Research On Three-dimensional Point Cloud Classification And Detection Based On Deep Learning

With the popularity of 3D data acquisition devices such as lidar sensors and depth sensors,point cloud has become a common type of computer vision data.Research on how to use point cloud data to serve various applications has become a new hotspot in the field of computer vision in deep learning.Different from the multi view method and voxel method,which will lead to conversion loss,most researchers now focus on how to directly use 3D point cloud as network input.However,due to the sparsity,non structure,disorder and rotation invariance of point cloud,convolutional neural network can not be directly applied to point cloud feature extraction,and some traditional feature enhancement methods such as attention mechanism are not directly applied to point cloud data.Based on the existing network model of directly using point cloud for feature learning,this paper analyzes the advantages and disadvantages of point cloud feature extraction scheme based on point-wise method.Aiming at the problems existing in the existing point cloud feature learning network model,such as construct neighborhood features,overcome the limitations of single angle pooling,and enhance the ability of network feature representation,this paper studies the feature extraction and enhancement of point cloud learning network based on manual and adaptive weighting.The work and achievements of this paper are as follows:(1)A weighted point cloud classification network wdgcnn based on dynamic graph convolution is proposed.The network structure is optimized by the idea of feature concat to realize the fusion of multi-level features;by designing an appropriate weighting function for the edge features composed of k-nearest neighbor graph,we can weaken the interference of far points and relatively strengthen the features of near points,and better learn the point cloud characteristics of the network;by virtue of the advantage of average pooling preserving global features and combining maximum pooling and average pooling,a new pooling method is proposed to alleviate the information loss caused by maximum pooling alone.Through comparative analysis with dgcnn and other methods on modelnet40 point cloud classification data set,the classification accuracy of wdgcnn is improved from 91.61% to 93.22% compared with dgcnn.Compared with other directions,the proposed method has achieved equivalent or better results,which verifies the rationality of the optimization method design in this paper.(2)An attention mechanism directly applied to point cloud data is proposed.This method can adaptively generate weights through the attention mechanism.Through the parallel maximum pooling and average pooling of point cloud data,the multi-layer perceptron with shared weight is used to train the adaptive attention weight,which is multiplied with the input features to enhance the network feature representation ability,so as to improve the network performance.It can be widely used in the feature extraction stage of pointnet networks to improve the network representation ability.The attention mechanism designed in this paper can help the OA(Overall Accuracy)of Point Net(vanilla)and Point Net to improve 0.89% and 0.40%respectively in Model Net40 classification task;in Shape Net partial segmentation task,the m Io U(mean Intersection over Union)of Point Net can increase 1.38%;in the KITTI 3D detection task,the AP(Average Precision)of Frustum-Point Net in pedestrian and cyclist detection has been significantly improved.Experiments show that the designed attention mechanism is effective and lightweight in multiple point cloud processing tasks.