CN108764050A - Skeleton Activity recognition method, system and equipment based on angle independence - Google Patents

Abstract

The invention relates to the field of human behavior recognition, in particular to an angle-independent skeleton behavior recognition method, system and equipment, which only improves the accuracy of angle-independent skeleton behavior recognition. The angle-independent skeleton behavior recognition method of the present invention includes: designing a view-specific subnet based on the skeleton sequence of each view, focusing on key joints and key frames through spatial attention and time domain attention modules, and using multiple The layer-long short-term memory network learns the discriminative features of each view sequence; the output features of each specific view subnetwork are concatenated as the input of the public subnetwork, and the angle-independent features are further learned through the bidirectional long-term short-term memory network. The module focuses on key perspectives; a regularized cross-entropy loss function is proposed to promote the joint learning of multiple network modules. The invention effectively improves the recognition accuracy, and can automatically focus on learning the characteristics of the viewing angle with more information.

Description

Skeleton Behavior Recognition Method, System, and Equipment Based on Angle-Independence

technical field

The invention relates to the field of human behavior recognition, in particular to a skeleton behavior recognition method, system and equipment based on angle independence.

Background technique

As an important research field of computer vision, human behavior recognition is a system that classifies and recognizes human behavior through input data. From the point of view of the input and output of the system, the input is one or more data related to human behavior, and the data is a time series obtained by different sensors through sampling at a certain frequency. The output of the system is the recognition and classification result of human behavior. Generally speaking, the input of human action recognition system is divided into four forms of data: RGB time series, skeleton time series, depth map video and infrared video. With the rapid development of depth sensors, the acquisition of skeleton data is becoming more and more convenient and fast. Therefore, skeleton-based human behavior recognition has also received more and more attention. In behavior recognition research, complex data changes are a major challenge in this research. Skeleton data is more robust than traditional RGB data, but in terms of viewing angle changes, skeleton behavior recognition based on angle independence and angle-based Unrelated RGB action recognition is equally challenging.

The key to the angle-independent skeleton behavior recognition technology lies in three parts: on the one hand, how to extract highly discriminative features, on the one hand, how to reduce the impact of angle changes on behavior recognition, and on the other hand, how to use time-domain correlation Sex models the dynamics of behavioral movements. According to the modeling method of human behavior recognition, it can be divided into traditional method modeling and deep learning method modeling. The traditional method is divided into two processes: feature representation and action recognition and understanding. Traditional feature extraction methods include HOG (Histogram of Oriented Gradient), SIFT (Scale-Invariant Feature Transform) and other feature extraction methods; the extracted features are usually classified by common classification algorithms, such as Support Vector Machine (Support Vector Machine, SVM) identify. With the introduction and development of deep learning theory, deep learning algorithms are increasingly used in the research of human behavior recognition. Recurrent Neural Network (RNN) is dedicated to modeling video information, using the last few hidden layer data as the input of the current moment, allowing information in the time dimension to be preserved; based on Long Short Term Memory (Long Short Term The recurrent neural network (RNN) of Memory, LSTM) is an extension of the ordinary RNN model, which mainly solves the gradient disappearance phenomenon in the RNN model. Therefore, most of the human skeleton behavior recognition in recent years utilizes the LSTM-based RNN model in deep learning.

The main problem of the current LSTM-based angle-independent skeleton behavior recognition method is that it does not fully exploit all the information of the given sequence and the recognition accuracy needs to be improved. Specifically, LSTM is used to extract the discriminative features of video sequences under a single view, thus ignoring the connection between videos of the same behavior under multiple views; at the same time, each joint point in the multi-view skeleton data, each frame and each Views have different effects on angle-independent skeleton behavior recognition, but only through LSTM modeling methods, different composition structures of skeleton data have the same contribution to behavior recognition, which limits the accuracy of angle-independent skeleton behavior recognition.

Contents of the invention

In order to solve the above problems in the prior art, the present invention proposes an angle-independent skeleton behavior recognition method, system and equipment, which improves the accuracy of angle-independent skeleton behavior recognition.

In one aspect of the present invention, a method of skeleton behavior recognition based on angle independence is proposed, including:

Input the skeleton time series to be recognized into the trained skeleton behavior recognition model according to different perspectives;

Using the trained skeleton behavior recognition model to calculate the behavior category probability of the skeleton time series to be recognized;

in,

The skeleton behavior recognition model includes: a preset number of specific perspective subnets and public subnets;

The training method of described skeleton behavior recognition model, comprises the following steps:

Step S1, for each of the specific view subnets, input a frame of training data corresponding to the specific view, respectively calculate the spatial domain attention weight, the temporal domain attention weight, and then calculate the discriminative features of the specific view subnetwork ;

Step S2, concatenating the discriminative features of each of the specific view subnetworks into a view sequence, as the input of the public subnetwork, calculating the view independent features and view attention weights, and then calculating the behavior category of the training data The probability;

Step S3, judging whether all the training data has been input, if so, go to step S4; otherwise, go to step S1;

Step S4, calculating the loss function;

Step S5, judging whether the loss function is convergent, if so, the training ends, otherwise, go to step S6;

Step S6, adjust the parameters of the skeleton behavior recognition model, go to step S1.

Preferably, the view-specific subnetwork includes: a spatial attention module, a temporal attention module, and a discriminative feature extraction module;

The public subnet includes: a two-way long-short-term memory network, a perspective attention module, and a probability calculation module.

Preferably, in step S1, "calculate the spatial domain attention weight and the temporal domain attention weight, and then calculate the discriminative features of the specific viewing angle subnet", specifically including:

Assigning attention weights to each joint point through the spatial attention module;

Through the time domain attention module, assign time domain attention weights for each frame;

According to the training data and the spatial attention weight, the discriminative feature of the training data on the specific viewing angle is extracted by the discriminative feature extraction module;

Outputting the discriminative features of the specific view subnet according to the temporal attention weights and the discriminative features on the specific view.

Preferably, in step S2, "calculate angle-independent features and perspective attention weights, and then calculate the probability of the behavior category of the training data", specifically includes:

Outputting angle-independent features through the bidirectional long-short-term memory network;

Assigning different perspective attention weights to each of the specific perspectives through the perspective attention module;

The probability of the behavior category of the training data is obtained through the probability calculation module according to the angle-independent feature and the view weight.

Preferably, the loss function is:

in,

The first item is the cross-entropy loss of the entire network; _yi is the true label of the training data; The probability that the training data predicted for the public subnetwork belongs to the i-th behavior category; C is the number of behavior categories;

λ ₁ , λ ₂ and λ ₃ are parameters for balancing the entire network;

The second item is the regular term of the spatial attention module; K is the number of joint points; v is the number of the specific viewing angle; T is the number of frames of the input training data; is the spatial attention weight of the k-th joint point in the t-th frame under the j-th viewing angle;

The third item is the regular term of the temporal attention module; Be the time-domain attention weight of the t-th frame under the j-th viewing angle;

The fourth item is the regular term of the parameter; W _sv is the connection matrix of the network, and the L ₁ norm is used to prevent the entire network from overfitting.

Preferably, the spatial attention module consists of an LSTM layer, two fully connected layers, and a tanh activation unit;

Correspondingly, the method for calculating the airspace attention weight includes:

The input data passes through the airspace attention module to obtain the corresponding score of the K joint point in the tth frame:

Normalize the obtained scores to obtain the spatial attention weight of each joint point:

in,

W _es , W _xs , W _hs are parameter matrices that need to be learned; is the input data of frame t; Indicates the input data of frame t-1 Spatial hidden output through the LSTM layer; b _s and be _es are both bias items;

is the spatial attention weight of the k-th joint point in the t-th frame under the j-th viewing angle; is the corresponding score of the k-th joint point in the t-th frame under the j-th viewing angle; is the corresponding score of the l-th joint point in the t-th frame under the j-th viewing angle.

Preferably, the temporal attention module is composed of an LSTM layer, a fully connected layer, and a ReLU activation unit;

Correspondingly, the method for calculating the temporal domain attention weight includes:

in,

is the time-domain attention weight of the t-th frame; W _e1 and W _e2 are parameter matrices that need to be learned, Indicates the time-domain hidden output of the input data of the t-1th frame after passing through the LSTM layer; b _e is the bias item.

Preferably, the discriminative feature extraction module is composed of 3 layers of LSTM;

The input of the discriminative feature extraction module The element-wise dot product of the spatial attention weight and the input data is obtained:

in,

Respectively, the spatial domain attention weight and input data of the kth joint point in the tth frame under the jth viewing angle;

The output of the discriminative feature extraction module Perform dot multiplication with the time-domain attention weight as the jth element of the public subnetwork input data:

Preferably, the training method of the skeleton behavior recognition model also includes the step of preprocessing the training data before step S1:

Step S0, group the skeleton sequences corresponding to the same behavior of the same subject in the same environment according to different viewing angles; keep the coordinates of each joint point to four decimal places; take the first 100 frames of the skeleton sequence for each viewing angle If there are less than 100 frames, the data of the last frame will be used to make up.

In the second aspect of the present invention, a storage device is provided, in which a program is stored, and the program is suitable for being loaded and executed by a processor, so as to realize the above-mentioned angle-independence-based skeleton behavior recognition method.

In a third aspect of the present invention, a processing device is proposed, including: a processor and a storage device; wherein, the processor is adapted to execute a program; the storage device is adapted to store the program; and the program is adapted to be executed by the processor Load and execute to realize the angle-independent-based skeleton behavior recognition method described above.

In the fourth aspect of the present invention, an angle-independent skeleton behavior recognition system is proposed, including: a control unit, and a skeleton behavior recognition model;

The control unit is used to train the skeleton behavior recognition model, and use the trained skeleton behavior recognition model to calculate the behavior category probability of the skeleton time series to be recognized;

The skeleton behavior recognition model includes: a preset number of specific perspective subnets and public subnets;

in,

The specific view subnet includes: a spatial attention module, a temporal attention module, and a discriminative feature extraction module;

The public subnet includes: a two-way long-short-term memory network, a perspective attention module, and a probability calculation module.

Beneficial effects of the present invention:

The invention designs a view-specific subnetwork based on the skeleton sequence of each view, focuses on key joints and key frames through the spatial attention and temporal attention modules, and learns the discriminability of each view sequence through a multi-layer long-short-term memory network Features; the output features of each specific view subnetwork are concatenated as the input of the common subnetwork, and the view-independent features are further learned through the bidirectional long short-term memory network, and the key view points are focused on through the view attention module; a regularized cross-entropy loss function is proposed Promote the joint learning of multiple modules of the network. The invention integrates specific view subnets and public subnets, fully mines all information of a given multi-view sequence, and adds spatiotemporal attention and view attention modules at the same time, effectively improving the accuracy of behavior recognition, and can automatically focus on learning information Multi-view features fully exploit all the information of a given sequence.

Description of drawings

Fig. 1 is the schematic flow chart of the embodiment of the skeleton behavior recognition method based on angle independence of the present invention;

Fig. 2 is the schematic flow chart of the training method embodiment of skeleton behavior recognition model of the present invention;

Fig. 3 is a schematic diagram of the signal flow of the skeleton behavior recognition model embodiment of the present invention;

Fig. 4 is a schematic diagram of the composition of a specific view subnet in an embodiment of the skeleton behavior recognition model of the present invention;

Fig. 5 is a schematic diagram of the composition of the public subnet in the skeleton behavior recognition model embodiment of the present invention;

FIG. 6 is a schematic diagram of an embodiment of the angle-independent-based skeleton behavior recognition system of the present invention.

Detailed ways

Preferred embodiments of the present invention are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present invention, and are not intended to limit the protection scope of the present invention.

In order to solve the problem that the existing skeleton behavior recognition technology does not fully mine all the information of the given sequence and the recognition accuracy needs to be improved. The present invention proposes an angle-independent skeletal behavior recognition method based on the deep network of spatio-temporal perspective attention, which integrates specific perspective subnetworks and public subnetworks to fully mine all the information of a given multi-viewpoint sequence, and simultaneously adds spatiotemporal attention and perspective attention , to improve the accuracy of behavior recognition. The design idea of the method is as follows: (1) Design a view-specific subnetwork based on the skeleton sequence of each view, focus on key joints and key frames through the spatial attention and temporal attention modules, and learn each view through the discriminative feature extraction module. Discriminative features of each view sequence; (2) The output features of each specific view subnetwork are used as the input of the public subnetwork, and the view-independent features are further learned through the bidirectional long-short-term memory network Bi-LSTM, and the view-attention module focuses on Key perspective; (3) A regularized cross-entropy loss function is proposed to promote the joint learning of multiple network modules.

FIG. 1 is a schematic flow chart of an embodiment of the method for recognizing skeleton behavior based on angle independence in the present invention. As shown in Figure 1, the identification method of this embodiment includes the following steps:

Step A1, input the skeleton time series to be recognized into the trained skeleton behavior recognition model according to different perspectives;

Step A2, using the trained skeleton behavior recognition model to calculate the behavior category probability of the skeleton time series to be recognized.

Among them, the skeleton behavior recognition model includes: a preset number of view-specific subnets and public subnets.

Fig. 2 is a schematic flowchart of an embodiment of the training method of the skeleton behavior recognition model of the present invention. Fig. 3 is a schematic diagram of the signal flow of the embodiment of the skeleton behavior recognition model of the present invention. As shown in Figure 2, the training method of the present embodiment includes steps S0-S6:

In step S0, the training data is preprocessed: the skeleton sequences corresponding to the same behavior of the same subject in the same environment are grouped according to different viewing angles; the coordinates of each joint point are kept to four decimal places; each Take the first 100 frames of the skeleton sequence of each perspective (if the sample is less than 100 frames, take the last frame data to complete; if the sample is more than 100 frames, take the first 100 frames).

As shown in Figure 3, the training data set is divided into v groups, which are sent to v view subnetworks respectively. Each view-specific subnet includes: spatial attention module, temporal attention module, and discriminative feature extraction module; the outputs of v view subnetworks are serially input into the public subnetwork, and the public subnetwork includes: bidirectional long-short-term memory network, perspective Attention module, probability calculation module.

In the j-th specific viewing angle subnetwork, the input data is the K joint point information in the t-th frame under the j-th viewing angle for a given action, as shown in formula (1):

Among them, K represents the number of joint nodes in a frame

In step S1, for each specific view subnetwork, input a frame of training data corresponding to the specific view, calculate the spatial domain attention weight and temporal domain attention weight respectively, and then calculate the discriminative features of the specific view subnetwork . This step is specifically divided into steps S11-S14:

In step S11, assign attention weights to each joint point through the Spatial Attention Module (SAM).

Fig. 4 is a schematic diagram of the composition of a specific view subnetwork in the embodiment of the skeleton behavior recognition model of the present invention. As shown in Figure 4: the spatial attention module consists of an LSTM layer, two fully connected layers (FC), and a tanh activation unit. Input the previous frame into data Hidden output from LSTM layer As the historical information of the current frame, the historical information and the input data of the current frame Through the fully connected layer and the nonlinear operation of the activation layer, the corresponding scores of the K joint points in the current t-th frame are obtained as shown in the formula (2):

The calculation method of is shown in formula (3):

The resulting score corresponds to each joint point, indicating how important each joint point is to the model. Then normalize the obtained scores to obtain the joint point selection gate of each joint point, that is, the spatial attention weight. For the kth joint point, the joint point selection gate is shown in formula (4):

in:

W _es , W _xs , W _hs are parameter matrices that need to be learned; is the input data of frame t; Indicates the input data of frame t-1 Spatial hidden output through the LSTM layer; b _s and be _es are both bias items; is the spatial attention weight of the k-th joint point in the t-th frame under the j-th view; is the corresponding score of the k-th joint point in the t-th frame under the j-th viewing angle; is the corresponding score of the l-th joint point in the t-th frame under the j-th viewing angle.

In step S12, a temporal attention weight is assigned to each frame through a temporal attention module (Temporal Attention Model, TAM).

It can also be seen from Figure 4 that the temporal attention module consists of an LSTM layer, a fully connected layer, and a ReLU activation unit. The method of calculating the temporal attention weight is shown in formula (5):

in,

is the calculated time-domain attention weight of the t-th frame; W _e1 and W _e2 are parameter matrices that need to be learned, Indicates the time-domain hidden output of the input data of the t-1th frame after passing through the LSTM layer; b _e is the bias item.

In step S13, according to the training data and the spatial attention weight, the discriminative feature extraction module extracts the discriminative features of the training data on the specific viewing angle.

It can also be seen from Figure 4 that the discriminative feature extraction module is composed of 3 layers of LSTM. The module's input As shown in formula (6):

enter It is obtained by element-wise dot multiplication of the spatial attention weight and the input data, as shown in formula (7):

are the spatial attention weight and input data of the k-th joint point in the t-th frame from the j-th viewpoint, respectively. In this step, the spatial attention weight is applied to the input data of the discriminative feature extraction module, so that the network can automatically and selectively learn key joint points.

In step S14, according to the temporal attention weight and the discriminative feature on the specific view, the discriminative features of the specific view subnetwork are output.

The output of the discriminative feature extraction module Multiplied with the time-domain attention weight, the discriminative feature of the specific view subnetwork is obtained, which is used as the jth element of the input data of the public subnetwork, as shown in formula (8):

In step S2, the discriminative features of each view-specific sub-network are concatenated into a view sequence as the input of the common sub-network, and the view-independent features and view attention weights are calculated, and then the probability of the behavior category of the training data is calculated. This step is specifically divided into steps S21-S23:

In step S21, the angle-independent feature is output through a bidirectional long-short-term memory network Bi-LSTM.

Concatenate the output of the preset number (v) specific view subnetworks into a view sequence as the input of the bidirectional long-short-term memory network in the public view subnetwork, as shown in formula (9):

z=[α ¹ ,α ² ,...,α ^v ] (9)

Fig. 5 is a schematic diagram of the composition of the public subnet in the embodiment of the skeleton behavior recognition model of the present invention. As shown in Figure 5: Bi-LSTM learns the potential shared features of the same behavior in multiple perspectives, that is, angle-independent features, that is, according to the context information in the jth perspective, the forward and reverse hidden states are calculated and As shown in formulas (10) and (11):

Then the hidden states in both directions and Concatenation constitutes a hidden state h _j . Among them, W _j is the weight parameter that needs to be learned in the bidirectional LSTM.

In step S22, the output sequence of each view of Bi-LSTM is used as the input of VAM, as shown by the dotted arrow in Figure 5, and v hidden states are formed into a hidden state set, as shown in formula (12):

H=(h ₁ ,h ₂ ,...,h _v ) (12)

Then assign different perspective attention weights to each specific perspective through the View Attention Module (VAM). It can be seen from Figure 5 that the perspective attention module in this embodiment includes two fully connected layers FC and a Tanh activation layer , the calculated perspective attention weight is shown in formula (13), which assigns a weight value to each perspective:

β=(β ¹ ,β ² ,...,β ^V ) (13)

In step S23, according to the angle-independent feature and the angle-of-view attention weight, the probability calculation module is used to obtain the probability of the behavior category of the training data. It can be seen from FIG. 5 that the probability calculation module in this embodiment includes a fully connected layer and a Softmax layer.

Step S3, judging whether all T frames of training data have been input, if so, go to step S4; otherwise, go to step S1;

In step S4, the regularized cross-entropy loss function is calculated, as shown in formula (14):

in,

The first item is the cross-entropy loss of the entire network; y _i is the real label of the training data; The probability that the training data predicted for the public subnetwork belongs to the i-th behavior category; C is the number of behavior categories.

λ ₁ , λ ₂ and λ ₃ are parameters for balancing the entire network.

The second item is the regular term of the spatial attention module, so that the skeleton behavior recognition model can dynamically focus on the key joint points in each frame in the sequence corresponding to each view; K is the number of joint points; v is the number of specific view points number; T is the number of frames of the input training data; is the spatial attention weight of the k-th joint point in the t-th frame under the j-th viewpoint.

The third term is the regular term of the temporal attention module, which enables the skeleton behavior recognition model to dynamically focus on key frames; is the temporal attention weight of the t-th frame under the j-th viewing angle.

The fourth item is the regular term of the parameter; W _sv is the connection matrix of the network, and the L ₁ norm is used to prevent the entire network from overfitting.

In step S5, it is judged whether the loss function is convergent, if yes, the training ends, otherwise, go to step S6.

In step S6, adjust the parameters of the skeleton behavior recognition model, and go to step S1.

An embodiment of a storage device of the present invention stores a program therein, and the program is suitable for being loaded and executed by a processor to realize the angle-independent-based skeleton behavior recognition method described above.

A processing device of the present invention includes: a processor and a memory. Wherein, the processor is suitable for executing the program; the storage device is suitable for storing the program; the program is suitable for being loaded and executed by the processor to realize the above-mentioned angle-independent-based skeleton behavior recognition method.

In this embodiment, the specific operating hardware and programming language are: the experiment is based on the Ubuntu 14.04LTS system, and the server configuration used is Xeon E5-2630V4 2.2GHZ processor, 128G memory and four NVIDIA Tian-X GPUs with 12G video memory. The experiment uses the Keras deep learning framework, the TensorFlow backend, the integrated development environment is Pycharm, and uses the stochastic gradient descent (SGD) algorithm to train our network.

In this embodiment, the NTU RGB+D data set, the largest multi-view skeleton public data set at present, is selected as the training data and the test data. The dataset contains 56,880 video samples, 60 behavior classes, and 40 subjects. Each frame of human body data is represented by the coordinates of 25 joint points. Using the standard cross-subject test method, the videos made by 20 actors in the data set are used as the training set, and the rest are used as the test set.

FIG. 6 is a schematic diagram of an embodiment of the angle-independent-based skeleton behavior recognition system of the present invention. As shown in Figure 6, the skeleton behavior recognition system of this embodiment includes: a control unit 100, and a skeleton behavior recognition model 200;

The control unit 100 is used to train the skeleton behavior recognition model 200, and use the trained skeleton behavior recognition model 200 to calculate the behavior category probability of the skeleton time series to be recognized;

The skeleton behavior recognition model 200 includes: a preset number of specific perspective subnetworks 210, and a public subnetwork 220;

in,

Specific view subnetwork 210 includes: spatial domain attention module 211, temporal domain attention module 212, and discriminative feature extraction module 213; public subnetwork 220 includes: bidirectional long-short-term memory network 221, perspective attention module 222, probability calculation module 223 .

For the functional configuration of the control unit 100, refer to the narration of steps A1-A2 and steps S1-S6; spatial domain attention module 211, time domain attention module 212, discriminative feature extraction module 213, bidirectional long-short-term memory network 221, perspective attention module 222. For the structure and functions of the probability calculation module 223, please also refer to the previous relevant descriptions, which will not be repeated here.

Those skilled in the art should be able to realize that the method steps, modules, and units of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the electronic hardware and software interchangeability, the composition and steps of each example have been generally described in terms of functions in the above description. Whether these functions are performed by electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may implement the described functionality using different methods for each particular application, but such implementation should not be considered as exceeding the scope of the present invention.

So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings, but those skilled in the art will easily understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to relevant technical features, and the technical solutions after these changes or substitutions will all fall within the protection scope of the present invention.

Claims (12)

1. A skeleton behavior recognition method based on angle independence, is characterized in that, comprises: Input the skeleton time series to be recognized into the trained skeleton behavior recognition model according to different perspectives; Using the trained skeleton behavior recognition model to calculate the behavior category probability of the skeleton time series to be recognized; in, The skeleton behavior recognition model includes: a preset number of specific perspective subnets and public subnets; The training method of described skeleton behavior recognition model, comprises the following steps: Step S1, for each of the specific view subnets, input a frame of training data corresponding to the specific view, respectively calculate the spatial domain attention weight, the temporal domain attention weight, and then calculate the discriminative features of the specific view subnetwork ; Step S2, concatenating the discriminative features of each of the specific view subnetworks into a view sequence, as the input of the public subnetwork, calculating the view independent features and view attention weights, and then calculating the behavior category of the training data The probability; Step S3, judging whether all the training data has been input, if so, go to step S4; otherwise, go to step S1; Step S4, calculating the loss function; Step S5, judging whether the loss function is convergent, if so, the training ends, otherwise, go to step S6; Step S6, adjust the parameters of the skeleton behavior recognition model, go to step S1. 2. skeleton behavior recognition method according to claim 1, is characterized in that, The specific view subnet includes: a spatial attention module, a temporal attention module, and a discriminative feature extraction module; The public subnet includes: a two-way long-short-term memory network, a perspective attention module, and a probability calculation module. 3. The skeleton behavior recognition method according to claim 2, characterized in that, in step S1, "calculate the spatial attention weight and the temporal attention weight, and then calculate the discriminative features of the specific viewing angle subnet", specifically including : Assigning attention weights to each joint point through the spatial attention module; Through the time domain attention module, assign time domain attention weights for each frame; According to the training data and the spatial attention weight, the discriminative feature of the training data on the specific viewing angle is extracted by the discriminative feature extraction module; Outputting the discriminative features of the specific view subnet according to the temporal attention weights and the discriminative features on the specific view. 4. The skeleton behavior recognition method according to claim 2, wherein, in step S2, "calculate the angle-independent feature and the attention weight of the angle of view, and then calculate the probability of the behavior category of the training data", specifically comprising: Outputting angle-independent features through the bidirectional long-short-term memory network; Assigning different perspective attention weights to each of the specific perspectives through the perspective attention module; The probability of the behavior category of the training data is obtained through the probability calculation module according to the angle-independent feature and the view weight. 5. skeleton behavior recognition method according to claim 1, is characterized in that, described loss function is: in, The first item is the cross-entropy loss of the entire network; _yi is the true label of the training data; The probability that the training data predicted for the public subnetwork belongs to the i-th behavior category; C is the number of behavior categories; λ ₁ , λ ₂ and λ ₃ are parameters for balancing the entire network; The second item is the regular term of the spatial attention module; K is the number of joint points; v is the number of the specific viewing angle; T is the number of frames of the input training data; is the spatial attention weight of the k-th joint point in the t-th frame under the j-th viewing angle; The third item is the regular term of the temporal attention module; Be the time-domain attention weight of the t-th frame under the j-th viewing angle; The fourth item is the regular term of the parameter; W _sv is the connection matrix of the network, and the L ₁ norm is used to prevent the entire network from overfitting. 6. The skeleton behavior recognition method according to claim 3, wherein the spatial domain attention module is made up of an LSTM layer, two fully connected layers, and a tanh activation unit; Correspondingly, the method for calculating the airspace attention weight includes: The input data passes through the airspace attention module to obtain the corresponding score of the K joint point in the tth frame: Normalize the obtained scores to obtain the spatial attention weight of each joint point: in, W _es , W _xs , W _hs are parameter matrices that need to be learned; is the input data of frame t; Indicates the input data of frame t-1 Spatial hidden output through the LSTM layer; b _s and be _es are both bias items; is the spatial attention weight of the k-th joint point in the t-th frame under the j-th viewing angle; is the corresponding score of the k-th joint point in the t-th frame under the j-th viewing angle; is the corresponding score of the l-th joint point in the t-th frame under the j-th viewing angle. 7. the skeleton behavior recognition method according to claim 6, is characterized in that, described temporal domain attention module, is made up of LSTM layer, a fully connected layer, and a ReLU activation unit; Correspondingly, the method for calculating the temporal domain attention weight includes: in, is the time-domain attention weight of the t-th frame; W _e1 and W _e2 are parameter matrices that need to be learned, Indicates the time-domain hidden output of the input data of the t-1th frame after passing through the LSTM layer; b _e is the bias item. 8. skeleton behavior recognition method according to claim 7, is characterized in that, described discriminant feature extraction module is made of 3 layers of LSTM; The input of the discriminative feature extraction module The element-wise dot product of the spatial attention weight and the input data is obtained: in, Respectively, the spatial domain attention weight and input data of the kth joint point in the tth frame under the jth viewing angle; The output of the discriminative feature extraction module Perform dot multiplication with the time-domain attention weight as the jth element of the public subnetwork input data: 9. The skeleton behavior recognition method according to claim 1, characterized in that, the training method of the skeleton behavior recognition model also includes the step of preprocessing the training data before step S1: Step S0, group the skeleton sequences corresponding to the same behavior of the same subject in the same environment according to different viewing angles; keep the coordinates of each joint point to four decimal places; take the first 100 frames of the skeleton sequence for each viewing angle If there are less than 100 frames, the data of the last frame will be used to make up. 10. A storage device, wherein a program is stored, wherein the program is adapted to be loaded and executed by a processor, so as to realize the angle-independent skeleton behavior recognition based on any one of claims 1-9 method. 11. A processing device comprising: a processor adapted to execute programs; and a storage device adapted to store the program; It is characterized in that the program is adapted to be loaded and executed by a processor to realize the angle-independent-based skeleton behavior recognition method according to any one of claims 1-9. 12. A skeleton behavior recognition system based on angle independence, comprising: a control unit, and a skeleton behavior recognition model; The control unit is configured to train the skeleton behavior recognition model, and use the trained skeleton behavior recognition model to calculate the behavior category probability of the skeleton time series to be recognized; The skeleton behavior recognition model includes: a preset number of specific perspective subnets and public subnets; in, The specific view subnet includes: a spatial attention module, a temporal attention module, and a discriminative feature extraction module; The public subnet includes: a two-way long-short-term memory network, a perspective attention module, and a probability calculation module.