Research On Multi-station Point Cloud Data Registration Algorithm Based On Graph Optimization

In the process of 3D laser scanning,due to the limitations of scanning range,scanning angle and complexity of measured objects,multi-station scanning is needed to do while the target object is measured.Only in this way can mutually independent point cloud data of coordinate system on each station be obtained.In order to the integrity of point cloud data,multi-station point cloud data must be registered.According to the difference stations of point cloud,point cloud registration can be divided into two-station registration and multi-station registration.The algorithm of two-station registration is mature.The process of registration includes coarse registration and fine registration.For multi-station registration,sequential registration is adopted at first because this method is feasible.However,existing errors of two-station point cloud registration will cause error accumulation in the process of registering the multi-station.In recent years,some scholars have proposed multi-station registration based on closed condition constraint.The global optimal coordinate conversion parameters are estimated by the adjustment method to improve accuracy and reliability of overall registration.However,this method is focus on the multi-station registration of single closed loop.Regional network data including the multiple closed loop only remains in the superficial level.Focus on the existing problems of multi-site cloud registration,the main work of this paper is as follows:Firstly,according to the steps of the two site cloud data registration,coarse registration was been done at first.Anticipated Registration point clouds from two different perspectives were respectively regarded as the source point set and object point set.Three methods-sampling consistency method based on FPFH,Super4pcs,index point-were adopted to register coarsely point cloud data.The experimental results showed that these three methods could initially register the point cloud data.Secondly,rigid body transformtion matrix was gotten by making use of previous coarse registration,which can be treated as initial iteration value.It can continue to be accurately registered by using of ICP algorithm and NDT algorithm.In the process,the method of voxel grid was adopted to simplify the point cloud data.The ICP algorithm and NDT algorithm were compared and analyzed from the aspect of registration precision and registration efficiency.Thirdly,based on the graph-based optimization,the paper proposed the closed multi-site cloud registration algorithm.Focus on the closed multi-site cloud data,the position of each site scanner was regarded as apex and transformtion matrix gotten by registering the two sites was regarded as sides.Therefore,pose graph was established and optimized.With the data measured by Ground 3D laser scanner,experiment was done.It turned out that this method was effective and feasible.