结合核相关滤波器和深度学习的运动相机中无人机目标检测  

UAV detection in motion cameras combining kernelized correlation filters and deep learning

作　　者：梁栋[1,2,3] 高赛 孙涵 刘宁钟[1,2,3] LIANG Dong;GAO Sai;SUN Han;LIU Ningzhong(College of Computer science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;MIIT Key Laboratory of Pattern Analysis and Machine Intelligence,Nanjing 211106,China;Collaborative Innovation Center of Novel Software Technology and Industrialization,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京211106 [2]模式分析与机器智能工业和信息化部重点实验室,南京211106 [3]软件新技术与产业化协同创新中心,南京211106

出　　处：《航空学报》2020年第9期154-166,共13页Acta Aeronautica et Astronautica Sinica

基　　金：国家自然科学基金(61601223);国防科技创新特区项目。

摘　　要：针对无人机与相机快速相对运动造成的运动模糊问题,以及小型无人机外观信息缺失和背景复杂造成漏警和虚警问题,提出了一种新的无人机检测-跟踪方法。针对成像尺寸小于32像素×32像素的无人机目标,提出改进的多层特征金字塔的分类和目标框回归器作为目标检测器,克服漏警。利用检测结果初始化基于核相关滤波的目标跟踪器,并持续修正跟踪结果,跟踪结果为剔除检测器虚警提供依据。在跟踪过程中,引入对观测场景纹理自适应的相机运动补偿策略实现目标重定位。多场景下的实验结果表明:提出的方法在对高速运动小目标的检测和跟踪指标上显著优于传统方法,且运动补偿机制的引入进一步增强了方法在极端复杂场景下的鲁棒性。To solve the problem of motion blurs caused by the rapid relative movement of the UAV and the camera,and the missed detection and false detection problems resulted from the lack of appearance information and complex background of small drones,a new drone detection-tracking method is proposed.Aiming at UAV targets with imaging sizes less than32 pixel×32 pixel,an improved multi-layer feature pyramid classification and a target box regressor are proposed as target detectors to overcome missed detection.The detection result is used to initialize the target tracker based on kernelized correlation filters,and continuously modify the tracking result which provides a basis for the elimination of false detection.During the tracking process,a camera motion compensation strategy adaptive to the observed scene texture is introduced to achieve target relocation.Experimental results in multiple scenarios show that the proposed method is significantly better than traditional ones in the detection and tracking of small high-speed moving targets.In addition,the introduction of motion compensation mechanism further enhances the robustness of the method in extremely complex scenarios.

关 键 词：多层特征金字塔 核相关滤波器 相位相关 随机抽样一致 光流场 

分 类 号：V249[航空宇航科学与技术—飞行器设计] TP391[自动化与计算机技术—计算机应用技术]