球面全方位目标测量及跟踪理论

In recent years , due to the development of the robot navigation technology and the virtual reality technology,the demand of achieving large field of view and even three-dimensional perception has been growing. Confronted with this situation, we create this project aiming to set up the full automatic target recognition theory based on the spherical imaging measurement theory.. Firstly we need to creat rhe basic theory of the spherical imaging measurement theory,including the spherical projection imaging mathematics model, spherical fundamental matrix and nuclear line constraint equations theory, the Homograph relationship of sphere and the plane coordinate system and the theory of spherical camera measuring the coordinate of the ground and the vertical height, spherical forward intersection, resection and relative orientation.With this basic theory, we can set up the full automatic target recognition theory which is the aimed theory. This project based on the arithmetic of TLD for intergration of Tracking,Learing and Detecting three modules so as to realize the optimization tracking algorithm .Tracking module will take the continuous movement inspection of the target as positive samples and take the false alarm which is far away from the trajectory as negative samples.Besides,it does real-time optimization to the classifier of Learning module so as to feedback and instruct Detection module to get more accurate test results..Panorama of measuring and target localization theory can be used in many domains such as three-dimensional visualization probe navigation and target tracking, rover, space exploration data collection, polar environment monitoring, comprehensive military targets reconnaissance and automatic warning, the vision mobile measuring system and so on .

近年来，由于机器人导航技术、虚拟现实技术的发展，对大视场甚至全视场场景的立体感知的需要日益增强。针对此需求，本项目旨在研究基于球面测量理论的全方位目标检测与跟踪的方法。.首先建立全方位目标定位所需的球面测量理论，包括：球面基础矩阵及核线约束理论、球面与外平面坐标系间的Homograph映射关系等理论。在此基础上研究全景视频条件下，跟踪Track、学习Learning、检测Detect三模块集成的自优化全景目标跟踪算法理论。跟踪模块将检测模块得到的连续运动目标作为正样本，远离运动轨迹的虚警作为负样本，对学习模块的分类器做实时优化，进而反馈指导检测模块得到更准确的检测结果。.本课题将研究针对球面全景视频模式下的TLD自优化目标定位和跟踪理论。研究成果将可应用在空间探测器三维可视化导航及目标跟踪、月球车、火星车周边环境探测、极地环境监控、全方位军事目标侦察等领域。

本项目主要研究基于全景成像设备的球面全景摄影测量及目标跟踪检测，以及配套的目标视觉跟踪-学习-检测算法。从2013年1至2016年12月，本团队历经4年不懈努力，在以下相关领域做出了很多进展：.(1).全景目标跟踪-学习-检测算法及软件；.(2).全景实时单目定位及测高原理；.(3).3维全景陀螺仪拼接算法理论； .(4).连续不变矩特征转换成离散特征学习算法；.(5).全景目标识别可信度算法研究；.(6).全景近邻可信度跟踪系统；.(7).街景全景实时运动目标三维重现；.(8).全景目标学习及特征提取算法；.(9).针对全景目标的卷积神经网目标学习平台。. 本项目已获得专利授权4项，申请专利9项，已发表论文3篇。本项目已完成了包括全景目标跟踪及测量系统、便携式全景摄影测量系统、单目目标实时跟踪及3维精密测量系统在内的多个有实际应用价值的实验系统。本技术还在不断发展，在智慧城市测量、测绘数据采集、移动全景测量等领域有着广泛应用。