3D Point Cloud Data Acquisition Based On Sequence Images

3D modeling is pervasively used in various practical applications,such as video games,film visual effect,urban and landscape design,visual environment,and so on.Sequential images based 3D modeling is an active research filed in the computer vision.To achieve the3 D modeling of the object,the 3D point cloud data over the surface of the object are needed.However,there exist three problems in current sequential images based 3D point cloud data acquisition: 1)the accuracy of the camera self-calibration is low;2)the matching between a large number of images are computational expensive;3)the accuracy on the reconstruction of the 3D cloud point is low.To solve such problems,this paper makes the following contributions:(1)To handle the low accuracy of the camera self-calibration,we propose a local-global hybrid iterative optimization method for the camera self-calibration by using bundle adjustment algorithm and the spatial connection between the matching points and 3D points in the sequential images.Firstly,the bidirectional SIFT(Scale-invariant feature transform)feature matching is employed to the sequential images,then a pair of the most suitable image pair is found based on the image matching.Subsequently,the initial parameters of the camera can be computed based on the antipode geometrical relationship between this pair of the images,and the initial 3D point set can be achieved by the triangle survey principle.Further,the Bundle Adjustment algorithm is applied to optimize the initial parameters.Afterwards,by exploiting the spatial relationship between the matched images and the 3D point sets,the matched image is added iteratively.After calibrating the cameras of the all images,the final camera parameters are acquired by the bundle adjustment algorithm.In experiments,several evaluations are conducted on Valbonne Church sequential images by various self-calibration algorithms.In the experiments,our method yields high accuracy and provide sparse 3D point cloud.(2)We propose a neighboring image matching strategy to address the problem in the multiple image matching.During the procedure of picturing the target object,the sequential order of the all recorded images is saved.The overlapped regions of the target image are few when the angle between a pair of the images is large.Using the images,whose overlaps with the target image are large,within the neighborhood of the target image to do the image matching,which can improve the matching speed while preserving the matching accuracy.Experiments show the matching time only takes up 12.40% of the original method when applying the eightneighborhood to the 70 images.(3)After getting the sparse 3D point cloud,we need to extend the sparse data for acquiring the dense ones.This research develops an improved 3D point cloud acquisition method based on PMVS(Patched-based Multi-View Stereo)algorithm.To improve the confidence of the sparse3 D point cloud,the clustering is employed to the sequential images,meanwhile the filtering and classification strategies are improved.During the selection procedure of the patches,the color and texture normalization regularization is used to reduce the wrong seeds of the patches.Moreover,the normal vector of the patch is modified in the phase of the patch filtering,which reduces the false filtering of the patches.Through comparative experiments,our improved 3D point cloud acquisition method can improve the number of the 3D points by 21.67% of the dense3 D point cloud.