Citation

Reddy, Shiva Shankar and Janarthanan, Midhunchakkaravarthy and Khan, Inam Ullah and Amrutha, Kankanala (2026) A Deep Learning Framework for Real-Time Pothole Detection from Combined Drone Imagery and Custom Dataset Using Enhanced YOLOv8 and Custom Feature Extraction. Mathematics, 14 (5). p. 898. ISSN 2227-7390

Text mathematics-14-00898.pdf - Published Version Restricted to Repository staff only Download (2MB)

Abstract

Road safety depends heavily on the timely identification and repair of potholes; however, detecting potholes is challenging due to various lighting and weather conditions. This work presents an attention-enhanced object detection framework for aerial pothole detection design that relies on a pre-trained backbone, YOLOv8, and a custom feature-extraction network, the Feature Pyramid Network (FPN). An enhanced detection head is used to make the model aware of discriminative areas in space to get accurate localization of a pothole to overcome the major limitations of the standard YOLOv8 used in aerial road inspection, irrespective of the road surface. The underlying architecture incorporates a purpose-built data layer and a preprocessing engine that can accommodate scenarios such as seasonal changes and bad weather. To further enhance learning dynamics, a customized loss function and a new optimizer framework are incorporated to improve convergence towards overall detection reliability. Specifically, a custom differential optimizer that uses layer-wise adaptive learning rates and momentum-based gradient updates to help suppress false positives and accelerate convergence. Conversely, the IoU-based personal loss function, combined with real-time validation, stabilizes training across a range of road conditions. A major feature of the proposed system is its ability to process aerial imagery from unmanned drone platforms. Empirical analysis proves a good result: an average precision of 0.980 with the IoU of 0.5 and an F1-score of 0.97 with a confidence threshold of 0.30. Precision is high (0.97 at the 90-percent confidence level). These metrics show how well the model will be able to balance false positives and false negatives—a critical need in a safety-critical deployment. The results make the framework a potential, scalable, and reliable candidate for integrating smart transportation systems and autonomous vehicle navigation.

Item Type:	Article
Uncontrolled Keywords:	Feature Pyramid Network (FPN), drone images, labelImg
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines
Divisions:	Faculty of Computing and Informatics (FCI)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	02 Apr 2026 02:34
Last Modified:	02 Apr 2026 02:34
URII:	http://shdl.mmu.edu.my/id/eprint/15618

Downloads

Edit (login required)