CN115499777A

CN115499777A - Indoor passive positioning method for multi-posture users based on CSI

Info

Publication number: CN115499777A
Application number: CN202211120917.6A
Authority: CN
Inventors: 杜金锴; 章丁祥
Original assignee: Nanjing University of Posts and Telecommunications
Current assignee: Nanjing University of Posts and Telecommunications
Priority date: 2022-09-15
Filing date: 2022-09-15
Publication date: 2022-12-20

Abstract

The invention discloses an indoor passive positioning method for multi-modal users based on CSI, belonging to the technical field of wireless perception, and the method comprises the following steps: firstly, learning the feature distribution of an image sample by using a decoupling VAE model, and generating feature data of users with different posture simulated by an artificial sample; secondly, amplitude values and phase difference data of the three CSI antennas are extracted and given to three channels of the image to form an RGB image so as to obtain richer space physical information; and finally, establishing a positioning model based on the five-layer neural network to estimate the spatial position. The method converts the CSI positioning into the image classification problem, realizes the stable positioning of users with different body states in a data augmentation mode, simplifies the model complexity, reduces the operation difficulty and improves the robustness of the CSI positioning.

Description

Indoor passive positioning method for multi-posture users based on CSI

Technical Field

The invention belongs to the technical field of wireless perception, and particularly relates to an indoor passive positioning method for multi-modal users based on CSI.

Background

With the burst of intelligent house theory and the vigorous development of internet of things, the demand of indoor positioning has appeared in all kinds of reality scenes, because WIFI has characteristics such as with low costs, the deployment is convenient, the penetrability is good, so based on WIFI's passive location has gradually been aroused in recent years.

Thanks to the distribution of the CSI-Tool kit, by modifying the network card drive, researchers can obtain Channel State Information (CSI) from commercial WIFI devices, the CSI-containing Channel physical layer Information records the amplitude and phase of each subcarrier, and has better time resolution and spatial resolution, so as to better describe the communication link attribute of signals from a transmitting end to a receiving end, reflect multi-path Information such as reflection and diffraction when the signals propagate in an indoor environment, reveal Channel Information such as scattering, environmental attenuation, power attenuation, and the like, and have finer granularity, so that indoor positioning research based on the CSI gradually becomes the mainstream, and most attempts are made to analyze the fixed representation of the positioning Information in the signals through discriminant models such as Support Vector Machine (SVM), neural network, and the like, so as to achieve the target of passive positioning.

However, because the WIFI signal is a reaction to the overall physical layer of the space, not only the information of the indoor space is recorded, but also the information of the body shape, the shape, etc. of the positioning main body is recorded, most of the existing researches are carried out for a certain specific user, even if the positioning accuracy is excellent, the existing researches have certain limitations, and when the height and the weight of the user change, the shielding and penetrating influences caused by the signal will change under the condition of different body shapes. Due to the characteristic of fine granularity of the CSI data, when the discriminant model analyzed based on the fixed features meets the situation, the positioning accuracy requirement and stability of the system will be decreased rapidly, and in the past, in order to solve the problem, or by continuously collecting new user data, maintaining the fingerprint database for updating, or deepening the network to enhance the computation power, and exchanging the accuracy with longer time consumption is too complex, and time and labor are wasted.

The generative model is different from the discriminant model, and the purpose of parameter estimation is not to establish mapping between input and output by extracting features, but to learn the distribution of training data, so that the model can generate samples similar to the training data in the application stage, and the samples are usually very similar to real samples. The classical generation type model mainly comprises two categories of generation countermeasure Networks (GAN) and variation auto-encoders (VAE), wherein the GAN optimizes a generator based on a discriminator network, so that data distribution generated by the generator directly fits training data, mandatory requirements are not distributed, and the defects of mode collapse, difficulty in training and the like exist; the VAE is based on explicit distribution mapping and forcibly fits data to the Gaussian mixture distribution with finite dimensionality, so that the interpretability of the VAE to the features is better, model parameters can better describe the distribution of original features, and in recent years, the appearance of various VAEs enables the model to have better performance in interpretability and reconstruction reality.

How to utilize the characteristics of good interpretability and strong reconstruction authenticity of the VAE model and combine the channel state information with rich CSI to realize indoor positioning of users with different body states, achieve the purposes of simulating real data, saving the effort of maintaining a fingerprint library, improving the positioning precision and improving the stability and the applicability of CSI indoor passive positioning in a simpler way, and at present, no research on the aspect exists.

Disclosure of Invention

In order to solve the problems, the invention provides an indoor passive positioning method for multi-posture users based on CSI, which starts from the perspective of a fingerprint library, simulates the feature distribution of a real space by a generative model, decouples and learns mapping features, and generates artificial sample simulation real data, thereby overcoming the defect that the traditional mode is greatly influenced by the posture change of the user during positioning work; the CSI positioning is converted into the image classification problem, stable positioning of users with different body states is achieved in a data augmentation mode, model complexity is reduced, operation difficulty is reduced, and robustness of the CSI positioning is improved.

The invention relates to an indoor passive positioning method for multi-modal users based on CSI, which comprises the following steps:

step 1, acquiring CIS fingerprint data of an indoor under the condition of presence and absence of people, and constructing an existing fingerprint database;

step 2, based on the existing fingerprint library, generating artificial CSI samples simulating various posture users by using a beta-VAE generative model, and expanding the number of the samples of the fingerprint library;

step 3, preprocessing the CSI data obtained after the expansion in the step 2, and converting the preprocessed CSI data into an RGB image as a positioning decision sample;

and 4, constructing a convolutional neural network, and inputting the RGB image to perform positioning decision.

Further, the specific steps of step 1 are:

step 1-1, selecting a plurality of mark points in a target positioning area, measuring the layout of the area, calculating the distance between the mark points, configuring a wireless router to be placed at the corner of the area, using the wireless router as a signal receiving end and a transmitting end, and collecting CSI fingerprint data when an indoor unmanned scene is detected;

step 1-2, calibrating point positions of users with different body states in a region according to a secondary station, simultaneously acquiring data of a receiving end at each user calibration, and extracting amplitude and phase data of CSI information from the data;

and 1-3, performing stability calculation on the collected CSI fingerprint data of the unmanned scene and the CSI fingerprint data of users with different body states, and screening out antenna pairs insensitive to the positioning scene.

Further, the steps 1 to 3 are specifically as follows: the wireless router has 3x3 antenna pairs, errors of the antenna pairs are mutually independent, the change degree of CSI data on each antenna pair before and after the user is positioned is calculated, and sub-carrier serial numbers which are large in change degree and sensitive to a human body are screened out to form fingerprints.

Further, step 2 specifically comprises:

step 2-1, defining and collecting limited user data as an existing fingerprint library, training a beta-VAE generative model by using all data of the fingerprint library, generating an artificial sample, and performing training convergence by presetting a loss function; carrying out high-dimensional mapping on different body states of a user and CSI signal characteristics, and gradually learning the influence of body state change on the CSI characteristics in iteration to realize characteristic decoupling;

and 2-2, sampling Gaussian noise distribution, manually adjusting the distribution range, inputting a beta-VAE generation model, generating manual CSI samples for simulating users with various body states, and expanding the number of samples in a fingerprint database.

Further, the step 2-1 specifically comprises:

step 2-1-1, training an encoder, mapping the fingerprint data to be normal distribution, and outputting a two-dimensional hidden variable simulation mean variance in a dimensionality reduction mode;

step 2-1-2, training a decoder, sampling randomly from the normal distribution of the fingerprint data, and outputting an artificial sample through the ascending dimension of the decoder;

and 2-1-3, setting an information entropy bottleneck, modifying the weight of the loss function by using an annealing strategy in order to take reconstruction authenticity and interpretability into consideration, repeating the steps until iteration times are reached, and training the network parameters to be stable in convergence.

Further, step 3 specifically comprises:

3-1, filtering the CSI amplitude data by using a Hampel filter to remove outliers;

step 3-2, subtracting the phase according to the antenna arrangement to obtain phase difference data, and discharging carrier offset errors;

and 3-2, dividing the amplitude and phase difference information of the CSI vector of the fingerprint database into data packets according to the number of subcarriers, cutting out a data matrix, zooming to 0-255 gray scale values, distributing and giving three channels to form gray scale subgraphs, and combining the gray scale subgraphs to form an RGB image serving as a positioning decision sample.

Further, in step 4, training a convolutional neural network, setting output neurons as the number of calibration point locations, and inputting an RGB image for positioning decision;

the convolutional neural network comprises three convolutional layers and two full-connection layers, data is input from the first convolutional layer, and decision is made by the last full-connection layer; in the convolutional layers, each layer takes a modified linear unit as an activation function, max posing is used for pooling, the size of FeatureMap is reduced by half, the output of the previous layer is processed in a multi-layer nesting mode, and low-resolution features are obtained; in the fully-connected layer, the first layer flattens the input of the convolutional layer, reduces the dimensionality of FeatureMap at the same time, and inputs the convolutional layer to the last layer for decision making. The number of the neurons of the last full-connection layer is the same as the number of the point positions, when the output value of the neuron is the maximum value, the neuron is judged as a decision result of the network, the loss is obtained through cross entropy loss function calculation with label, and the network is optimized through back propagation.

The invention has the beneficial effects that:

1. the data used by the method is Channel State Information (CSI) data, the data granularity is fine, the information such as multipath effect, time-frequency response and amplitude-frequency response of signals transmitted in space is contained, the information can be collected through commercial WIFI equipment, the threshold is low, and compared with a traditional sensing mode, the method has the advantages that peripheral equipment does not need to be worn, the precision is higher, the cost is lower, the privacy is protected, the user is not sensed, and the like;

2. the influence of user posture change on a wireless channel is analyzed, a generative model is used for learning and simulating CSI data characteristic distribution, high-dimensional mapping is carried out on posture information and CSI amplitude phase information, the internal relation between the posture change and the channel information change is described, and the basic performance limit of a CSI positioning system in a multi-user posture scene is revealed after the mapping distribution is successful;

3. the invention researches the passive positioning based on the CSI, and particularly evaluates the performance of positioning decision for different user attitudes under the scenes of various users; if the CSI passive positioning system can be stably operated for users with different body states, the position of the user can be sensed in real time, the risk of invading privacy is avoided, the operation stability in user scenes with different body states is known, complex training and models are not needed, and the difficulty in deployment work is also very important to be reduced.

Drawings

FIG. 1 is a schematic overall flow chart of a step 2beta-VAE model;

FIG. 2 is a schematic diagram of an overall positioning decision flow;

FIG. 3 is a schematic diagram of a typical experimental environment.

Detailed Description

In order that the present invention may be more readily and clearly understood, a more particular description of the invention briefly described above will be rendered by reference to specific embodiments that are illustrated in the appended drawings.

The invention relates to an indoor passive positioning method aiming at multi-modal users based on CSI, which is implemented specifically by the following steps: firstly, fingerprint point locations are marked in an indoor environment, wireless signal data of different posture personnel on the point locations are collected, and CSI original data are extracted from the wireless signal data; secondly, dividing original CSI data to obtain amplitude values and phase data of different point locations of different users; secondly, building a beta-VAE generating model, respectively training by using amplitude and phase data, reconstructing artificial data, calculating similarity loss, and iteratively optimizing the model; then, gaussian noise distribution is adjusted, the distribution discrete range is expanded, sampling is carried out from the distribution, the sampling is used as a hidden variable to be input into a trained generative model, artificial amplitude phase data of a multi-body state is obtained, and therefore an initial fingerprint database is expanded; then, carrying out data preprocessing on the amplitude phase data in the fingerprint database, including removing outliers and removing channel noise; the preprocessed data is scaled to a gray value interval to form a gray image, and then the gray image is combined into an RGB image which is used as the input of a positioning decision network; and finally, training a convolutional neural network, converting the positioning problem into an image classification problem for decision making, and inputting an image to obtain the position estimation of the user.

An indoor passive positioning method for multi-modal users based on CSI specifically comprises the following steps:

step 1, acquiring indoor CSI fingerprint data, specifically:

step 1-1, marking a regular area in the experimental environment shown in fig. 3, marking 10 point locations according to the area size, measuring the area layout, calculating the distance between the marked points, and configuring a wireless router to be placed at the corner of the area to serve as a signal receiving and sending end. The acquisition equipment adopted by the method is two WDR4310 type routers carrying Atheros wireless network cards, the external parts of the two routers are respectively provided with 3 external antennas which are divided into a receiving end and a transmitting end, and the receiving end and the transmitting end are connected to a computer host through network cables, and the host sends commands through the network cables to control data acquisition. In order to ensure the stability of the WIFI channel in data transmission, the communication between the routers is set to be in a Monitor mode. The receiving end collects an original CSI data packet, and the analyzed CSI data stream is 3x 56 and represents the number of receiving end antennas, the number of transmitting end antennas and the number of subcarriers;

step 1-2, calibrating point positions of users with different heights and weights in a region according to the secondary station, simultaneously acquiring data of a receiving end, extracting amplitude and phase data of CSI information from the data, and not limiting the orientation action of a human body during acquisition;

step 1-3, performing stability calculation on the two CSI fingerprints, and screening out subcarrier sequences insensitive to a positioning scene; the method specifically comprises the following steps:

the equipment is provided with 31 x3 antenna pairs, the errors of the antenna pairs are independent, and the change degree of CSI data on each antenna pair before and after the positioning of a user is calculated; with 3 antennas at the transmitting end, the data on each antenna pair can be formulated as:

CSI _old ＝[S ₁ ，S ₂ ，S ₃ ]

wherein, the CSI _old And CSI _new Data before and after environmental change, S _i The value of the ith antenna pair is a row vector, S ', with dimension 1x 168' _i The method comprises the steps that a numerical value of an ith antenna pair after a user participates in positioning is shown, wherein i represents an antenna sequence of a transmitting end and is limited by equipment, i =1,2,3 and loc represents the position of the user under the condition of someone, and in order to reduce calculated amount, a plurality of indoor point positions are randomly selected;

the degree of difference between the changes before and after the CSI data can be modeled by the following formula:

wherein

Calculating the correlation coefficient of each antenna before and after the user to CSI data for the environment, wherein D (S) _i ) For the variance of the data of the ith antenna pair, the higher the correlation coefficient, the higher the number of the antenna pairAccording to the smaller the variation is, the proper threshold value beta is set

And if so, retaining the data, otherwise, removing the data of the antenna pair.

Step 2, expanding the user multi-modality fingerprint database, wherein the process is as shown in fig. 1, and specifically comprises the following steps:

step 2-1, limited user data is defined and collected as an existing fingerprint library, a beta-VAE generative model is trained by all data of the fingerprint library, an artificial sample is generated, training convergence is carried out through a preset loss function, different body states of a user and CSI signal characteristics are subjected to high-dimensional mapping, the model is gradually learned to influence of body state change on the CSI characteristics in iteration, and characteristic decoupling is achieved. The method specifically comprises the following steps:

and 2-1-1, training an encoder, mapping the fingerprint data to be normal distribution, and outputting a two-dimensional hidden variable simulation mean variance in a dimensionality reduction mode. Setting a CSI data vector as X, inputting the CSI data vector into an encoder of a beta-VAE generative model to obtain an output hidden variable vector set z;

step 2-1-2, training a decoder, sampling randomly from the normal distribution of the fingerprint data, and outputting an artificial sample through the dimensionality increase of the decoder; decoder for generating artificial CSI data vector X by inputting z into beta-VAE generating model _m′ Simultaneously computing a loss function, minimizing X and X _m′ Distance D (X, X) _m′ )；

Wherein D (X, X) _m′ ) Measuring the distance by KL divergence, specifically

D is the dimension of a hidden variable z which approaches normal distribution in iteration, mu and sigma ² The ith mean vector and the variance vector in normal distribution;

2-1-3, setting an information entropy bottleneck, modifying the weight of the loss function by using an annealing strategy in order to take reconstruction authenticity and interpretability into consideration, repeating the steps until iteration times, and training the convergence stability of network parameters;

the classical VAE model distinguishes the model in an explicit mapping modeDistribution q of _φ (z | X) true posterior probability distribution p of approximation data _θ (z | X) and by KL divergence D _KL (q _φ (z|X)||p _θ (z | X)) measures the degree of similarity between two distributions, which can be formulated as:

the objective function of a classical VAE can be formulated as:

wherein gamma (theta, phi; x) ⁱ ) Is the lower bound of variation of the edge probability, D _KL (q _φ (z|x ⁱ )||p _θ (z)) represents the KL divergence of the two posterior distributions, representing the loss of distribution similarity,

the target function of the beta-VAE is modified based on the formula, and the enhancement of the characteristic decoupling capacity is realized by introducing parameters beta and C, as shown in the formula:

the value of the beta controls the decoupling capacity of the model, and when the value is 1, the model is degraded into a classical VAE model; when the value is less than 1, the constraint on the information entropy bottleneck is weakened, the model focuses more on the reconstruction similarity of the data, and even if the output is extremely similar to the input, the characteristic change is not learned, and only the 'copy' of the original data is realized; when the value is greater than 1, the constraint on the information entropy bottleneck is enhanced, so that the model has better feature decoupling capability, the model is more sensitive to the variation in the sample features, the original data distribution is easier to learn, and the features are encoded into hidden variables in the model; meanwhile, the reconstruction similarity of the output data is compressed, so that a value C is introduced, the value C controls the information entropy bottleneck capacity, and when beta is greater than 1, the information entropy bottleneck needs to be improved, so that the model has the characteristic decoupling capacity and can also take into account the data representation capacity, and the system performance is improved. In actual training, an annealing strategy is adopted for C, the annealing strategy is gradually increased to a preset threshold value in an iterative process, and the attention degree of a model to decoupling and characterization is dynamically adjusted, so that a real sample is better simulated.

And 2-2, sampling Gaussian noise distribution, manually adjusting the distribution range, inputting a beta-VAE generating model, generating artificial CSI samples for simulating multiple posture users, and expanding the number of samples in a fingerprint library.

Step 3, preprocessing CSI data and converting the CSI data into images

Step 3-1, filtering the CSI amplitude data by using a Hampel filter to remove outliers, specifically:

because the acquisition device router does not directly obtain the CSI data, but obtains the CSI data through the driving calculation, and because the device Power amplification is not certain (PAU), the limited Power Amplifier resolution makes it difficult to keep the signal at a stable level after channel attenuation and Power Amplifier amplification, which causes a uniform deviation between the partial amplitude data of the subcarrier and the true value, but is limited in a single data packet, and the deviation of different data packets is random, so it is difficult to remove the threshold, and the Hampel filter is more suitable for solving the problem of outliers in the partial waveform data.

Step 3-1-1, in each amplitude vector, setting a window base number as k =3 and a window size as 2k +1, and calculating a median m of amplitude data | Am | in a window range of (i, i +2k + 1) _i And standard deviation σ _i Where i ∈ (1, 168), is the subcarrier index 56x3 for 3 antennas;

step 3-1-2) multiplying the standard deviation by the normal distribution standard deviation estimated value 1.4826 to obtain the estimated standard deviation sigma 'of each window' _i (ii) a Calculating the distance between the two adjacent rows according to the formula c = | | Am- _i+k -m _i |-3*σ′ _i Wherein | Am | _i+k For window intermediate values, calculating the medianIf the difference value with the median exceeds 3 estimation standard deviations, the difference value is regarded as exceeding a threshold value, and the median is replaced by the median;

and 3-1-3, processing the amplitude data in each data packet, and removing outliers in the amplitude of the whole link.

Step 3-2, obtaining phase difference data by differentiating the phases according to the antenna arrangement, and eliminating carrier offset errors, specifically:

the error of the phase data is linear error, and comes from the time frequency offset problem caused by SFO and CFO, and the measured phase value of the ith subcarrier of a formula is obtained according to the definition

Wherein < Ph _i For the true phase value, Δ t is the time offset, β is the position phase offset, k _i Represents the ith subcarrier index, k is the (1, 56), and N is the fast Fourier transform sampling number 64;

because the phase shift caused by SFO and CFO exists on each antenna and the number of subcarriers of each antenna is 56, the relative offset can be eliminated by adopting the method of carrier phase difference between antennas, and the phase difference data can be obtained

And 3-3, dividing the amplitude and phase difference information of the CSI vector of the fingerprint database into data packets according to the number of subcarriers, cutting out a data matrix, zooming to 0-255 gray scale values, distributing and giving three channels to form gray scale subgraphs, and combining the gray scale graphs to form an RGB image serving as a positioning decision sample.

Step 4, the neural network decides the fingerprint position, and the flow is as shown in figure 2;

and 4-1, training a convolutional neural network, setting output neurons as the number of the calibration point positions, and inputting an RGB image to perform positioning decision. The method comprises the following specific steps:

the network has five layers including three convolution layers and two full connection layers, and data is input from the first convolution layer and decided by the last full connection layer. In the convolutional layers, each layer takes a modified linear unit as an activation function, the Max Pooling is used for pooling, the size of FeatureMap is reduced by half, and the output of the previous layer is processed in a multi-layer nesting mode to obtain low-resolution characteristics. And (4) flattening the maximum pooling layer by fully connecting the layers, and outputting a classification result to make a decision after two layers of full connection.

The indoor passive positioning method for the multi-posture user based on the CSI starts from the perspective of a fingerprint library, simulates the feature distribution of a real space by a generating model, decouples and learns the mapping features, and generates artificial sample simulation real data, so that the defect that the traditional mode is greatly influenced by the posture change of the user during positioning work is overcome; the data source is expanded into amplitude and phase difference, three channels of images are formed by extracting amplitude and phase difference data of the three CSI antennas to form RGB images, a sample conversion positioning problem is formed in an image fusion mode to classify the images, space multipath information is further mined, and model decision efficiency and accuracy are improved; compared with the traditional positioning method, the method does not need an excessively complex decision model and maintenance and update of the fingerprint database, simplifies the complexity of the model, reduces the operation difficulty and improves the robustness of CSI positioning.

The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention further, and all equivalent variations made by using the contents of the present specification and the drawings are within the scope of the present invention.

Claims

1. An indoor passive positioning method for multi-modal users based on CSI is characterized by comprising the following steps:

step 1, acquiring indoor CIS fingerprint data under the condition of presence and absence of people, and constructing an existing fingerprint database;

step 3, preprocessing the CSI data obtained after the expansion in the step 2, and converting the preprocessed CSI data into RGB images as positioning decision samples;

2. The indoor passive positioning method for multi-modal users based on CSI as claimed in claim 1, wherein the specific steps of step 1 are:

1-1, selecting a plurality of mark points in a target positioning area, measuring the area layout, calculating the distance between the mark points, configuring a wireless router to be placed at the corner of the area, and collecting CSI fingerprint data when an indoor unmanned scene is used as a signal receiving end and a transmitting end;

3. The indoor passive positioning method for multi-modal users based on CSI as claimed in claim 2, wherein steps 1-3 specifically are: the wireless router is provided with 3x3 antenna pairs, errors of the antenna pairs are mutually independent, the change degree of CSI data on each antenna pair before and after the positioning of a user is calculated, and subcarrier serial numbers which are large in change degree and sensitive to a human body are screened out to form fingerprints.

4. The indoor passive positioning method for multi-modal users based on CSI as claimed in claim 1, wherein step 2 specifically comprises:

5. The indoor passive positioning method for the multi-modal user based on the CSI as claimed in claim 4, wherein the step 2-1 specifically comprises:

step 2-1-1, training an encoder, mapping fingerprint data into normal distribution, and outputting a two-dimensional hidden variable simulation mean variance in a dimensionality reduction mode;

and 2-1-3, setting an information entropy bottleneck, modifying the weight of the loss function by using an annealing strategy in order to take reconstruction authenticity and interpretability into consideration, repeating the steps until iteration times, and training the convergence stability of network parameters.

6. The indoor passive positioning method for multi-modal users based on CSI as claimed in claim 1, wherein step 3 specifically comprises:

and 3-2, dividing the amplitude and phase difference information of the CSI vector of the fingerprint database into data packets according to the number of subcarriers, cutting out a data matrix, zooming to 0-255 gray scale values, distributing and giving three channels to form gray scale sub-images, and combining the gray scale sub-images to form RGB images serving as positioning decision samples.

7. The indoor passive positioning method for the multi-modal user based on the CSI as recited in claim 1, wherein in step 4, a convolutional neural network is trained, an output neuron is set as a calibration point location number, and an RGB image is input for positioning decision;

the convolutional neural network comprises three convolutional layers and two full-connection layers, data is input from the first convolutional layer, and decision is made by the last full-connection layer; in the convolutional layers, each layer takes a modified linear unit as an activation function, max posing is used for pooling, the size of FeatureMap is reduced by half, the output of the previous layer is processed in a multi-layer nesting mode, and low-resolution features are obtained; in the full-connection layer, the first layer flattens the input of the convolution layer, reduces the dimensionality of FeatureMap at the same time, and inputs the input to the last layer for decision making; the number of neurons of the last full-connection layer is the same as the number of point positions, when the output value of the neuron is the maximum value, the neuron is judged as a decision result of the network, the decision result and the label are calculated through a cross entropy loss function to obtain loss, and the network is optimized through back propagation.