Towards Generation of Effective 3D Surface Models from UAV Imagery Using Open Source Tools

There has been increasing popularity in large scale mapping for deriving 3D surface and elevation models of earth and building structures. The techniques of computer vision comprising feature detections and matching and photogrammetry play an important role in deriving near accurate 3D reconstruction of scenes from 2D images. Since the images captured by the unmanned aerial vehicle (UAVs) are of high resolution, there is need for more sophisticated processing and analysis of the imagery to generate 3D models and other useful imagery products. The open source softwares are excellent tools for research and can be modified or changed to suit our model, as specific or combinations of algorithms behave differently based on the nature of UAV image scene to be processed. Though many algorithms are available for performing feature extractions from images, few studies have been carried out to identify suitable detector algorithms to be used based on the nature of image or scene that the UAV captures. An attempt has been made to understand and analyse the suitability of feature detection and descriptor algorithms for different scene types. This article also describes the popular technique called structure from motion process pipeline for sequential processing of UAV images with high overlapping, which involves the estimation of 3D point clouds from the keypoint correspondences. The relative accuracy of the 3D point cloud derived from our approach is comparable with similar output from other state-of-the-art UAV processing systems and is found to match with high precision.