Automated volumetric damage detection and quantification using region-based convolution neural networks and an inexpensive depth camera.

摘要

Structural health monitoring has become an outstanding tool to perform structural condition assessments, once performed solely by trained experts. In this study, a methodology utilizing an inexpensive depth sensor to detect and quantify volumetric damages within concrete surfaces is proposed. To allow automatic damage detection, a Faster Region-based Convolutional Neural Network (Faster R-CNN)-based method is implemented. A database of 444 images with resolution of 853 × 1440 pixels annotated for concrete spalling is developed. The network is modified, trained and validated using the proposed database. Damage quantification is automatically performed using the depth data output by the sensor. The surface of the analyzed element is extracted by merging the bounding boxes output by the Faster R-CNN onto the depth map. A polystyrene test rig containing damage simulations of known volume was utilized to test the accuracy of volume calculation. In addition to that, a concrete beam was also used to test the entire system. The Faster R-CNN yielded an average precision (AP) of 77.97% for damage detection. Damage quantification routine presents error of 9.45% in volume quantification of samples located within 100 cm and 250 cm away from the sensor plane. On top of that, maximum depth measurements of damages show a mean precision error (MPE) of 3.24% considering the same distance range. The implemented method allows for damage segmentation and quantification regardless of the distance between the sensor and the analyzed element.