Automated Real-Time Pothole Detection Using ResNet-50 for Enhanced Accuracy under Challenging Conditions

Abstract

The research paper aims to develop an automated, real-time pothole detection system using deep learning techniques, specifically the ResNet-50 architecture, to improve detection accuracy and system efficiency. The primary objective is to address the limitations of current systems, which often face challenges in detecting potholes under varying traffic and weather conditions, such as rain, high-speed traffic, or snow. Existing methods either lack real-time capabilities or fail to maintain consistent accuracy, requiring manual inspections or high computational resources. This research introduces a novel approach that combines image preprocessing techniques, such as noise reduction and contrast enhancement, with a deep learning model to achieve more accurate and reliable results in diverse environments. The methodology involved collecting a comprehensive dataset of road images, including various conditions like potholes under different lighting and weather scenarios. The data was preprocessed, and the ResNet-50 model was trained using transfer learning to reduce the training time while improving accuracy. The system was tested in both controlled environments and real-world scenarios, where it achieved a high detection accuracy of 94.5% during validation and maintained 95% accuracy in real-time deployment. However, slight performance drops were noted in more challenging situations, such as high-speed traffic and rain, where detection accuracy fell by approximately 10-15%. Despite these challenges, the system demonstrated significant improvements over existing models by reducing false positives and providing faster detection. The achievements of this research lie in creating a more practical and scalable pothole detection system that can be applied in real-time on roads, with future enhancements focused on improving performance in adverse conditions and integrating additional real-time communication features.