UAV image target localization method based on outlier filter and frame buffer  

作　　者：Yang WANG Hongguang LI Xinjun LI Zhipeng WANG Baochang ZHANG 

机构地区：[1]Institute of Unmanned System,Beihang University,Beijing 100191,China [2]School of Electronics and Information Engineering,Beihang University,Beijing 100191,China [3]Institute of Artificial Intelligence,Beihang University,Beijing 100191,China

出　　处：《Chinese Journal of Aeronautics》2024年第7期375-390,共16页中国航空学报（英文版）

基　　金：financial support from the National Key Research and Development Program of China (Nos. 2022YFB3904303 and 2020YFB0505602);the National Natural Science Foundation of China (Nos. 62076019, 62022012, U2233217, 62101019 and 62371029);the Civil Aviation Security Capacity Building Fund Project, China (Nos. CAAC Contract 2020(123), CAAC Contract 2021(77) and CAAC Contract 2022(110))

摘　　要：With rapid development of UAV technology,research on UAV image analysis has gained attention.As the existing techniques of UAV target localization often rely on additional equipment,a method of UAV target localization based on depth estimation has been proposed.However,the unique perspective of UAVs poses challenges such as the significant field of view variations and the presence of dynamic objects in the scene.As a result,the existing methods of depth estimation and scale recovery cannot be directly applied to UAV perspectives.Additionally,there is a scarcity of depth estimation datasets tailored for UAV perspectives,which makes supervised algorithms impractical.To address these issues,an outlier filter is introduced to enhance the applicability of depth estimation networks to target localization.A frame buffer method is proposed to achieve more accurate scale recovery,so as to handle complex scene textures in UAV images.The proposed method demonstrates a 14.29%improvement over the baseline.Compared with the average recovery results from UAV perspectives,the difference is only 0.88%,approaching the performance of scale recovery using ground truth labels.Furthermore,to overcome the limited availability of traditional UAV depth datasets,a method for generating depth labels from video sequences is proposed.Compared to state-of-the-art methods,the proposed approach achieves higher accuracy in depth estimation and stands for the first attempt at target localization using image sequences.Proposed algorithm and dataset are available at https://github.com/uav-tan/uav-object-localization.

关 键 词：Object localization Deep learning Depth estimate Scale recovery Unmanned Aerial Vehicle(UAV) 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] V279[自动化与计算机技术—计算机科学与技术]