Research Of Plant Point Cloud Registration Algorithm Based On Hypothesis Test Matching Constraints

The three-dimensional point cloud reconstruction technology is a research hotspot in graphics and virtual reality,and has received extensive attention and application in the field of agricultural research.In this technology,the surface information of the plant is mainly obtained by scanning device,and its three-dimensional model is reconstructed,in order to research the growth law of the plant and formulate the correct cultivation management plan.Due to the limited view range of the scanning device,it is impossible to obtain all the surface information of the plant in one scan.In order to obtain the complete model of the plant,point cloud registration technology needs to be used to align the point clouds of different perspectives into the same coordinate system,which plays a crucial role in the quality of the final plant model.However,due to the complex shape of the plant point cloud and the mutual occlusion between branches and leaves,many existing registration algorithms cannot accurately and efficiently register the plant point cloud.Aiming at the problems of low efficiency,large error,and weak noise resistance in plant point cloud registration,a new registration algorithm based on hypothesis test matching constraints is proposed,and it includes initial registration and fine registration.The main contents and innovations of this research are as follows:?1?Aiming at the influence of outliers on initial registration,an improved algorithm for initial registration of point clouds based on PCA was proposed.The normal distribution was used to statistically analyze the neighborhood distance average of vertices,and non-outliers were extracted for calculate the principal directions of the two point clouds,and then the principal directions were aligned as two new reference frames to complete the initial registration.The experimental results show that compared with the traditional PCA-based point cloud initial registration algorithm and other initial registration algorithms,the improved algorithm for initial registration of point clouds based on PCA can effectively reduce the influence of outliers and improve the overall registration efficiency by about 30%?93%.The extraction rate of non-outlier points of the improved algorithm on the standard point cloud reaches 98.6%?99.6%,which will not destroy the original structure of the point cloud.?2?Aiming at the point pair selection process in fine registration,a new constraint of matching point pairs based on hypothesis test and a candidate point pair search strategy based on uniform distribution were proposed.The former used t test to analyze the neighborhood distance distribution of candidate point pairs and eliminated wrong point pairs.It has reduced the number of point pairs and improved the efficiency and accuracy of point pairs selection.The latter abandoned the previous method of searching for candidate point pairs according to the index order of the points in the previous research,and searched for candidate point pairs based on uniform distribution,so that the selected point pairs searched in each iteration were distributed as much as possible in each part of the point cloud model to ensure the complete morphological registration of point clouds.Compared with the ICP?Iterative Closest Point?algorithm,the improved method for point pair selection improves the registration efficiency and accuracy by about 50%and 40%.?3?Aiming at the unexpected situation of algorithm convergence,a multiple convergence detection method was proposed.Because the candidate point pair search strategy based on uniform distribution caused a certain randomness in the selection of point pairs,during the iteration process,a batch of point pairs searched for a certain iteration had high matching accuracy,which met the algorithm's convergence criteria,but the overall point cloud model was not completely aligned,and the registration algorithm ended early.To solve this problem,the multiple convergence detection method counted the number of times that the algorithm met the convergence criterion during the iteration process.The iteration process was ended only when the number of convergence times equalled the set threshold,and the adverse influence of the randomness of the point pair on the algorithm was reduced.Compared with the traditional method of using a single iteration to judge whether the algorithm converges,although the multiple convergence detection method reduces the registration efficiency by about 8%,it further improves the registration accuracy by about 10%.?4?The improvement schemes of point pair selection and algorithm convergence were combined with the ICP algorithm,and the new algorithm was named T-ICP?Test-Iterative Closest Point?fine registration algorithm.In order to verify the improved registration algorithm,Stanford Bunny,Dragon,Happy Buddha and other standard point clouds in the 3D model scanning library of Stanford University were used to verify the effectiveness of the algorithm and the rationality of the registration parameters set,and the corresponding registration error reaches the level of 10^-4mm and the registration time is about 1 minute.Furthermore,based on the point cloud models of apple trees and magnolia trees scanned by Kinect 2.0 equipment,the proposed registration algorithm was compared with other algorithms to verify the superiority of the proposed algorithm in plant point cloud registration.Experimental data shows that in the application of plant point cloud registration,the proposed algorithm improves the efficiency and accuracy by about 2%?60%and 9%?50%,respectively,and has better noise immunity and robustness,compared with ICP and some improved registration algorithms in recent years.Finally,chair and potted plant point cloud models were used to verify the generality of the proposed algorithm.The corresponding registration time is within 10s,and the error reaches the level of 10^-1mm.