Aplicação de aprendizagem profunda para detecção automática de trincas em pavimentos

Abstract

The inventory of pavement defects is extremely important because it allows the responsible organizations to plan maintenance and repair actions on the pavement in order to allow the adequate level of efficiency and quality necessary for the full development of economic and social activities in the country. However, this data collection process can prove to be quite time consuming and costly. Therefore, some computer vision techniques can help to automate this collection process, also allowing a reduction in the degree of subjectivity of some evaluations. Therefore, the present research aims to present a proposal for automated detection of cracks in the pavement, considering that this surface defect rapidly evolves into holes, which is the most critical defect for road safety. The You Only Look Once (YOLO) algorithm was used for this activity, in its two most stable current versions, which are YOLOv3 and YOLOv4, the latter being the one that presented the best results, with an image dimension of 416x416 pixels and 8000 iterations, whose general metrics were: F1-score of 0.49, IoU of 41.08% and mAP@0.50 of 47.71%