Research On Point Cloud Classification Technology Based On Deep Learning

As the method of obtaining point cloud data becomes more and more simple,the processing of point cloud data is gradually becoming a hot spot.Although many excellent jobs have appeared at this stage,they still face great challenges in terms of overall accuracy and practicality.On the one hand,how to use deep learning technology to further improve the accuracy of point cloud classification;on the other hand,how to reduce the number of network parameters so that the network can meet the needs of practical applications while ensuring the accuracy.This article has conducted in-depth research on the above problems in point cloud data processing and proposed new solutions.The main tasks are as follows:(1)We propose a multi-scale point cloud classification network Multi-scale Point Cloud Classification Network(MSP-Net)to further improve the accuracy of point cloud classification.First,the characteristics of the point cloud are analyzed,and a new method for local partitioning of point cloud data is proposed.Different levels of the point cloud are used as input to obtain local regions of different scales.Then,the operating principle of the convolutional neural network is simulated,and a multi-scale point cloud classification network consisting of 4 modules is constructed.Finally,tests are performed on the standard public datasets ModelNet10 and ModelNet40,which verifies the feasibility and effectiveness of the algorithm in this paper.(2)We propose an efficient point cloud classification network Efficient Significant Point Clouds Processing Convolutional Network(ESP-Net),which satisfies the greatly reduced parameter quantity while ensuring accuracy.The network consists of multiple multi-scale saliency feature extraction blocks(MS-SFE).Each MS-SFE block contains a valid point sampling module(SPS)and an MS-SFE module.These modules can be flexibly divided and added to other networks.Under the premise of ensuring that the parameter amount is only 0.3×10~6,the accuracy rate on the standard public data set ModelNet40 reaches 92.42%;the ESP-Net network is directly applied to the segmentation task and has achieved comparable results to the existing methods.Experiments have shown the tremendous potential of our network in real-time applications such as autonomous driving.Based on the above research,this paper designs and implements a point cloud classification system based on deep learning under Windows.Based on the research content of this paper,the system uses the standard 3D point cloud data set as training and test samples to realize the classification of 3D point clouds.Finally,the effectiveness of the algorithm is evaluated through comparison and ablation experiments;the system designed in this paper is comprehensively evaluated in a qualitative and quantitative manner.The experimental results show that the proposed algorithm achieves good results on the standard data set of point cloud classification tasks.The system implemented in this paper perfectly supports the needs of point cloud classification tasks,and is flexible,convenient,and easy to use.