Position And Orientation Estimation Of Spherical Panorama Image

After being launched as a kind of immersive visual scene rendering product,the spherical panorama image has triggered a new wave of technological revolution about map.The network of global street-view map has been established one after the other,and the spherical panorama image has been widely used in intelligent transportation of digital city,virtual tourism,cultural relics filing and so on.The Spherical panoramic image has richer texture detail and omnidirectional perspective,thus in addition to the further excavation of the immersive experience potential which is presented in,it also has a greater development potential on three-dimensional scene model reconstruction in close-range photogrammetry,texture mapping,measurement,as well as GIS data integration.At present,the collection and construction of the global street-view map network is generally carried out by integrating the GPS/IMU system and panoramic camera on the car platform.The accuracy of position and orientation estimation with street-view image is quite a little rough that it just meets the needs of immersive rendering,and the texture mapping with spherical panorama image need to be accompanied by manual editing,so it can't meet the requirements of automatic precise three-dimensional scene model reconstruction.It can be seen that the theory and technique about estimating the position and orientation of spherical panoramic image and reconstructing three-dimensional scene need to be explored to better exploit its role and value.Due to its special imaging methods,the spherical panoramic image is faced with avoiding and controlling the stitching error,matching error,front intersection error and other problems.In this paper,the projection model of spherical panoramic image imaging is constructed,and with that the pose of spherical panoramic image is described.As for the problem of estimating the position and orientation of spherical panoramic image in single station,an improved EPnP algorithm and a resection algorithm are proposed based on the collinearity condition.Through the simulation experiment and the real data experiment,the two algorithms are compared and analyzed to find that the improved EPnP algorithm is more efficient and stable,and the projection noise of object point caused by stitching error has greater impact on both two algorithms,so it should choose the spherical panoramic image with less stitching error and strictly control the distance distribution between the object point and the center of spherical panoramic image,to avoid the impact as possible as it can.At the same time,the polar geometric relation of spherical panoramic image pair is deduced by coplanar condition in this paper,the essential matrix is used to describe the polar geometric constraint and to optimize the matching between spherical panoramic images.The experimental results show that the optimization can eliminate the gross matching error and robustly estimate the essential matrix.In the automatic position and orientation estimation of spherical panoramic images in multi-station,this paper avoids the blind spot(sky and car-body)and obtains the initial matching point pairs through the SIFT feature extraction and matching algorithm,and the polar geometric constraint is applied to optimize the matching result.The incremental SfM algorithm is used to automatically estimate the positon and orientation,in the progress the newly added image is positioned and oriented with the improved EPnP algorithm,afterwards the results areoptimized as a whole by the beam method adjustment in single station and multi-station.The experimental results show that after estimating the position and orientation of spherical panoramic image with the radius of 1m,the reprojection standard deviation of the three-dimensional scene can reach mm level,and the measurement accuracy can reach dm level or higher.