语义信息与地理配准相结合的实例目标定位  

Instance object localization based on semantic information and geo-registration

作　　者：吕可枫 张永生[2] 于英[2] 闵杰 L Kefeng;ZHANG Yongsheng;YU Ying;MIN Jie(Troops 61290,Shanghai 200422,China;Institute of Geospatial Information,Information Engineering University,Zhengzhou 450001,China;Troops 96863,Luoyang 471000,China)

机构地区：[1]61290部队,上海200442 [2]信息工程大学地理空间信息学院,河南郑州450001 [3]96863部队,河南洛阳471000

出　　处：《测绘学报》2023年第8期1375-1386,共12页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(42071340);中原学者首席科学家工作室专项(2018007)。

摘　　要：随着地理空间科学的迅速发展,面向地理空间智能感知应用的研究日益增多。本文以三维模型为先验数据,提出了一种基于语义信息和地理配准的实例目标智能感知与定位的方法。首先,使用载体上搭载的GNSS和IMU(inertial measurement unit)获取传感器的位姿,并利用初始位姿从先验的三维模型实时渲染图像,同时用深度相机获取当前帧的真实场景图像;其次,对真实图像和渲染图像进行全景分割,使用语义分割结果对两幅图进行匹配,获取匹配点对;然后,使用匹配点信息得到两个图像的运动关系,并根据运动参数和三维模型的地理坐标信息对传感器位姿进行优化;最后,使用更新后的传感器位姿、实例分割结果及对应深度信息对目标进行感知和定位。针对不同类型三维模型,使用不同类型图像数据进行了测试,并与多种匹配算法进行了对比,结果表明本文算法能够提高匹配准确率和定位精度,并能有效对目标进行感知和定位。With the rapid development of geospatial science,there are more and more researches towards geospatial intelligent perception.Based on geospatial 3D model as priori data,an instance object perception and localization method based on semantic information and geo-registration is proposed.First,GNSS and IMU(inertial measurement unit)are used to obtain its initial position and orientation,which are then used to render the images from 3D models.Second,the panoptic segmentation and the match network are together used to segment and match the 3D model rendered image and the truth image.Third,the matching results between two images are used to restore the camera motion,which then is used to refine the position of the camera with the geographic coordinate information of the 3D models.Finally,the objects can be detected and localized using the results of instance segmentation,depth information and the refined camera position.The method was tested on different types of 3D models and images.The results demonstrate that the proposed method can improve the matching and localization accuracy,and can detect and localize objects effectively.

关 键 词：地理空间智能 地理配准 全景分割 目标定位 

分 类 号：P208[天文地球—地图制图学与地理信息工程]