Researches On Depth Map Based Point Cloud Fusion Methods And System

With the development of the stereo matching and the prevalence of the depth sensors,depth maps is more convenient,which makes it possible to restore the object point cloud model.Nowadays,point clouds are of great significance in the application,like virtual reality,3D printing.Point cloud fusion based on depth map is very useful,but there are number of redundant information and unrelated background points,which will compromise the quality of the point cloud.In this paper,above problems are deal with via a multiply depth fusion system.The fusion algorithm is proposed based on the hypothesis test to solve redundant information issue.Firstly,the initial correspondence between cross views is established by homograph matrix via cameras parameter.To optimize this unreliable relation,a Gaussian distribution mode,based on error of disparity from multiview,is estimated and the confidence of each point is computed jointly by the model and difference of normal vectors.In fusion stage,the point will be accepted if its confidence is within confidence interval.Then,to ensure real-time running of algorithm,a GPU framework is set up and has been tested both on public and our own multiview datasets.The experiments show that our fusion algorithm can eliminate and merge redundant points and ensure a point cloud model of high quality.The multi cameras joint calibration algorithm is proposed to handle non-coplanar cameras setting in our depth fusion system.Combined with object based calibration and selfcalibration,our calibration algorithm can output coarse camera parameters and refine them using geometry prior of camera array.By minimize the distance between estimated circle of cameras center and outliers,external camera parameter is updated iteratively.A depth fusion system is set up in the laboratory condition to achieve reconstruction of point cloud in an end to end workflow.Data acquisition,controller and process unit are included in this system as well as two mentioned algorithm above.Depth data is obtained by stereo matching method and our fusion method is applied to guarantee object model.A set of multi view stereopsis is established in our system for further use.