Uma aplicação de visão computacional para detecção automática de defeitos em pavimentação asfáltica urbana no Brasil

Abstract

Road infrastructure maintenance is a cornerstone of well-functioning cities, as it directly influences urban mobility, traffic safety, and the efficiency of transporting people and goods. However, the natural deterioration of pavements—compounded by increasing traffic and the lack of adequate preventive maintenance—leads to recurrent damage that affects both users and public management. In this context, efficient monitoring methods are indispensable to ensure roadway quality and to reduce costs associated with emergency interventions. This work aims to develop an automated system for detecting damage in asphalt pavements using computer vision and deep learning techniques. Among the principal challenges addressed are the accurate identification and classification of different defect types: fissures (cracks), deformations, ravelling (surface wear), potholes, and patches. We present a publicly available dataset of high-resolution digital images of urban pavements, structured in accordance with Brazilian standards and accompanied by per-image annotations curated by a technical committee. The study also proposes an efficient methodology for automated roadway condition assessment based on images collected with a smartphone mounted on the bumper of a passenger vehicle. The model trained with YOLOv5 achieved an mAP@50 of 84.4%. These results are intended to support public administrations in making decisions about maintenance and interventions, leading to improvements in the quality and safety of road infrastructure.