DETECÇÃO AUTOMÁTICA DE POÇOS DE VISITA ATRAVÉS DE DEEP LEARNING E RECONSTRUÇÃO AUTOMÁTICA DE REDE DE DRENAGEM

Abstract

The lack of information about the location and hydraulic characteristics of drainage networks makes it difficult to manage infrastructure assets. In this dissertation, a methodology for the automatic reconstruction of drainage networks is presented, which involves two steps. The first step is to automatically obtain the drainage elements. In this study, different Deep Learning methods (FasterRCNN, Cascade-RCNN, RetinaNet, ATSS, SABL, PAA, and VFNET) were compared for the detection of inspection wells. The second step is to consider the location of the inspection wells as nodes and automatically recreate the drainage network. The method was tested in the city of Campo Grande, MS, and the shape and topology of the reconstructed network were compared with the existing network, resulting in an excellent fit.