Development and Evaluation of a Comprehensive Dataset for Pothole Depth Estimation of Indian Roads Using Smartphone Camera Approach

In this paper, we introduce “Road Anomaly Detection (RAD),” a comprehensive machine learning dataset collected in the city of Bengaluru, Karnataka, India. The key objective of this dataset is to accurately represent different kinds of Indian road damages. This dataset consists of 364 high-resolution videos and 10,000 labeled images with 40,368 instances of six different classes. It is a supervised learning dataset targeting road damages. Diverse data was collected using a smartphone mounted on the windshield, and meticulous preprocessing steps were applied to ensure its quality and veracity. This dataset underwent augmentation to produce 11,800 images with 52,568 instances for six classes, namely LMVs, HMVs, road damages, unsurfaced roads, pedestrians, and speed bumps. Our dataset can be used to make object detection and depth estimation models that can identify road damages. This data can be given to the authorities and can provide crucial information about the road surface, enabling autonomous vehicles to navigate and adapt to irregularities effectively which could help bring autonomous vehicles that are accustomed to Indian roads. However, the dataset comes with some challenges such as weather and time. The current state of available datasets for Indian roads is limited, which hinders the development of models tailored specifically to the unique terrain and challenges of Indian roads. We aim to address this gap and provide a standardized dataset that can enable the development of more relevant and effective models for road damage detection.