CN-115170682-B - Method for processing point cloud data and target processing point cloud data model

CN115170682BCN 115170682 BCN115170682 BCN 115170682BCN-115170682-B

Abstract

The embodiment of the application provides a method for processing point cloud data and a target processing point cloud data model, wherein the method comprises the steps of obtaining point cloud data to be compressed, wherein the point cloud data to be compressed are obtained by collecting point cloud data of a target scene or a target object; inputting the point cloud data to be compressed into a compression model, and obtaining compressed point cloud data through the compression model, wherein the compression model comprises a plurality of layers of feature extraction modules, each layer of feature extraction modules further comprises a local feature extraction module and a global feature extraction module connected with the local feature extraction module, the global feature extraction module at least comprises a self-attention module and a feedforward neural network, and image processing is carried out according to the compressed point cloud data. According to the method and the device for compressing the point cloud data, the compression of the point cloud data can be achieved, global features of the point cloud data can be obtained, and further accuracy of subsequent image processing can be improved.

Inventors

GAO WEI
XIE LIANG
LI GE

Assignees

北京大学深圳研究生院
北京大学深圳研究生院

Dates

Publication Date: 20260421
Application Date: 20220426
Priority Date: 20220426

Claims (7)

1. The method for processing the point cloud data is characterized by being applied to a target processing point cloud data model, wherein the target processing point cloud data model comprises a compression model, a super prior encoder, a decoding module and an identification module, the target processing point cloud data model is obtained through target loss function training, the super prior encoder is used for obtaining Gaussian distribution parameters according to compressed point cloud data, the decoding module is used for carrying out up-sampling decoding on the compressed point cloud data through the Gaussian distribution parameters to obtain a reconstructed image, the target loss function is characterized by a reconstruction result loss value and an identification result loss value, the reconstruction result loss value is determined by the compression model, the super prior encoder and the decoding module, and the identification result loss value is determined by the compression model, the super prior encoder and the identification module; The method comprises the following steps: Acquiring point cloud data to be compressed, wherein the point cloud data to be compressed are obtained by acquiring point cloud data of a target scene or a target object; Inputting the point cloud data to be compressed into a compression model, and obtaining compressed point cloud data through the compression model, wherein the compression model comprises a plurality of layers of feature extraction modules, each layer of feature extraction modules further comprises a local feature extraction module and a global feature extraction module connected with the local feature extraction module, and the global feature extraction module at least comprises a self-attention module and a feedforward neural network; The compressed point cloud data is input into a super prior encoder, the super prior encoder performs up-sampling and down-sampling operation on the compressed point cloud data to obtain super prior point cloud data characteristics, the super prior point cloud data is continuously compressed, the super prior down-sampling global characteristics are obtained through combination of a multi-layer perceptron and a position-free global characteristic extraction module, the super prior down-sampling global characteristics are subjected to up-sampling operation to obtain super prior point cloud data characteristics, the position-free global characteristic extraction module is built based on a Transformer frame, the input of the position-free global characteristic extraction module is the space coordinates of the point cloud data and the output characteristics of the multi-layer perceptron, and the position-free global characteristic extraction module does not need to input local characteristics; according to the super prior point cloud data characteristics, gaussian distribution parameters are obtained, wherein the Gaussian distribution parameters comprise a mean value and a variance; performing arithmetic coding and decoding operation on the compressed point cloud data based on the Gaussian distribution parameters to obtain target compressed point cloud data; and performing image processing according to the target compression point cloud data.
2. The method of claim 1, wherein the global feature extraction module further comprises a first normalization module and a second normalization module; Wherein the first normalization module is configured to: inputting the output of the self-attention module, the downsampled local features and the spatial coordinates of the point cloud data into the first normalization module; adding the output of the self-attention module, the downsampled local features and the space coordinates of the point cloud data through the first normalization module to obtain first added features; Normalizing the first summation feature to obtain a first normalization feature; The second normalization module is configured to: inputting the output of the feedforward neural network and the first normalization feature to the second normalization module; Adding the output of the feedforward neural network and the first normalization feature through the second normalization module to obtain a second added feature; and normalizing the second summation feature to obtain a downsampled global feature.
3. The method of claim 2, wherein the obtaining compressed point cloud data by the compression model comprises: inputting the output of the i-1 global feature extraction module into an i local feature extraction module to extract local features, so as to obtain an i downsampled local feature; inputting the output of the i-1 global feature extraction module and the i downsampled local feature into an i global feature extraction module to obtain an i downsampled global feature; The method comprises the steps that i is an integer greater than or equal to 2, the ith downsampling global feature is data obtained by compressing the point cloud data by a target multiplying power, and the target multiplying power is determined by all downsampling convolution kernels in an ith global feature extraction module; Repeating the steps until the compression code rate of the point cloud data meets the requirement, and obtaining the target downsampling global feature; And inputting the target downsampling global features into a multi-layer perceptron to obtain compressed point cloud data.
4. The method of claim 1, wherein the super a priori encoder comprises a multi-layer perceptron and a location-free global feature extraction module; the step of performing downsampling operation on the compressed point cloud data to obtain a super prior downsampled global feature includes: Inputting the output of the ith-1 non-position global feature extraction module into an ith multi-layer perceptron to obtain an ith perception feature; Inputting the ith perception feature and the space coordinate into an ith position-free global feature extraction module to extract global features, and obtaining an ith super priori downsampled global feature; Repeating the steps until the prior compression code rate of the point cloud data meets the requirement, and obtaining the super prior downsampling global feature.
5. The method of claim 1, wherein the performing image processing according to the target compressed point cloud data comprises: classifying according to the target compression point cloud data, and/or Reconstructing the target scene or the target object according to the target compression point cloud data.
6. The method of claim 5, wherein reconstructing the target scene or the target object from the target compressed point cloud data comprises: And inputting the target compression point cloud data into a decoding module, and obtaining the target scene or target object through the decoding module, wherein the decoding module comprises a multi-layer point cloud decoding module, and each layer of point cloud decoding module further comprises a decoding convolution layer and a position-free global feature extraction module connected with the decoding convolution layer.
7. The target processing point cloud data model is characterized by comprising a compression model, a super prior encoder, a decoding module and an identification module, wherein the target processing point cloud data model is obtained by training a target loss function, the super prior encoder is used for obtaining Gaussian distribution parameters according to compressed point cloud data, the decoding module is used for carrying out up-sampling decoding on the compressed point cloud data through the Gaussian distribution parameters to obtain a reconstructed image, the target loss function is characterized by a reconstruction result loss value and an identification result loss value, the reconstruction result loss value is determined by the compression model, the super prior encoder and the decoding module, and the identification result loss value is determined by the compression model, the super prior encoder and the identification module; The model comprises: The reconstruction module is used for compressing and reconstructing point cloud data to be compressed to obtain a reconstructed image, wherein the reconstruction module comprises: The compression model is used for carrying out downsampling encoding on the point cloud data to obtain compressed point cloud data, inputting the compressed point cloud data into a super priori encoder, and carrying out upsampling and downsampling operation on the compressed point cloud data through the super priori encoder to obtain super priori point cloud data characteristics, wherein the compressed point cloud data is subjected to downsampling operation, the super priori downsampling global characteristics are obtained through the combination of a multi-layer perceptron and a position-free global characteristic extraction module, and the super priori downsampling global characteristics are subjected to upsampling operation to obtain the super priori point cloud data characteristics; The super prior encoder is used for obtaining Gaussian distribution parameters according to the compressed point cloud data and obtaining Gaussian distribution parameters according to the super prior point cloud data characteristics, wherein the Gaussian distribution parameters comprise a mean value and a variance; the decoding module is used for carrying out up-sampling decoding on the compressed point cloud data through the Gaussian distribution parameters to obtain the reconstructed image, wherein the reconstructed image is a target scene image or a target object image, and carrying out arithmetic coding decoding operation on the compressed point cloud data based on the Gaussian distribution parameters to obtain target compressed point cloud data; and the identification module is used for obtaining the type of the point cloud data by classifying the compressed point cloud data and carrying out image processing according to the target compressed point cloud data.

Description

Method for processing point cloud data and target processing point cloud data model Technical Field The embodiment of the application relates to the field of point cloud processing, in particular to a method for processing point cloud data and a target processing point cloud data model. Background Three-dimensional point clouds are an important representation of real world digitization. The huge amount of data in the three-dimensional point cloud brings great challenges to the storage and transmission of the data, and therefore, the subsequent steps are usually required to be carried out after the point cloud data are compressed. In the related technology, aiming at static point cloud compression, compression is realized by utilizing local features of point cloud data, but the local features cannot accurately express real features of the point cloud data, so that the accuracy of the subsequent steps of identification, reconstruction and the like is reduced. Therefore, how to accurately express the characteristics of the point cloud data in the process of compressing the point cloud data by adopting the local characteristics becomes a problem to be solved. Disclosure of Invention The embodiment of the application provides a method for processing point cloud data and a target processing point cloud data model, which can at least realize compression of the point cloud data and obtain global characteristics of the point cloud data through some embodiments of the application, thereby improving the accuracy of subsequent image processing. The embodiment of the application provides a method for processing point cloud data, which comprises the steps of obtaining point cloud data to be compressed, inputting the point cloud data to be compressed into a compression model, obtaining compressed point cloud data through the compression model, wherein each layer of feature extraction module further comprises a local feature extraction module and a global feature extraction module connected with the local feature extraction module, the global feature extraction module at least comprises a self-attention module and a feedforward neural network, and performing image processing according to the compressed point cloud data. Therefore, unlike the method of compressing the point cloud data only through the local features in the related art, the embodiment of the application can extract the global features of depth in the point cloud data through the global feature extraction module connected with the local feature extraction module, so that the accuracy of subsequent image processing is improved. With reference to the first aspect, in one implementation mode of the application, the global feature extraction module further comprises a first normalization module and a second normalization module, wherein the first normalization module is configured to input the output of the self-attention module, the downsampled local feature and the spatial coordinates of the point cloud data into the first normalization module, perform addition operation on the output of the self-attention module, the downsampled local feature and the spatial coordinates of the point cloud data through the first normalization module to obtain a first addition feature, perform normalization operation on the first addition feature to obtain a first normalization feature, and the second normalization module is configured to input the output of the feedforward neural network and the first normalization feature into the second normalization module, perform addition operation on the output of the feedforward neural network and the first normalization feature through the second normalization module to obtain a second addition feature, and perform normalization operation on the second addition feature to obtain the downsampled global feature. Therefore, the embodiment of the application fuses the down-sampling local feature and the feature corresponding to the self-attention through the first normalization module and the second normalization module in the global feature extraction module, so that the global feature of the point cloud data can be obtained, and the accuracy of the subsequent downstream task can be improved. In combination with the first aspect, in one embodiment of the present application, the obtaining of compressed point cloud data through the compression model includes inputting an output of an i-1 global feature extraction module into an i local feature extraction module to perform local feature extraction to obtain an i downsampled local feature, inputting an output of the i-1 global feature extraction module and the i downsampled local feature into the i global feature extraction module to obtain an i downsampled global feature, where i is an integer greater than or equal to 1, the i downsampled global feature is data obtained by compressing the point cloud data by a target multiplying power, the target multiplying power is determined by all downsa