Research On Key Technologies In 3D Point Cloud Registration

In the field of photogrammetry and computer vision,point cloud registration technology has an important impact on other subsequent processing as an important step of point cloud data processing.According to its process,point cloud registration technology can be divided into four steps: preprocessing,feature extraction,outlier removal and attitude estimation.With the continuous improvement of hardware level,the increasing perfection of deep learning theory and the increasing richness of data acquisition methods,the advantages of deep learning in feature extraction are gradually highlighted,and it shows great application potential in 3D point cloud registration field.Based on the deep learning method,this paper conducts an in-depth research on point cloud registration technology according to the four steps.The main research contents and innovations are as follows:(1)Aiming at the problem of inaccurate noise distribution estimation caused by the lack of iteration in current point cloud denoising methods based on deep learning,a twostage point cloud denoising network based on noise distribution estimation is proposed in this paper.The network realizes iterative refinement of feature distribution and noise estimation by adding noise distribution features and network iterations.The experimental results show that:(1)When the noise is less than 0.25%,the best denoising effect can be obtained by estimating the mean and variance statistics of the noise distribution simultaneously.When the noise is between 0.25% and 1.5%,only estimating the mean statistic has the best effect,and it is suitable for many types of noise conditions.(2)For the point cloud with a density of 10 K to 50 K,the number of optimal neighboring points is positively correlated with the density of point clouds.(3)The refined prediction network module has a greater impact on the denoising accuracy than the initial prediction network module,and the number of neurons has a greater impact than the number of modules.(2)Aiming at the problem that traditional artificially designed feature descriptor FPFH cannot learn based on data and has poor generalization ability,and the feature description of GNN based on deep learning has weak robustness,this paper proposes a weight selflearning FPFH＿pro method,so that the weight of neighborhood point features can be selflearned based on data,which will increase the generalization of the method.On this basis,features extracted from the FPFH＿pro and GNN and their correlation features are deeply fused through the twin network to form more robust features.The experimental results show that:(1)The registration accuracy of FPFH＿pro method is higher than that of the classical FPFH method.(2)If the two methods with similar accuracy are fused,the performance will be improved.The fused features have stronger generalization when the noise intensity is in the range of 0% to 3%.(3)The single iteration speed of the fusion method is similar to that of the FPFH method,but the convergence speed is faster,which improves the training efficiency of the model.(3)Aiming at the problem that traditional point cloud registration methods are difficult to take into account registration accuracy and computational efficiency,this paper proposes a registration method combining point cloud filtering and adaptive fireworks algorithm.This algorithm adds the point cloud filtering process and adaptive fireworks rough registration process on the basis of KD-tree ICP algorithm,which can simultaneously improve the registration accuracy and registration efficiency.The experimental results show that:(1)The ICP algorithm based on KD-tree can simultaneously improve the registration accuracy and speed of the registration algorithm.(2)Before using the KD-tree ICP algorithm for accurate registration,adding the coarse registration process based on the adaptive fireworks algorithm can improve the registration accuracy by one order of magnitude.(3)The adaptive registration accuracy is related to the number of fireworks,and reaches the optimal value when the number of fireworks is 15.(4)Aiming at the problem that point clouds with tiny noises have different registration points,and the point pair constraint of loss function in the unsupervised deep learning registration method which leads to insufficient registration accuracy.In this paper,an unsupervised deep learning method based on latent surface is proposed.The method can predict the latent surface of the point cloud through the deep learning network,and then register two point clouds through the latent surface.The experimental results show that:(1)The latent surface prediction module and the consistent constraint of the latent surface can improve the registration accuracy of the model.(2)When the noise is within a certain range,the registration accuracy can be effectively improved by using the predicted latent surface.When the noise exceeds the threshold,it may lead to the deviation of the prediction surface,which will affect the final registration result.The denoising network in this paper can be used to reduce the noise,so as to improve the registration accuracy.(3)The unsupervised deep learning point cloud registration method base on the latent surface is superior to the traditional point cloud registration optimization method proposed in this paper in terms of accuracy and performance.The former takes a long time to train the model,while the latter is suitable for scenes with few data and low precision.