Research On The Scene 3D Point Clouds Classification Based On Deep Learning

In recent years,with the emergence of low-cost miniaturized 3D sensors,the acquisition of 3D shape models is becoming easier and easier,and the recognition and analysis of 3D shape models has become an important research direction.With the rapid development of deep learning,the research on 3d shape has shifted from manual feature extraction to deep neural network for classification and detection.3D shapes can be represented by multi-view images,voxel,point clouds,polymesh and so on.In this thesis,the classification task of 3D point cloud data is discussed.Existing deep learning methods include Point Net,Point Net++,Point CNN and kd-network.Each method has advantages and disadvantages.Based on these models,three deep learning architectures based on Tensor Flow was proposed.(1)An efficient point clouds classification model based on KD tree was proposed to solve the time-consuming problem of Point Net++ hierarchical model.KD tree structure was introduced in the grouping layer,constructed on three-dimensional point clouds to conducts range query to create local area set,which was then input into mini-PointNet for feature extraction.At the same time,to avoid the overfitting problem in the training process of the original network,the dropout algorithm was introduced to inhibit the overfitting of network to training samples and reduce the convergence time.The classification experiment of the improved model was carried out on ModelNet40.The experimental results showed that the classification accuracy reached 92.5% when the query radius was0.3m,which was better than Point Net++ hierarchical model.The semantic parsing task on Stanford 3D semantic parsing dataset also presented better robustness,where the MIoU reached 57.2%.(2)A deep learning point clouds classification model based on improved softmax classifier was proposed.To solve the problem of parameter redundancy in softmax model,an improved softmax classifier was obtained by adding weight attenuation into softmax cost function.Then connected this improved softmax to the fully connected layer of PointNet to normalize the k class scores to the real probability of(0,1)for easily understanding.The experiment evaluated the model on the 2D dataset MNIST and the 3D dataset ModelNet40.For 2D shape classification,the model's classification accuracy reached 99.7% and for 3D shape classification,the model's output accuracy was 91.1%,both exceeding Point Net classification model,which proved the effectiveness of the proposed model in three-dimensional point clouds classification.(3)A new point clouds classification model based on GCN is proposed.To overcome the shortcomings of the Point Net model,a kNN graph layer is designed to construct a k-nearest neighbor graph in the point clouds space.Then encode the local information using the graph structure,so as to make more effective use of the local information of the point clouds.The classification experiment was carried out on ModelNet40,and the effects of different neighbor values k on the model classification accuracy were compared.The results show that when k is 20,the classification accuracy is the highest,reaching 93.2%.