Research On Method Of 3D Point Cloud Model Recognition

With the rapid development of computer vision,3D pattern recognition has become a hot research direction.People have put forward higher requirements for machine vision system,and more emphasis on the extraction of image information and the accuracy of object recognition.At this stage,a large number of research results derived on the basis of in-depth learning often focus on improving the accuracy of the model,while ignoring the practicality of the model.With the development of modern industry,building a real-time and efficient point cloud recognition model has become an urgent problem to be solved.With the increasing popularity of sensor devices,the application of point clouds is also becoming widespread.This article is based on deep learning and deeply explores the problems in 3D model recognition,and proposes solutions.When using deep learning technology to process 3D point cloud models,the main challenges are as follows: the distribution of point clouds is irregular,limited by the influence of collection devices,and the obtained point cloud data is usually uneven in density,which greatly hinders the progress of research.This experiment uses the standard dataset Model Net-40 publicly available from Princeton University.The internal data is manually drawn using CAD to avoid the phenomenon of "near dense and far sparse" point cloud data caused by equipment such as Li DAR during the collection process,which affects the experimental results.On the basis of the original data,this article uses PCA technology for dimensionality reduction processing,adding normal vector information to it to improve the model recognition accuracy;The point cloud data is unordered,that is,the point cloud data is independent of its internal point cloud arrangement order.This thesis refers to a series of function construction methods,and finally decides to use symmetric function to solve the problem that the point cloud distribution is independent of its arrangement order;The point cloud data is unstructured,and its internal points are independent of each other rather than fixed.It is very difficult to process the point cloud data with neural networks,and the spatial structure characteristics of the point cloud cannot be extracted.This thesis uses a multi-scale strategy to extract the local spatial characteristics of the point cloud.Its principle is similar to the Receptive field in convolutional networks.Influenced by residual networks,this article constructs a neural network framework based on residual modules.By introducing residual modules,a deep neural network is constructed to enhance the model’s ability to extract local information from point cloud data and improve the recognition accuracy of the model;Utilize the "collision detection" strategy to reduce the number of point clouds in the point cloud model and accelerate inference speed.This method did not affect the recognition accuracy of the model,as it effectively preserved the "key point" data during the recognition process.Experimental results show that the recognition accuracy of the algorithm proposed in this thesis reaches 91.5%,which is 2.3% higher than Point Net,and nearly 1.3% higher than Point Net++;The algorithm also accelerates the reasoning speed of the model,and its running efficiency is nearly 1/3 higher than that of Point Net++.This algorithm can be applied to point cloud model recognition,especially in scenes with real-time requirements,such as object detection on pipeline,3D object perception,etc.