首页|全球导航卫星系统拒止条件下无人机影像快速级联检索方法

全球导航卫星系统拒止条件下无人机影像快速级联检索方法

扫码查看
提出一种全球导航卫星系统(GNSS)拒止条件下的无人机影像快速级联检索方法.基于影像的全局特征,利用哈希小波方法,根据相似度对无人机和卫星影像进行初次检索、筛选和重排序;根据SuperPoint网络设计轻量化特征提取网络(ISFFRN),基于一次检索结果和影像局部特征进行二次检索和排序,以精准筛选目标影像.选用NVIDIA Jetson AGX Orin模拟无人机机载处理设备,在公开的Inria航空影像数据集上开展实验验证.实验结果表明,所提出的级联检索方法在实验区域内能够高效、准确地完成异源影像的快速检索任务,检索过程耗时少于0.5 s,具备满足机载GNSS设备射频广播频率的实时化应用潜力.
Fast Cascade Retrieval Method for UAV Images with GNSS Rejection
Objective Our research is of great importance for improving the autonomous positioning ability of UAVs,advancing the development of remote sensing image technology,and enhancing the widespread application of UAVs.Simultaneously,it addresses the challenge of GNSS signal interference,enhances the accuracy and efficiency of image retrieval,and meets the real-time processing needs of UAVs.In complex environments where GNSS signals are disrupted or rejected,traditional positioning methods prove ineffective.The fast cascade retrieval method we proposed significantly strengthens retrieval accuracy and efficiency by comprehensively utilizing local image features and adopting a two-level retrieval strategy.The method demonstrates feasibility and practicality in real-time processing,which provides reliable positioning support for UAVs in surveying and mapping,firefighting,agriculture,transportation,rescue,and military applications.This research holds crucial practical significance and immediate necessity.Methods We propose a fast cascade retrieval method for UAV images under GNSS rejection conditions,which achieves rapid positioning through a two-stage retrieval process(Fig.1).Initially,a subset of aerial image datasets from Vienna,Austria,and Chicago,USA,along with corresponding commercial satellite image data,is selected for the experiment.To ensure data consistency and comparability,these images are cropped and size-normalized for subsequent retrieval.Subsequently,hash codes and feature point data are extracted from satellite images to construct hash and feature point databases,respectively,thereby improving retrieval efficiency and avoiding redundant calculations.In the initial retrieval stage,the wavelet hashing method is employed by incorporating spatial information.This method extracts low-frequency spatial components from the image using wavelet transform,generates hash codes,and utilizes these codes for rapid retrieval.The specific steps are as follows.First,the satellite image undergoes wavelet transformation to extract low-frequency components,and then the resulting hash codes are stored in the hash database.Similarly,the UAV image is processed to generate corresponding hash codes,which are used for rapid matching to identify candidate sets of satellite images similar to that of UAVs.In the secondary retrieval stage,an improved superpoint fast feature retrieval network(ISFFRN,Fig.2)is employed to further retrieve and rank images based on local features of both UAV and satellite images.The steps involve using the ISFFRN to extract local feature points from UAV and candidate satellite images,followed by matching these feature points and calculating the number of matched pairs.The retrieval results are then sorted based on the number of matching point pairs to determine the final retrieval outcomes.The entire experiment is conducted using the Jetson AGX Orin hardware device to evaluate the method's feasibility and real-time performance on an actual airborne platform.Results and Discussions The fast cascade retrieval method for UAV images with GNSS rejection proposed in this paper performs well on Jetson AGX Orin hardware devices,with an average processing time controlled within 0.5 s to meet the requirements of real-time positioning tasks.In the first retrieval stage,the wavelet hashing method is used to generate hash codes by extracting low-frequency spatial information from the image,and a candidate set of similar satellite images is quickly selected through matching(Figs.4 and 5).Experimental results demonstrate that the wavelet hashing algorithm effectively captures global features and texture information from the image,which exhibits good adaptability and robustness.In the second retrieval stage,the ISFFRN is employed to determine the final retrieval results by extracting local feature points from UAV images and candidate satellite images,followed by matching and sorting.Experiments show that the ISFFRN completes the retrieval based on the ToCP@1 index(Figs.6 and 7),confirming the feasibility of the ISFFRN for heterogeneous image retrieval(Fig.8).Regarding performance evaluation,the entire experimental process is conducted on image data from Vienna,Austria,and Chicago,USA,utilizing Jetson AGX Orin hardware equipment,with experiment times recorded(Tables 4 and 5).Experimental results show that the cascade retrieval method can successfully achieve fast retrieval of UAV images.In the first search,ToCP@1 is 0,which indicates that the correct UAV image is not directly matched by the first satellite image after sorting the search results.However,ToCP@6 is 50%,which indicates that the correctly matched satellite image ranks second among the first six retrieved images.Although it does not achieve direct matching initially,it significantly narrows down the retrieval range and successfully locates the target image in subsequent results.In the second retrieval,ToCP@1 is 100%following the screening from the first retrieval.This demonstrates the cascade retrieval method's effectiveness in accurately identifying and retrieving the correct satellite image,which ranks first in the similarity ranking.In summary,the fast cascade retrieval method for UAV images under GNSS rejection conditions achieves real-time airborne retrieval of UAV images by leveraging both global and local image features.The method comprehensively considers the characteristics of multi-source remote sensing images and effectively addresses the challenge of fast UAV image retrieval in GNSS-rejected environments.Conclusions Aiming at the problem of UAV positioning with GNSS rejection,we propose a fast cascade retrieval method that combines global and local features of satellite and UAV images.Its feasibility is verified using airborne equipment.Experimental results demonstrate that the two-stage cascade retrieval method achieves an average processing time of under 0.5 seconds on Jetson AGX Orin hardware,which meets the real-time requirements for GNSS RF and makes it suitable for real-time positioning of airborne equipment.In the first retrieval stage,the wavelet hash image retrieval algorithm,which integrates spatial information,shows robustness under the ToCP@6 index.However,adjustments to the hash length are necessary to ensure accuracy when dealing with images of varying scales.In the secondary retrieval,the ISFFRN successfully completes retrieval under the ToCP@1 index,thus confirming the feasibility of ISFFRN in heterogeneous image retrieval.To conclude,by leveraging both global and local image features,our method achieves real-time airborne retrieval of UAV images,which effectively tackles the challenge of fast UAV image retrieval in GNSS-denied environments.It exhibits high accuracy and robustness,meeting experimental expectations.

global navigation satellite system rejectionremote sensing image retrievalheterogeneous remote sensing imageunmanned aerial vehicle

高寒、于英、李力、李磊、宋亮、张磊

展开 >

信息工程大学地理空间信息学院,河南 郑州 450001

231016 部队,北京 100000

全球导航卫星系统拒止 遥感影像检索 异源遥感影像 无人机

2024

光学学报
中国光学学会 中国科学院上海光学精密机械研究所

光学学报

CSTPCD北大核心
影响因子:1.931
ISSN:0253-2239
年,卷(期):2024.44(24)