| With the rapid development of UAV photogrammetry and sensor technology,massive UAV aerial image data processing has become one of the current research hotspots.Whether the matching image pairs can be obtained quickly and effectively in massive image data sets has become a key problem in 3D reconstruction.The attitude angle of the images obtained by UAV low-altitude photogrammetry is large and the route arrangement is irregular,so the accurate camera information can not be obtained without camera calibration,which leads to the traditional aerial photogrammetry data processing methods gradually failing to meet the needs of rapid processing of massive image data.In addition,in recent years,the new data acquisition methods such as imitating flight,surrounding flight and close flight make the traditional image matching methods difficult to meet the demand.Incremental Structure From Motion(Incremental SFM)is one of the most commonly used algorithms.The method does not require precise camera parameters,and has high automation degree and precision of reconstruction results,so it is worth digging deep into its research value and application prospect.In order to solve the problem of fast processing of massive UAV image data,this paper studies the fast processing of image data from two aspects: fast extraction of image features and fast selection of image matching pairs,and uses incremental SFM algorithm to reconstruct sparse point cloud of UAV aerial images with different acquisition methods,focusing on improving the efficiency and robustness of incremental SFM algorithm.The main research contents as follows:(1)In view of the lack of effective feature extraction algorithm,the UAV image has certain stability in the face of various terrain conditions,and can quickly extract the features from massive image datasets,this paper designed two experiments for research.Firstly,four feature extraction algorithms,SIFT,SURF,ORB and Sift GPU,are used to carry out UAV image matching experiments,and the various algorithms are compared and analyzed in combination with relevant algorithm evaluation indexes.Secondly,the efficiency of image feature extraction is compared between SIFT and SIFTGPU algorithms.The experimental results show that the Sift GPU algorithm can greatly improve the efficiency of UAV image feature extraction and has a high accuracy of feature extraction,and it is suitable for the process of feature point extraction of large-scale UAV image(2)Aiming at the problem that the conventional vocabulary tree algorithm is unefficient and low accurate in extracting matching images of large-scale UAV,this paper proposes a word tree retrieval algorithm based on spatial distance weighting By using the relative position relationship between UAV images provided by POS data and studying the score distribution of image retrieval results.The experimental results show that this algorithm can effectively eliminate the redundant matching of image pairs to be matched,and improve the precision and recall rate of UAV image retrieval,which is suitable for the practical application scenarios of large-scale UAV matching image pair extraction.(3)The vocabulary tree retrieval algorithm based on spatial distance weighting proposed in this paper is applied to incremental 3D point cloud reconstruction.Through the scene 3D reconstruction experiment of UAV images acquired by different survey areas and different photography methods,the results not only prove the feasibility of the proposed algorithm,but also can effectively complete the 3D reconstruction of UAV images with different flying heights and terrain undulations.The given method can automatically and effectively process massive UAV image data,and has no requirement on image data collection methods,so it has strong applicability and scalability.And the given method greatly improving the processing efficiency of 3D reconstruction of large-scale scenes,and has certain application value in the practical problems of massive aerial photogrammetry data processing. |
PDF Full Text Request