor data from electric cars with high-precision reference data from specialized survey vehicles, offering a rich resource for developing and validating road condition prediction models.
What Problem Does This Paper Address?
Accurate and real-time assessment of road conditions is crucial for road maintenance, traffic safety, and optimizing vehicle performance (e.g., energy consumption). However, developing robust models for this requires extensive, high-quality data that links readily available in-vehicle sensor information to standardized road condition parameters. Existing datasets often lack this direct linkage or the comprehensive nature needed for advanced machine learning and pavement engineering applications. This paper addresses this gap by presenting LiRA-CD, a dataset designed to bridge this divide and facilitate research in this area.
Methods
The LiRA-CD dataset was collected using two primary methods:
- In-vehicle sensor data: This was gathered from a fleet of electric vehicles (Renault Zoe) equipped with an AutoPi Telematics Unit. This unit, featuring a Raspberry Pi, GPS, and accelerometer, captured over 50 different sensor signals by interfacing with the vehicle's Controller Area Network (CAN) bus. Data was transmitted periodically to a cloud-based system.
- Reference road condition data: This high-quality data was collected by standard survey vehicles operated by the Danish Road Directorate (DRD). These vehicles were equipped with specialized instruments, including:
- P79 Profilometer: Measures longitudinal and transverse road profiles using lasers.
- ARAN9000: A road scanning vehicle that uses cameras and a Laser Cracking Measurement System (LCMS) to quantify road defects.
- VIAFRIK: A skid resistance device compliant with CEN/TS 15,901-5.
The paper also details the data processing steps, including on-board translation of raw sensor data into physical units using OBD protocols and the CanZE application, and provides formulas for further translation of specific sensor signals.
Datasets
The primary dataset introduced is the Live Road Assessment Custom Dataset (LiRA-CD). This dataset comprises:
- In-vehicle sensor data: Collected from electric vehicles, stored in
.hdf5and.txtformats. - Reference road condition data: Collected from standard survey vehicles, stored in
.csvformats.
LiRA-CD covers approximately 230 km of highway and urban roads in Copenhagen, Denmark. The dataset is structured into three main subsets:
- Car sensor data for road condition modeling.
- Reference data for road condition modeling.
- Data from a platoon friction test involving a car driving behind the VIAFRIK vehicle.
This dataset serves as a new benchmark by directly linking in-vehicle sensor data to detailed, high-quality road condition measurements.
Key Insights and Ideas
- Bridging the gap: The most significant insight is the direct linkage between readily available in-vehicle sensor data and precise, standardized road condition parameters. This is crucial for developing practical, wide-area road condition monitoring systems that can leverage the growing number of connected vehicles.
- Richness of data: The dataset's breadth, with over 50 in-vehicle sensors and multiple types of reference measurements (friction, roughness, damage, texture), allows for multi-faceted analysis and the development of sophisticated predictive models.
- Template for future research: LiRA-CD is explicitly designed to be a template for future studies, enabling researchers and pavement engineers to benchmark algorithms and explore new methodologies for road condition assessment.
Key Results
The paper primarily focuses on describing the dataset itself rather than presenting specific model performance results. However, it highlights the dataset's utility for:
- Developing prediction models for surface friction, texture, road roughness, road damages, and energy consumption.
- Training, validating, and testing machine learning algorithms for estimating road conditions.
- Benchmarking the performance of different algorithms.
The dataset is freely available for researchers and pavement engineers.
Limitations and Future Work
- Data Pre-processing: The authors note that the raw data has not been aligned or structured, requiring users to perform several pre-processing steps. These include re-orientation of accelerometer axes, map-matching, interpolation of GPS coordinates, and re-sampling of sensor readings to ensure consistency.
- Image Data: While the reference data includes some visual information (e.g., from ARAN9000), the paper suggests that incorporating additional image datasets could further enhance the quality of training data and enable vision-based methods.
- Future Studies: The dataset is presented as a foundation for future research, encouraging its use for developing new interpretation schemes, exploring physical models, and advancing machine learning applications in road condition monitoring.
Summary updated