International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 06 | Jun 2025
p-ISSN: 2395-0072
www.irjet.net
A Novel Approach to Pothole Detection Using RT-DETR for Smart Road Maintenance Ashwini Bhosale1, Rohit Deokate2, Atharva Pawar3, Parth Sawant4, Nikee Kumar5, Nikhil Mhaske6 1,2,3,4,5,6 Department of Computer Engineering,
JSPM’s Rajarshi Shahu College of Engineering, Tathawade, Pune-411033, Maharashtra, India, 1,2,3,4,5,6 JSPM’S RAJARSHI SHAHU COLLEGE OF ENGINEERING Pune, India ----------------------------------------------------------------------***-------------------------------------------------------------------
Abstract -Potholes pose a significant issue for road safety, contributing to numerous accidents and being extremely costly to repair annually. Traditional methods we currently discover potholes, such as individuals visualizing or having sensors, are usually slow, not great, and prone to errors. To address this problem, we are proposing a novel, intelligent way RT-DETR (Real-Time Detection Transformer). This technology employs special Transformer networks that are capable of rapidly and extremely accurately detecting objects even if it is dark or the weather is inclement. What's unique about RT-DETR is that it can actually pay close attention to what matters that is potholes and not worry about extraneous background noise, so there are fewer false positives. It employs something known as self-attention mechanisms that actually locate potholes with incredible precision. As soon as we implemented it, it scored an astonishing 92% accuracy (0.92 mAP at 0.5 IoU), showing that it works great on real-world scenarios. This proves that using Transformer-based Pothole Detection can be very advantageous for road care and intelligent transportation. By real-time detection of potholes, RT-DETR can help prevent accidents, enhance roads, and streamline road maintenance. The combination of deep learning and transformers with real-time processing makes it an extensive solution for road safety that can be used on a large scale, operates effectively, and is cost-effective. This work makes its contribution to the development in smart cities and smart transport systems to render our roads safer and more effective. Keywords: Potholes, Accidents, RT-DETR, Transformers, Self attention mechanism, Real-Time Processing, Machine Learning, Computer Vision, Smart Cities, Object Recognition, Data Augmentation, Feature Extraction
1. INTRODUCTION Potholes are a common road problem all over the globe, causing accidents, damage to vehicles, and expensive repairs. Pothole detection in an early phase is significant in order to maintain road safety and maintain smooth traffic flow. Conventional pothole detection techniques like manual inspection are time consuming, require immense human efforts, and are prone to human errors. With the latest development of computer vision and deep learning, computerized pothole detection systems are now at the forefront as speedier and more accurate substitutes. YOLO (You Only Look Once) is among the popular deep learning architectures used for real-time object detection. YOLO can detect an entire image in a single pass because it is faster and computation-efficient than the conventional techniques like R-CNN. Nevertheless, YOLO has its drawbacks in picking up fine details and complex structures, especially under conditions of varying road surfaces. RT-DETR (Real-Time Detection Transformer) is applied here. RT-DETR is a Transformer object detection model that improves performance and accuracy by steering clear of region proposal networks. Unlike traditional CNN-based models, RT-DETR performs well in capturing contextual information and long-range dependencies and, therefore, is highly suitable for pothole detection in different conditions. It utilizes the strength of self-attention mechanisms such that it can pay attention to the important features of potholes while suppressing background noise. This research delves into the strengths of RT-DETR compared to traditional models for pothole detection. It tackles the usual problems like variations in lighting, varying road textures, occlusions, and varying vehicle speeds. By combining deep learning, Transformers, and real-time processing, RT-DETR provides a scalable, accurate, and efficient solution to pothole detection. The aim is to improve road maintenance practices, reduce accidents, and help build safer transport systems by leveraging the latest AI-based solutions.
2. RELATED WORK [1] Ping et al. states YOLO V3, SSD, HOG, and SVM being used in conjunction with Faster R-CNN, a deep learning-based approach for detecting street potholes is presented in this study. The main emphasis is on evaluating their accuracy and
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 516