PathCare: A Dataset for Road Fault Diagnosis

Published: 29 October 2024| Version 2 | DOI: 10.17632/6p52w7d5xd.2
Contributors:
,
,
,
,
,

Description

This dataset containing 386 images and 4 videos. The images were captured using a GoPro Hero 9 camera mounted on a moving vehicle. The purpose of the recording was to monitor road conditions. The recorded videos, featuring various road types, were later converted into individual frames. The dataset is categorized into three distinct classes: Crack, Pothole, and Uneven Surface

Files

Steps to reproduce

The data collection process utilized a GoPro Hero 9 camera, set to capture at 24 frames per second with a resolution of 2704 x 1520. The camera was mounted on a car at an optimized height, carefully determined based on the road's characteristics to ensure accurate and comprehensive data capture. The vehicle was driven over a 28-kilometer stretch of road, maintaining a consistent speed to achieve uniformity in the captured data. Environmental factors such as road conditions, lighting, and any obstacles encountered were meticulously documented to account for potential anomalies. Once collected, the footage was processed using specialized video analysis software, preparing the dataset for further stages of the research. Before implementing YOLOv7 instance segmentation, the Roboflow tool was employed for data annotation, ensuring that all instances of cracks, potholes, and uneven surfaces were accurately labeled. The annotated data was then processed for machine learning tasks. The analysis was performed on a PC equipped with an Intel i7 9th generation processor, 16GB of RAM, and an NVIDIA RTX 2060 Super GPU, ensuring efficient processing of the dataset and rapid execution of machine learning models. With this setup, YOLOv7 instance segmentation was applied to detect the aforementioned road surface issues. The model achieved an impressive 98.6% classification accuracy and 99% mean average precision (mAP), validating the robustness of the data and methodology. No reagents were used in this workflow, and the process adhered to a structured sequence of data collection, annotation, processing, and analysis to ensure reproducibility and precision.

Institutions

University of Arkansas at Little Rock, Universidad de Malaga, Mehran University of Engineering and Technology

Categories

Real-Time System, Detection Technique, Change Detection, Features Detection, Real-World Knowledge, Road Safety

Funding

University of Malaga

SINDH HIGHER EDUCATION COMMISSION

(RESEARCH)/SHEC/1-11/2021

HEC-NCRA-CMS Lab

2(1076)/HEC/M&E/2018/704

Licence