Article: Pothole detection and localisation from images using deep learning Journal: International Journal of Information and Decision Sciences (IJIDS) 2025 Vol.17 No.4 pp.357 - 370 Abstract: The existence of potholes threatens road safety and contributes to a significant portion of accidents worldwide. It takes a lot of work to constantly patch potholes and keep track of when new ones appear. Our goal in this work is to create a pothole detection system that would make it simpler to accurately detect potholes from images. The system can potentially save human lives and assist the government authorities to fix the potholes. In order to achieve this objective, we first make use of a pre-trained deep learning model (VGG-16) and thereafter, propose a novel convolutional neural network (CNN) model. This work employs a publicly available dataset, Nienaber Potholes 2 (Complex), for experiments. The proposed model provides 98.87% accuracy on pothole classification task in images and outperforms recent state-of-the-art approaches in the literature. Further, since no past work has been done on this dataset to detect bounding boxes for potholes, we use YOLO-v3 and YOLO-v5 to generate bounding box predictions on this dataset and evaluate the results. The bounding box task achieves 83.23% mAP and 87.45% precision. Due to the absence of significant existing results, these results for bounding box prediction may be considered as a benchmark. Inderscience Publishers - linking academia, business and industry through research

Title: Pothole detection and localisation from images using deep learning

Authors: Archit Dhiman; Mohit Kumar; Arun Kumar Yadav; Divakar Yadav

Addresses: Department of Computer Science and Engineering, NIT Hamirpur (HP), India ' Department of Computer Science and Engineering, NIT Hamirpur (HP), India ' Department of Computer Science and Engineering, NIT Hamirpur (HP), India ' SOCIS, IGNOU, New Delhi, India

Abstract: The existence of potholes threatens road safety and contributes to a significant portion of accidents worldwide. It takes a lot of work to constantly patch potholes and keep track of when new ones appear. Our goal in this work is to create a pothole detection system that would make it simpler to accurately detect potholes from images. The system can potentially save human lives and assist the government authorities to fix the potholes. In order to achieve this objective, we first make use of a pre-trained deep learning model (VGG-16) and thereafter, propose a novel convolutional neural network (CNN) model. This work employs a publicly available dataset, Nienaber Potholes 2 (Complex), for experiments. The proposed model provides 98.87% accuracy on pothole classification task in images and outperforms recent state-of-the-art approaches in the literature. Further, since no past work has been done on this dataset to detect bounding boxes for potholes, we use YOLO-v3 and YOLO-v5 to generate bounding box predictions on this dataset and evaluate the results. The bounding box task achieves 83.23% mAP and 87.45% precision. Due to the absence of significant existing results, these results for bounding box prediction may be considered as a benchmark.

Keywords: pothole detection; pothole; convolutional neural network; CNN; bounding box; you only look once; YOLO; Nienaber.

DOI: 10.1504/IJIDS.2025.150097

International Journal of Information and Decision Sciences, 2025 Vol.17 No.4, pp.357 - 370

Received: 06 Sep 2022
Accepted: 04 Aug 2023
Published online: 01 Dec 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Pothole detection and localisation from images using deep learning

Keep up-to-date