Sfm Algorithm For Multi-lens Combined Panoramic Camera

Panoramic cameras are one of the main visual sensors in the field of computer vision and photogrammetry due to the characteristics of 360 ° full-view imaging,such as mobile measurement vehicles,autonomous vehicles and Mars detection robots.With the improvement of 3D image reconstruction technology,especially the Structure From Motion(SFM)algorithm can accurately,stably and quickly recover the 3D structure of the target scene From the image.The 3D reconstruction of panoramic images has high efficiency and good continuity,and is suitable for the reconstruction of large-scale scenes.At present,the poses of panoramic images are mainly obtained through external platforms like GPS / IMU,or a motion recovery structure algorithm based on a spherical panoramic camera model.Although the latter is not affected by wrong or weak GPS signals,and the reconstruction effect is better.However,multi-lens combined panoramic cameras with special imaging methods face difficulties such as stitching errors,matching errors,and frontal intersection error control in orientation,making it difficult to meet the requirements for accurate 3D scene reconstruction.This paper focuses on the three-dimensional reconstruction strategy of multi-lens combined panoramic cameras,and proposes an SFM algorithm to achieve more efficient and accurate pose estimation and model reconstruction.In order to avoid problems such as errors in panoramic image stitching,this article uses a strict non-spherical imaging model,which can construct the extended collinear equation From the pose relationship between each lens and express continuous and unified spatial information of the scene more accurately.Based on this imaging model,the improved SFM algorithm is used to estimate the pose of images,and this work is divided into two cases: single-lens and multi-lens.Aiming at how to complete the pose estimation of the single-lens image group more efficiently,this paper proposes an improved incremental SFM algorithm.The real data experiments show that the improved algorithm is more efficient and stable.At the same time,this paper proposes an Eye-to-Eye camera calibration method to restore a unified scene coordinate system through the rigid relative pose constraint between multi-lens combination cameras,and then realizes more poses through the resection of the multi-lens SFM algorithm.Experimental results show that the multi-lens camera pose estimation method can successfully achieve all image orientations and restore a good three-dimensional structure.The bundle adjustment is the last but not least step of the SFM algorithm.Based on the extend imaging model of the multi-lens combined panoramic camera,this paper re-derives the minimized re-projection error equation for cost function and reduces the model's unknown parameters.As the experimental results show,the optimization process in this paper can quickly converge and the results are correct,compared with the traditional beam adjustment.In the analysis of the material accuracy of the image reconstruction optimization results,control point constraints are introduced in this paper.The point deviations of the check points indicate that the optimized results of the pose estimation can reach centimeter level.