Detecção de pavimento e análise de qualidade de sinalização horizontal na Rodovia BR-163

Abstract

This work investigated computational solutions for pavement detection and analysis of horizontal road marking quality on the BR-163 highway, a strategic infrastructure for the state of Mato Grosso. The research, carried out in partnership with the concessionaire Nova Rota do Oeste, proposed a hybrid approach that integrates deep learning and classical image processing. The methodology involved the use of the YOLOv11-Seg model, trained with different datasets captured by the concessionaire, ranging from small manually labeled sets to 265,000 automatically pre-labeled images. The results showed precision and recall metrics close to 1.0 in all trainings, confirming the robustness of the model and the feasibility of large-scale automatic labeling. External validation with images from a highway in the direction of the state of São Paulo reinforced its generalization capacity. Complementarily, the Canny algorithm was applied to the segmented regions, allowing the measurement of marking intensity and classification into good, medium, and poor. This analysis objectively confirmed the differences observed manually. In addition to the academic contribution, the model was incorporated by the Artificial Intelligence team of Nova Rota do Oeste to isolate the region of interest (ROI) of the pavement before analyzing the markings, demonstrating immediate practical impact on road management.