Depth Prediction-Based Enhancement of Feature Matching in Urban Drone Imagery Geonseok Lee (a), Junhee Youn (b), Kanghyeok Choi (c*)
a) Program in Smart City Engineering, Inha University, Incheon 22212, Republic of Korea
b) Department of Future & Smart Construction Research, Korea Institute of Civil Engineering and Building Tech, Gyeonggi-do 10223, Republic of Korea
c) Department of Geoinfomatic Engineering, Inha University, Incheon 22212, Republic of Korea
*cwsurgy[at]inha.ac.kr
Abstract
Recent drone imagery has been actively utilised across various domains, including disaster monitoring and geospatial information acquisition. In particular, its application has been expanding in urban areas for purposes such as urban monitoring and construction site management. To effectively leverage drone imagery, the role of feature matching algorithms, which identify accurate correspondences between images, is critically important. These algorithms serve as fundamental components in 3D reconstruction and trajectory estimation techniques, such as Structure from Motion (SfM) and Visual Simultaneous Localization and Mapping (Visual SLAM). However, complex urban environments, characterisedby densely clustered high-rise buildings, abrupt terrain variations, and repetitive structural patterns, significantly degrade the accuracy and robustness of conventional 2D-based feature matching algorithms. Even with the incorporation of outlier rejection techniques such as Random Sample Consensus (RANSAC), mismatches frequently occur, adversely affecting the precision of downstream tasks such as 3D modelling and localisation. To address these limitations, this study proposes a novel feature matching enhancement method by introducing 3D spatial constraints through the integration of deep learning-based depth estimation into existing feature matching pipelines. Experimental scenarios were designed to reflect real-world urban complexities, including dense built-up areas, waterfront ecological zones, and road intersections, with considerations for variations in elevation and lighting conditions. The proposed approach demonstrated significant improvements in matching precision and geometric consistency compared to traditional feature matching algorithms. This research presents a methodology for improving feature matching performance in complex urban environments, and it is anticipated to serve as a foundational technology for high-precision applications such as large-scale map generation and urban spatial analysis.
