Construction Site Segmentation Using Drone-Based Ortho-Image and Convolutional Encoder-Decoder Network Model

Document Type

Conference Proceeding

Publication Date

2022

Source Publication

Construction Research Congress 2022

Source ISSN

9780784483985

Original Item ID

DOI: 10.1061/9780784483961.115

Abstract

This paper presents a convolutional encoder-decoder network model for pixelwise segmentation of drone-based ortho-images of construction sites. The input ortho-image disassembling and output label-image prediction assembling algorithms are supplemented with the encoder-decoder to process high-resolution inputs. Parameter analyses were conducted to evaluate the performances between differently sized small-patches and different model training epochs. Testing results showed using the 64 × 64-pixel patch with 100-epoch can produce a well-trained encoder-decoder for pixelwise segmentation with a pixel accuracy of 0.98 in validation and 0.93 in testing, and had weighted average results of the precision, recall and f1-score larger than 0.98 in validation, and larger than 0.93 in testing. The developed method was also applied to object detection in drone photogrammetric orthophoto for objects’ contour extraction, then the extracted contours were used for as-built modeling in AutoCAD. The results of this work can benefit construction professionals in automatic as-built modeling and earthwork estimation if they have the corresponding elevations.

Comments

Published in the proceedings of the conference Construction Research Congress 2022, March 9-12, 2022: 1096-1105. DOI.

Link to Full Text

Share

COinS