3D Point Cloud Processing Based On Deep Learning

Point cloud processing is an important research direction in the field of computer vision.It is widely used in the fields of intelligent driving,robot technology and smart city.The point cloud processing method represented by Point Net uses MLP with shared weight to extract the point cloud features,and aggregates the features through the maximum pool symmetry function to solve the disorder problem of point cloud,and has achieved competitive results in the current benchmark of point cloud shape classification and point cloud shape part segmentation.However,in the process of point cloud feature extraction,the number of channels is increasing.There are always some channels that contain useless information that is not conducive to feature extraction,and there are always some key points in the point cloud that are very important for shape recognition.They should be given more attention.At the same time,the aggregation feature of maximum pool symmetric function only retains the maximum value of point cloud feature,resulting in a large amount of feature information loss.To solve these problems,this thesis does the following work:(1)Channel attention mechanism is introduced to help the network automatically select the channel containing useful information.Point Net uses MLP with shared weight to map point cloud features from low-dimensional space to high-dimensional space,which will increase the number of point cloud feature channels,and some channels will contain a large number of redundant features that are unfavorable to feature extraction.To solve this problem,this thesis introduces the channel attention mechanism into the network,by giving greater weight to those useful channels,so as to select the useful channels and suppress those channels with a large number of redundant features.(2)A key point attention mechanism is proposed to help the network automatically select the key points in the point cloud data.From the perspective of human cognition of the shape of point cloud,we finds that only some key points are needed,people can correctly recognize the shape of point cloud,that is,for point cloud data,there are some key points,which are very important for shape recognition.Therefore,this thesis proposes a key point attention mechanism,which can give higher weight to the key points in the point cloud.(3)A higher-order statistics aggregation(HOSA)module is proposed to help the network aggregate more representative features.Most point-based point cloud processing methods simply use maximum pooling or average pooling to aggregate global features.From a statistical point of view,the maximum or mean value only describes one aspect of the characteristics of random variables,and many details of the probability distribution of random variables are ignored in this process.Therefore,this thesis proposes a high-order statistics aggregation module to replace the maximum pooling or average pooling symmetric function to calculate and aggregate multiple high-order and low-order statistics.(4)The HOSA network proposed in this thesis is analyzed theoretically,the general approximation ability of the network to the continuous set function is proved theoretically,and the stability of the network is analyzed theoretically.On the data set of point cloud shape classification and shape part segmentation,according to the two standards of classification accuracy and part segmentation m IOU,the comparative test is carried out with other point cloud processing algorithms.At the same time,the ablation experiment is carried out,and the robustness of point cloud data loss and noise is studied.Experimental results show that the proposed method can effectively improve the performance of shape classification and shape part segmentation.