Reconstruction of 3D Environments from UAV’s Aerial Video Feeds

Abstract

Graphical 3D models of a real-world scene could contribute to the knowledge, better understanding, and further investigations for a variety of scenarios such as accident sites, historical sites, construction progress monitoring, etc. This study proposed a proof of concept prototype for low-cost offline 3D reconstruction using 2D video frames obtained by a UAV using the method of Structure from Motion (SfM). The work presented consists of novel approaches for video frame selection and feature matching in the reconstruction pipeline to improve the accuracy, robustness, and efficiency which are also considered as major challenges in 3D reconstruction. The introduced video frame selection stage to extract frames is based on the affine transformation between views and it is to obtain an optimal set of video frames to be processed with the required amount of feature points while maintaining the performance. Introduced feature matching stage with improved and a novel algorithm is presented with the underline assumption of spatial cohesion which is the consecutive video frames have higher matchings between them. The algorithm is designed to perform feature matchings between consecutive frames to avoid unnecessary processing thus improving the performance and final output. The feature matching stage consists of multiple filtering steps to remove noise and outliers among matched features to obtain robust matches. The evaluation results show that the improved feature matching algorithm has a significant performance improvement in terms of time and memory usage as well as the increased final output model accuracy. The evaluation results of Hausdorff distance comparisons between ground truth models and the output 3D models of the implemented prototype and between the output of existing 3D reconstruction tools show that the implemented prototype outperforms the existing tools.