Research On 3D Point Cloud Registration Algorithm Of Complex Surface Based On Local Features

3D vision measurement is a new and advanced technology developed from computer vision and precision measurement.It has become the optimization way to get3 D information of objects with the characteristics of high-precision,non-contact and high-efficiency.However,due to the limitation of the field of view of measurement system and the self-occlusion of the object,only part of 3D point cloud data can be acquired from a single view.To obtain complete 3D information of the object surface,multi-view 3D point clouds need to be registered into a same coordinate system.At present,the registration methods based on the cooperative target and auxiliary equipment have achieved remarkable results.However,there are still some deficiencies and application limitations for the feature-based automatic registration methods in terms of registration accuracy and stability.Therefore,this paper regards the description of local feature and the optimization of iterative closest point algorithm as the research direction on the base of registration idea from coarse to fine.The 3D point cloud registration algorithm for the complex surfaces with low overlap rate is studied.The main research contents of this paper are as follows:(1)For the problem that local features are easily influenced by noise,occlusion and resolution variation in the point cloud,a local feature descriptor based on the deviation angle of the grid normal is proposed.First,the local reference system,established by covariance analysis and feature transformation,provides the spatial information and the invariance to rigid transformation for descriptor.Then,the grid point generated by grid division is regarded as the coding unit to calculate the normal deviation angle.The impact of point disturbance for the performance of descriptor is effectively suppressed and the calculation efficiency is improved by this local feature description method.Meanwhile,the problem of dimension verbosity for the descriptor based on the local reference system is solved by feature statistics.(2)Targeting the 3D point cloud for complex surfaces with low overlap rate,the automatic registration algorithm is studied deeply.First,a local feature-based coarse registration algorithm is proposed to provide a good initial value for the subsequent fine registration,and a local surface completeness judgment method is designed to reduce the proportion of mismatches.Then,in order to compensate the shortcomings of the iterative closest point algorithm,the improved measurers of simplifying the point cloud and dynamically adjusting the distance threshold are proposed to improve the performance of the fine registration algorithm.Based on this,multi-view point clouds registrations are achieved by continuous registrations.(3)The 3D point cloud registration experiments based on public data sets and measured point cloud data are designed.The comparative experimental results suggest that the new descriptor presented in this paper has higher comprehensive matching performance on different types of point cloud data sets.And the point cloud registration results have higher accuracy and efficiency.The average registration accuracy is superior to 0.5mm and the time is at the second level.