Point cloud human body posture reconstruction method and system based on attention mechanism and convolutional neural network

Abstract

The invention discloses a point cloud human body posture reconstruction method based on an attention mechanism and a convolutional neural network, and the method comprises the steps: carrying out the mean cancellation, DC component removal, CFAR and IAA super-resolution angle measurement of radar echo data, obtaining the dense representation of a human body point cloud, and carrying out the collapse projection of the point cloud information to three groups of orthogonal planes, and taking a picture form as network input. Meanwhile, a neural network composed of a channel fusion module, a feature deep mining module and a posture decoding module is constructed, a deep learning technology is used for fitting input pictures and human body joint points, and finally real-time three-dimensional joint point reconstruction of human body postures is achieved.

Description

Technical Field

[0001] The present invention relates to the field of millimeter wave radar smart home and smart health care technology, and in particular to a point cloud human posture reconstruction method and system based on attention mechanism and convolutional neural network. Background Art

[0002] Human posture reconstruction technology originated from computer vision, and has shown great potential in the fields of human-computer interaction, smart home, gesture recognition, virtual reality, etc. With the development of deep learning technology, vision-based human posture reconstruction technology has become increasingly mature. However, vision-based systems have inherent defects, and their applications are greatly limited by the environment. They cannot work properly under poor lighting conditions and occlusion, and there are privacy and ethical risks. Therefore, in the field of posture reconstruction, non-contact human posture reconstruction methods based on radio frequency signals have emerged.

[0003] Radar-based sensing technology is less affected by environmental conditions such as light, temperature, and humidity, and does not pose a risk of privacy leakage. It does not require direct contact with the object being measured, freeing the object from the constraints of electronic instruments and equipment, making it an ideal monitoring method.

[0004] In radar detection, compared with traditional microwave frequency radar, millimeter wave radar has the natural properties of small wavelength and large bandwidth, which makes it easier to achieve system miniaturization design and high-precision measurement of targets. At the same time, it has the common advantages of radar, such as all-weather and all-day environmental adaptability, non-contact detection, and non-privacy infringement.

[0005] With the help of the good detection performance of millimeter-wave radar, this method focuses on the needs of non-contact human posture reconstruction in the fields of smart home and smart health care, conducts research on reconstruction technology based on millimeter-wave radar, breaks through the key technologies of target fall detection in diverse scenarios and human posture reconstruction under strong clutter backgrounds, and meets the urgent needs of application fields such as smart medical care, smart elderly care, and smart home. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a point cloud human posture reconstruction method based on attention mechanism and convolutional neural network, which adopts millimeter wave radar to realize real-time reconstruction of human posture.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] The method for reconstructing human body posture from point cloud based on attention mechanism and convolutional neural network provided by the present invention comprises the following steps:

[0009] S1: Collect radar echo data of human body;

[0010] S2: Process the radar echo data to obtain human body point cloud data;

[0011] S3: Projecting the human body point cloud data to obtain an orthogonal plane image;

[0012] S4: Construct a convolutional neural network model under the attention mechanism and obtain the trained human posture reconstruction network model;

[0013] S5: Input the orthogonal plane image data into the human body posture reconstruction network model for processing to obtain reconstructed human body posture data.

[0014] Further, the method for processing data in S2 is performed according to the following steps:

[0015] S21: using mean cancellation and DC component removal methods, static clutter is filtered out of the acquired intermediate frequency signal, and a two-dimensional fast Fourier transform is performed to obtain a range-velocity spectrum;

[0016] S22: Perform multiple two-dimensional CFAR detections based on the range-velocity spectra obtained in S21 to obtain range and velocity information, perform IAA super-resolution angle measurement in the pitch and azimuth channel dimensions to obtain target angle information, and complete the information representation of the target in the polar coordinate system;

[0017] S23: Mapping the data to a three-dimensional Cartesian coordinate system according to a coordinate transformation principle to obtain point cloud data information;

[0018] S24: Perform DBSCAN clustering based on the point cloud data information obtained in S23 to remove outlier noise points.

[0019] Further, the projection of the human body point cloud data in S3 is performed according to the following steps:

[0020] According to the representation of point cloud data information in three-dimensional space, it is projected onto three sets of orthogonal planes, xy, xz, and yz. The scene grid is discretized and the point cloud data information is saved in three sets of images. At the same time, the joint information coordinates captured by the motion capture device are used as labels to complete the network input and label parts in the data set, and the radar frame number is divided into training set and validation set.

[0021] Further, the construction of the convolutional neural network model in S4 is carried out according to the following steps:

[0022] S41: Channel information fusion, with convolutional layers as the backbone, integrating SE attention and deep high-resolution representation modules, extracting spatial features from point cloud projections and fusing information obtained from three sets of orthogonal planes;

[0023] S42: Feature deep mining, with ResNet18 as the network backbone, four groups of residual blocks with CA attention mechanism are stacked to perform deep mining based on the multi-scale features obtained in S41, complete the feature high-density representation, and obtain the feature encoding vector;

[0024] S43: Posture decoding, using a multi-layer perceptron to process the feature encoding vector described in S42, and output the joint point representation of the human body posture in the current frame.

[0025] Furthermore, the convolutional neural network model also includes a training process, and the specific steps are as follows:

[0026] The joint point representation of the human body posture and the corresponding label in the training set are obtained and sent into the loss function together to obtain the loss result. The gradient backpropagation is performed according to the loss result to update the parameter value of the neural network model. The above process is repeated according to the hyperparameters to finally obtain the convolutional neural network model under the attention mechanism.

[0027] Further, the method further comprises the following steps:

[0028] The loss function uses the Euclidean distance between the joint point and the label, that is, the L2 norm, and the loss function is calculated according to the following formula:

[0029]

[0030] Among them, P i represents the i-th joint point obtained by model reasoning, represents the i-th joint point in the label, and N is the number of joint points.

[0031] Furthermore, the two-dimensional CFAR detection is specifically performed in the following manner:

[0032] The threshold value of the evaluation unit is determined based on the reference unit level value and the false alarm rate, and compared with the actual level value. If the actual level value exceeds the threshold value, it is considered that there is a potential target in the current unit, and its distance and speed information is extracted and saved in the candidate target point queue, and the current number of CFAR operations is recorded.

[0033] Further, the channel information fusion is specifically performed according to the following steps:

[0034] Separate channels through group convolution;

[0035] Batch normalize the point cloud information;

[0036] The channel weight is obtained through SE channel attention;

[0037] Use multi-layer convolution and up-down sampling to fuse features of different scales, and use the CA spatial attention mechanism to filter features;

[0038] Output feature cube.

[0039] The point cloud human posture reconstruction system based on attention mechanism and convolutional neural network provided by the present invention includes a data acquisition module, a human point cloud data generation module, a human image data generation module, a convolutional neural network model construction module, and a human posture data reconstruction module;

[0040] The data acquisition module is used to collect radar echo data of the human body;

[0041] The human body point cloud data generating module is used to process the radar echo data to obtain human body point cloud data;

[0042] The human body image data generation module is used to project the human body point cloud data to obtain an orthogonal plane image;

[0043] The convolutional neural network model building module is used to build a convolutional neural network model under the attention mechanism and obtain a trained human posture reconstruction network model;

[0044] The human body posture data reconstruction module is used to input the orthogonal plane image data into the human body posture reconstruction network model for processing to obtain reconstructed human body posture data.

[0045] Further, the human body point cloud data generation module includes an adaptive two-dimensional CFAR detection module and an IAA super-resolution detection module;

[0046] The adaptive two-dimensional CFAR detection is used to extract human body echo data from radar echo data;

[0047] The IAA super-resolution detection module extracts the angle information of candidate target points from the human body echo data to obtain the point cloud feature expression of the human body.

[0048] The beneficial effects of the present invention are:

[0049] The point cloud human posture reconstruction method based on attention mechanism and convolutional neural network provided by the present invention generates human point cloud for radar echo signal through mean cancellation, incoherent accumulation, constant false alarm detection, IAA super-resolution angle measurement and DBSCAN clustering algorithm, and then collapses and projects the three-dimensional point cloud to three groups of orthogonal planes xy, xz, and yz, and sends them to the deep learning network in the form of three pictures, and finally obtains human posture information. The method can accurately identify various human actions and falls in some special cases, protect user privacy, is not affected by the environment, and can reproduce human posture information in real time.

[0050] This method can be used for single-person posture detection. In a contactless environment, it uses privacy-protected electromagnetic information to reconstruct the user's posture in real time, effectively recording the user's behavior dynamics, and the user can replay the data for self-monitoring of posture behavior.

[0051] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0053] Figure 1 Build a flow chart for the point cloud human pose reconstruction method.

[0054] Figure 2 Flowchart for dense point cloud generation.

[0055] Figure 3 It is the point cloud collapse projection method.

[0056] Figure 4 This is the structure diagram of the channel fusion module.

[0057] Figure 5 It is a residual block structure.

[0058] Figure 6 is the posture decoder structure.

[0059] Figure 7 The network training results are shown.

[0060] Figure 8 This is a display of network test results. DETAILED DESCRIPTION

[0061] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0062] Example 1

[0063] like Figure 1 As shown, this embodiment describes in detail a point cloud human posture reconstruction method based on an attention mechanism and a convolutional neural network, comprising the following steps:

[0064] Step 1, data and preprocessing process: receive millimeter-wave radar echo data, process the echo data into a data cube of sampling point dimension*chirp dimension*antenna dimension, and perform corresponding data preprocessing to obtain point cloud information. Use mean cancellation and DC component removal to filter static clutter on the obtained intermediate frequency signal, and perform two-dimensional fast Fourier transform to obtain a range-velocity spectrogram.

[0065] The mixer in the millimeter-wave radar hardware mixes and filters the transmit and receive signals to obtain an intermediate frequency signal, which is then processed into a three-dimensional complex cube of sampling point dimension*chirp dimension*antenna dimension. First, the data of each sampling point is averaged and subtracted from the chirp dimension to remove static clutter in the echo to improve the reliability of the point cloud. Then, two fast Fourier transforms are performed in the sampling point dimension and the chirp dimension to obtain a joint spectrum of distance and velocity of the echo data.

[0066] Step 2: The human body point cloud data is dense point cloud information data. The dense point cloud is generated. The specific process is as follows: Figure 2 As shown, the following steps are included:

[0067] Two-dimensional CFAR detection S1: Perform a two-dimensional CFAR detection based on the distance and speed joint spectrum obtained in step 1, that is, determine the threshold value of the evaluation unit based on the reference unit level value and false alarm rate, and compare it with the actual level value. If the actual level value exceeds the threshold value, it is considered that there is a potential target in the current unit, and the distance and speed information is extracted and saved in the candidate target point queue, and the number of CFAR operations currently performed is recorded;

[0068] Filter out target point S2: After obtaining the candidate target queue through the above operation, set the level value of the unit where its index is located in the joint spectrum of distance and speed to zero in turn;

[0069] Condition judgment S3: judge whether the number of completed CFARs reaches the preset maximum number of CFARs or no new candidate target points are generated. If one of the conditions is met, the distance and speed information of the candidate targets are screened and the point cloud generation is terminated. If none of the conditions is met, return to S1 and repeat S1, S2 and S3 to obtain the candidate target points.

[0070] If the condition required by condition judgment S3 is not met and the process returns to S1, the candidate target points obtained by performing two-dimensional CFAR detection S1 again are in parallel relationship with the candidate target points in the queue, and the former are added to the candidate target point queue in turn.

[0071] Then, according to the extracted indexes of the distance and speed dimensions, the three-dimensional complex cube in step 1 is converted into multiple one-dimensional complex vectors, i.e., channel dimension vectors, and converted into two sets of complex vectors in elevation and azimuth according to the antenna channel relationship. IAA super-resolution angle measurement is performed in the elevation and azimuth channel dimensions to obtain angle information, thus completing the information representation of the potential target in the polar coordinate system.

[0072] Step 3, point cloud and image conversion: According to the coordinate transformation principle, the data is mapped to the three-dimensional Cartesian coordinate system to obtain point cloud information, and the DBSCAN clustering algorithm is used to remove outlier noise points. Then, the point cloud information is collapsed (i.e., projected) using the orthogonal space, and the three sets of scene plane grids xy, xz, and yz are discretized, and the point cloud is filled into the grid to obtain three sets of pixelated scene images, such as Figure 3 shown.

[0073] The beneficial effects of this step are:

[0074] Compared with the method of voxelizing the space, projection can effectively improve the density information of the input space, that is, remove the redundancy of the input information, effectively reduce the number of network parameters and computational complexity, and improve the real-time performance of network reconstruction.

[0075] Step 4: Channel information fusion: With the convolutional layer as the main backbone, the SE attention and deep high-resolution representation modules are integrated to extract spatial features from the point cloud projection and fuse the information obtained from three sets of orthogonal planes; considering the differences between channels in terms of information volume, feature scale, etc., a channel fusion module is designed.

[0076] First, the channels are separated by group convolution, and the number of channel groups is 3;

[0077] Then the point cloud information is batch normalized to ensure the regularity of data distribution;

[0078] Different weights of channels are obtained through SE channel attention. This embodiment uses SE channel attention to assign different weights to channels to ensure the expression of different features is differentiated;

[0079] Then, multi-layer convolution and up-down sampling are used to fuse features of different scales, and the CA spatial attention mechanism is used to filter features;

[0080] Finally, the point cloud information feature cube is generated and output. Figure 4 shown.

[0081] Among them, the role of group convolution and SE channel attention is to strip off the channel dependency in the convolution process and avoid introducing noise information from the corresponding positions of other pictures during the convolution process, that is, only extracting features from the current picture; secondly, due to the dense characteristics of point cloud information, deep high-resolution representation and CA spatial attention modules are introduced to enable the model to maintain high spatial resolution when extracting features in different dimensions, prevent feature mixing, and thus improve the reconstruction accuracy.

[0082] Step 5: Feature deep mining: The residual block in ResNet18 is used as the module backbone, and its structure is as follows: Figure 5 As shown in the figure, the residual link is used to ensure the effective transmission of the gradient; the main structure is formed by stacking four groups of residual blocks with CA attention mechanism, and the multi-scale features obtained in step 4 are deeply mined to complete the high-density representation of the features and obtain the feature encoding vector;

[0083] Among them, the residual block is composed of three groups of convolutional layers and CA spatial attention. The two groups of convolutional layers on the top are for deep feature extraction, and the parallel convolutional layers on the bottom are to ensure the effectiveness of the gradient during back propagation. CA spatial attention is to highlight the areas with richer information in the feature map, which plays a role in suppressing the expression of redundant features.

[0084] Step 6, posture decoding: Based on the universal approximation theorem of the multi-layer perceptron, it is used here to process the feature encoding vector described in step 2, and the extracted features are mapped to joint point coordinates, that is, the decoding operation, and the joint point representation of the human body posture in the current frame is output and fitted with the label. The multi-layer perceptron consists of three fully connected layers and a ReLU activation function, and its structure is as follows Figure 6 shown.

[0085] This embodiment uses the point cloud information obtained by the motion capture device to perform real-time annotation to obtain data label information.

[0086] The model in step 6 is trained. The specific process includes the following steps:

[0087] The joint point representation of the human body posture obtained according to the above method is sent into the loss function together with the corresponding label in the training set to obtain the loss result, and the gradient back propagation is performed according to the loss result to update the parameter value of the neural network model. The above process is repeated according to the hyperparameters to finally obtain the convolutional neural network model under the attention mechanism.

[0088] In this embodiment, the point cloud generated by different human body movements is filtered out by the DBSCAN clustering algorithm to remove noise points and orthogonal space projection, and three groups of images for each movement are obtained as the input of the network module. The joint information coordinates synchronously captured by the motion capture device are used as corresponding labels to construct radar point cloud projection images and posture data sets, which are divided into training sets and verification sets; and both the images and labels are stored in the database to form a data set, and then the radar frame number is divided into training sets and verification sets in a ratio of 7:3.

[0089] The inference results of the trained model for different actions in the validation set are as follows: Figure 7 shown.

[0090] The trained model is deployed locally, and real-time posture reconstruction is performed based on radar data to obtain real-time inference results such as Figure 8 As shown in the figure, the first line is the measured camera data, which is used to compare the reconstruction results and does not participate in model reasoning. The second line is the real-time reconstruction result, which is the reasoning result of the model and the final effect achieved by this patent. The third line is the point cloud result generated by the dense point cloud generation method described in this patent for different actions, showing the quality of the point cloud and data set. The deployed real-time system can reach 50-60 frames per second on the NVIDIA GeForce GTX 1060 graphics card, which proves the effectiveness and real-time performance of the model.

[0091] Example 2

[0092] In this embodiment, the millimeter wave radar and the motion capture device constitute an acquisition system. The point cloud human posture reconstruction system based on the attention mechanism and the convolutional neural network provided in this embodiment includes a data acquisition module, a human point cloud data generation module, a human image data generation module, a convolutional neural network model construction module, and a human posture data reconstruction module;

[0093] The data acquisition module is used to collect radar echo data of the human body;

[0094] The human body point cloud data generating module is used to process the radar echo data to obtain human body point cloud data;

[0095] The human body image data generation module is used to project the human body point cloud data to obtain an orthogonal plane image;

[0096] The convolutional neural network model building module is used to build a convolutional neural network model under the attention mechanism and obtain a trained human posture reconstruction network model;

[0097] The human body posture data reconstruction module is used to input the orthogonal plane image data into the human body posture reconstruction network model for processing to obtain reconstructed human body posture data.

[0098] The human body point cloud data generation module includes an adaptive two-dimensional CFAR detection module and an IAA super-resolution detection module; wherein, the adaptive two-dimensional CFAR detection is performed through two-dimensional CFAR detection, target point filtering, condition judgment and other processing, and its function is to effectively extract radar echo data, make up for the defect that the simple CFAR threshold is too high and the weak target is submerged, and extract as much human body echo data as possible, so as to reduce the pressure on the subsequent network end;

[0099] The IAA super-resolution detection module is to accurately extract the angle information of candidate target points. Due to the limitation of radar hardware, the angle information obtained by fast Fourier transform has a large error. The IAA super-resolution algorithm can effectively solve the above problem and obtain more accurate information, which is beneficial to the expression of point cloud features.

[0100] The convolutional neural network model building module includes channel information fusion, feature depth mining and posture decoding;

[0101] Among them, the channel information fusion can be divided into group convolution to strip off channel dependency, batch normalization to ensure that data distribution is convenient for gradient direction propagation, SE channel attention module to ensure differentiated feature expression on different channels, and deep high-resolution representation module to extract and fuse features at different scales and ensure high spatial resolution. The above different structures are combined into a channel information fusion module, which solves the irrelevance between input channels, fuses features with fewer parameters, and ensures the deployability and real-time performance of the model.

[0102] The feature depth mining module uses the residual block as the basic structure, which has the beneficial effect of giving the model a network depth that can be determined by self-learning and facilitating the effectiveness of gradient back propagation, avoiding gradient diffusion and explosion to a certain extent. A CA spatial attention module is added between the residual blocks, which has the beneficial effect of ensuring the differentiation of feature expression in space, highlighting the expression of effective features while suppressing redundant features, so that the model can identify and effectively utilize features;

[0103] The posture decoding module is composed of a multi-layer perceptron, which uses the universal approximation theorem to decode the features extracted by the previous module and map them to the three-dimensional representation of the human body joints. It is a key part of the transformation between features and final results.

[0104] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A point cloud human posture reconstruction method based on attention mechanism and convolutional neural network, characterized by: The following steps are involved: S1: Collect radar echo data of human body; S2: Process the radar echo data to obtain human body point cloud data; S3: Projecting the human body point cloud data to obtain an orthogonal plane image; S4: Construct a convolutional neural network model under the attention mechanism and obtain the trained human posture reconstruction network model; S5: Input the orthogonal plane image data into the human body posture reconstruction network model for processing to obtain reconstructed human body posture data.

2. The method for reconstructing human body posture from point cloud based on attention mechanism and convolutional neural network according to claim 1, characterized in that: The method for processing data in S2 is performed according to the following steps: S21: using mean cancellation and DC component removal methods, static clutter is filtered out of the acquired intermediate frequency signal, and a two-dimensional fast Fourier transform is performed to obtain a range-velocity spectrum; S22: Perform multiple two-dimensional CFAR detections based on the range-velocity spectra obtained in S21 to obtain range and velocity information, perform IAA super-resolution angle measurement in the pitch and azimuth channel dimensions to obtain target angle information, and complete the information representation of the target in the polar coordinate system; S23: Mapping the data to a three-dimensional Cartesian coordinate system according to a coordinate transformation principle to obtain point cloud data information; S24: Perform DBSCAN clustering based on the point cloud data information obtained in S23 to remove outlier noise points.

3. The method for reconstructing human body posture from point cloud based on attention mechanism and convolutional neural network according to claim 1, characterized in that: The projection of the human body point cloud data in S3 is performed according to the following steps: According to the representation of point cloud data information in three-dimensional space, it is projected onto three sets of orthogonal planes, xy, xz, and yz. The scene grid is discretized and the point cloud data information is saved in three sets of images. At the same time, the joint information coordinates captured by the motion capture device are used as labels to complete the network input and label parts in the data set, and the radar frame number is divided into training set and validation set.

4. The method for reconstructing human body posture from point cloud based on attention mechanism and convolutional neural network according to claim 1, characterized in that: The construction of the convolutional neural network model in S4 is carried out according to the following steps: S41: Channel information fusion, with convolutional layers as the backbone, integrating SE attention and deep high-resolution representation modules, extracting spatial features from point cloud projections and fusing information obtained from three sets of orthogonal planes; S42: Feature deep mining, with ResNet18 as the network backbone, four groups of residual blocks with CA attention mechanism are stacked to perform deep mining based on the multi-scale features obtained in S41, complete the feature high-density representation, and obtain the feature encoding vector; S43: Posture decoding, using a multi-layer perceptron to process the feature encoding vector described in S42, and output the joint point representation of the human body posture in the current frame.

5. The method for reconstructing human body posture from point cloud based on attention mechanism and convolutional neural network according to claim 4, characterized in that: The convolutional neural network model also includes a training process, and the specific steps are as follows: The joint point representation of the human body posture and the corresponding label in the training set are obtained, and are sent into the loss function together to obtain the loss result. The gradient backpropagation is performed according to the loss result to update the parameter value of the neural network model. The above process is repeated according to the hyperparameters to finally obtain the convolutional neural network model under the attention mechanism.

6. The method for reconstructing human body posture from point cloud based on attention mechanism and convolutional neural network according to claim 5, characterized in that: The following steps are also included: The loss function uses the Euclidean distance between the joint point and the label, that is, the L2 norm, and the loss function is calculated according to the following formula: Among them, P i represents the i-th joint point obtained by model reasoning, represents the i-th joint point in the label, and N is the number of joint points.

7. The method for reconstructing human body posture from point cloud based on attention mechanism and convolutional neural network according to claim 1, characterized in that: The two-dimensional CFAR detection is specifically performed in the following manner: The threshold value of the evaluation unit is determined based on the reference unit level value and the false alarm rate, and compared with the actual level value. If the actual level value exceeds the threshold value, it is considered that there is a potential target in the current unit, and its distance and speed information is extracted and saved in the candidate target point queue, and the current number of CFAR operations is recorded.

8. The method for reconstructing human body posture from point cloud based on attention mechanism and convolutional neural network according to claim 4, characterized in that: The channel information fusion is specifically performed according to the following steps: Separate channels through group convolution; Batch normalize the point cloud information; The channel weight is obtained through SE channel attention; Use multi-layer convolution and up-down sampling to fuse features of different scales, and use the CA spatial attention mechanism to filter features; Generate and output a feature cube.

9. A point cloud human posture reconstruction system based on attention mechanism and convolutional neural network, characterized by: It includes data acquisition module, human point cloud data generation module, human image data generation module, convolutional neural network model construction module, and human posture data reconstruction module; The data acquisition module is used to collect radar echo data of the human body; The human body point cloud data generating module is used to process the radar echo data to obtain human body point cloud data; The human body image data generation module is used to project the human body point cloud data to obtain an orthogonal plane image; The convolutional neural network model building module is used to build a convolutional neural network model under the attention mechanism and obtain a trained human posture reconstruction network model; The human body posture data reconstruction module is used to input the orthogonal plane image data into the human body posture reconstruction network model for processing to obtain reconstructed human body posture data.

10. The point cloud human posture reconstruction system based on attention mechanism and convolutional neural network according to claim 9, characterized in that: The human body point cloud data generation module includes an adaptive two-dimensional CFAR detection module and an IAA super-resolution detection module; The adaptive two-dimensional CFAR detection is used to extract human body echo data from radar echo data; The IAA super-resolution detection module extracts the angle information of candidate target points from the human body echo data to obtain the point cloud feature expression of the human body.

Images