Practical implementation of pothole detection and dimension estimation using YOLOv8 and Mask R-CNN with a Flutter-based mobile application

Abstract

Road surface deterioration, particularly pothole formation, poses substantial threats to traffic safety and driving comfort, highlighting the need for timely and objective road condition assessment. However, past studies have predominantly focused on pothole detection without estimating pothole dimensions, while conventional inspection methods remain manual, inefficient, and subjective. To address these limitations, a deep-learning-based framework is proposed herein for joint pothole detection and dimension estimation using monocular road images. The proposed framework integrates the You Only Look Once (YOLO)v8 and Mask Region-Based Convolutional Neural Network (Mask R-CNN) models to detect potholes and estimate their geometric dimensions from road surface images. Image data were acquired using a GoPro Hero 8 camera across multiple road segments in Sumatra Selatan, Indonesia. The models were trained on Google Colaboratory using the YOLOv8x-seg architecture and subsequently integrated into a Flutter-based mobile application for practical deployment, while Leaflet JS was employed to interactively visualize pothole locations. Experimental results demonstrated detection confidence levels exceeding 92.22%, with stable performance observed on both local computing platforms and mobile devices. Application-level evaluations further confirmed robust cross-device functionality and yielded a usability score of 4.8875. These results indicate that the proposed approach provides an accurate and reliable solution for automated road condition monitoring, supporting more objective and data-driven road maintenance planning.