A speed-breaker image dataset, INR-SPEED-BUMP of Indian road surfaces in uncontrolled environment and its detection using Deep Learning

Road accidents are increasing worldwide, resulting in numerous deaths and disabilities each year all over the world. So road safety has become one of the major public safety concerns in a country like India. Sometimes such road accidents occur because drivers are unable to detect of road anomalies such as potholes, speed-breakers, etc on road surfaces in time and consequently the driver cannot reduce or control the speed of the vehicle at right time. This study aims to develop a computer vision-based model to assist in detecting the road anomalies and thereby contribute to reducing road accidents. In this direction, as a first step, it builds a new dataset of bump images of Indian road surfaces. All the images of national/state highways, local roads were captured under uncontrolled environmental and lighting conditions. This dataset comprises of high resolution road surface images captured in different states of India under diverse weather and lighting conditions. It specifically tries to capture the complexities and variations of real-world Indian road surfaces. The image dataset has a total of 9323 images comprising 6028 bump images and 3295 non-bump(normal) image. Moreover bump images cover a total of 24 categories and non-bump (normal) images cover 9 categories. After collection of images, in the next step, it has attempted a few Deep Learning hybrid models (MobileNetv2, EfficientNetB0, ResNet-50) are utilized for speed-breaker detection using the dataset. Also, the study leveraged specialized multi-scale feature extraction network Feature Pyramid Network, lightweight attention mechanism CBAM, feature aggregation module Horizontal Pyramid Pooling. Experimental results show that the hybrid pre-trained models achieves up to 97.75% accuracy on INR-SPEED-BUMP image dataset. To improve the interpretability and reliability of the model predictions, two Explainable Artificial Intelligence (XAI) techniques, namely Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP), are utilized. The image dataset of Indian road surface may serve as a valuable resource for academicians and researchers and contribute significantly for the development of computer vision-based system for advancing of road safety technologies in the Indian subcontinent.