An annotated street view image dataset for automated road damage detection

Miao Ren,Xianfeng Zhang,Xiaobo Zhi,Yuanjia Wei,Ziyuan Feng

DOI: https://doi.org/10.1038/s41597-024-03263-7

2024-04-22

Scientific Data

Abstract:Road damage is a great threat to the service life and safety of roads, and the early detection of pavement damage can facilitate maintenance and repair. Street view images serve as a new solution for the monitoring of pavement damage due to their wide coverage and regular updates. In this study, a road pavement damage dataset, the Street View Image Dataset for Automated Road Damage Detection (SVRDD), was developed using 8000 street view images acquired from Baidu Maps . Based on these images, over 20,000 damage instances were visually recognized and annotated. These instances were distributed in five administrative districts of Beijing City. Ten well-established object detection algorithms were trained and assessed using the SVRDD dataset. The results have demonstrated the performances of these algorithms in the detection of pavement damages. To the best of our knowledge, SVRDD is the first public dataset based on street view images for pavement damages detection. It can provide reliable data support for future development of deep learning algorithms based on street view images.

multidisciplinary sciences

What problem does this paper attempt to address?

The paper aims to address the issue of road damage detection. Specifically, it proposes a road damage detection dataset based on street view images—Street View Image Dataset for Automated Road Damage Detection (SVRDD), to promote the development of automated road damage detection technology.

### Main Objectives:
1. **Develop a New Dataset**: Create a dataset containing 8,000 street view images from Baidu Maps for the automatic detection of road damage.
2. **Annotate Diverse Damage Types**: Annotate over 20,000 damage instances in these images and classify them into six major damage types and one non-concrete pavement category.
3. **Evaluate Multiple Algorithm Performances**: Train and evaluate ten mainstream object detection algorithms using the SVRDD dataset, including Faster R-CNN, YOLOv5, etc., and analyze their performance with different numbers of training samples.

### Background and Significance:
- **Limitations of Traditional Methods**: Currently used manual visual inspection and vehicle-mounted road monitoring systems (PMS) are time-consuming, inefficient, and disruptive to traffic.
- **Insufficiency of Existing Datasets**: Existing public datasets mainly rely on aerial or wide-angle images, which have certain limitations in terms of data volume, resolution, and perspective.
- **Advantages of Street View Images**: Compared to other types of images, street view images have the advantages of wide global coverage, easy accessibility, and regular updates, making them suitable for road damage detection.

By introducing a new dataset based on street view images, this study aims to advance the technology of automated road damage detection and provide reliable data support for future deep learning-based road detection algorithms.