Date of Award
Summer 2020
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Civil, Construction, and Environmental Engineering
First Advisor
Bai, Yong
Second Advisor
Karshenas, Saeed
Third Advisor
Sheng, Wenhui
Abstract
Using deep learning technology to rapidly estimate depth information from a single image has been studied in many situations, but it is new in construction site elevation determinations, and challenges are not limited to the lack of datasets. This dissertation presents the research results of utilizing drone ortho-imaging and deep learning to estimate construction site elevations for excavation operations. It provides two flexible options of fast elevation determination including a low-high-ortho-image-pair-based method and a single-frame-ortho-image-based method. The success of this research project advanced the ortho-imaging utilization in construction surveying, strengthened CNNs (convolutional neural networks) to work with large scale images, and contributed dense image pixel matching with different scales.This research project has three major tasks. First, the high-resolution ortho-image and elevation-map datasets were acquired using the low-high ortho-image pair-based 3D-reconstruction method. In detail, a vertical drone path is designed first to capture a 2:1 scale ortho-image pair of a construction site at two different altitudes. Then, to simultaneously match the pixel pairs and determine elevations, the developed pixel matching and virtual elevation algorithm provides the candidate pixel pairs in each virtual plane for matching, and the four-scaling patch feature descriptors are used to match them. Experimental results show that 92% of pixels in the pixel grid were strongly matched, where the accuracy of elevations was within ±5 cm.Second, the acquired high-resolution datasets were applied to train and test the ortho-image encoder and elevation-map decoder, where the max-pooling and up-sampling layers link the ortho-image and elevation-map in the same pixel coordinate. This convolutional encoder-decoder was supplemented with an input ortho-image overlapping disassembling and output elevation-map assembling algorithm to crop the high-resolution datasets into multiple small-patch datasets for model training and testing. Experimental results indicated 128×128-pixel small-patch had the best elevation estimation performance, where 21.22% of the selected points were exactly matched with “ground truth,” 31.21% points were accurately matched within ±5 cm. Finally, vegetation was identified in high-resolution ortho-images and removed from corresponding elevation-maps using the developed CNN-based image classification model and the vegetation removing algorithm. Experimental results concluded that the developed CNN model using 32×32-pixel ortho-image and class-label small-patch datasets had 93% accuracy in identifying objects and localizing objects’ edges.