The construction industry is introducing autonomous heavy equipment to overcome labor shortages and improve productivity. For autonomous heavy equipment to work on earthmoving at sites, the equipment needs to recognize and understand ground surface types. However, the ground surface types are manually inspected in practice, and related studies are lacking. To address this issue, the authors developed and tested models that automatically classify ground surface types from images acquired by an unmanned aerial vehicle using a deep learning-based multi-label classification method that applies Binary Relevance (BR) and Label Powerset (LP) methods with Residual Neural Network (ResNet) and Vision Transformer classification network (VIT). The model performances were comparatively evaluated through experiments conducted on actual construction sites. The results showed that the BR model with ResNet is the best model in terms of automated ground surface type identification during earthmoving. The results are expected to broaden the understanding of complex and expansive construction sites for autonomous vehicles and thus facilitate deployment of autonomous heavy equipment by helping them to understand working areas and any obstacles on construction sites quickly and effectively, which will reduce the cost and time needed for on-site ground surface management.
UAV Imagery-based Automatic Classification of Ground Surface Types for Earthworks
KSCE J Civ Eng
KSCE Journal of Civil Engineering ; 28 , 6 ; 2121-2131
2024-06-01
11 pages
Aufsatz (Zeitschrift)
Elektronische Ressource
Englisch
Ground stabilisation on steeply sloping earthworks
British Library Online Contents | 1994
|Online Contents | 1997
Online Contents | 1997
Earthworks, Drainage, Subgrade (C12)
British Library Conference Proceedings | 1996
|Europäisches Patentamt | 2023
|