A highly compact method for long-term human motion prediction based on a full MLP architecture

Abstract

The present invention discloses a highly compact long-term human motion prediction method based on a full MLP architecture, comprising: obtaining raw time-series human skeleton data and performing preprocessing, wherein the preprocessing involves transforming the raw time-series human skeleton data from the time domain to the frequency domain and then mapping it to a high-dimensional space to form high-dimensional spatial features; establishing a full MLP architecture model, comprising N sequentially connected dynamic spatiotemporal feature perception modules; inputting the high-dimensional spatial features into the full MLP architecture model to obtain a first extracted feature; restoring the first extracted feature from the high-dimensional space to a three-dimensional space through layer normalization and a first linear layer; transforming the restored first extracted feature from the frequency domain to the time domain to form a second extracted feature; and inputting the second extracted feature into a multi-period joint prediction module to obtain a final prediction result. The lightweight design significantly reduces the number of model parameters, thereby reducing computational complexity and resource consumption, improving prediction accuracy and reliability, and achieving good versatility.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and machine learning technology, and specifically relates to a highly compact long-term human motion prediction method based on a full MLP architecture. Background Art

[0002] Human motion prediction technology has broad application prospects in the field of artificial intelligence, including but not limited to autonomous driving, human-computer interaction, motion monitoring, and public safety. Accurate human motion prediction can significantly improve the performance and user experience of these applications.

[0003] Existing technologies generally use the following technical solutions for human motion prediction, including: 1) methods that mainly rely on hidden Markov models and linear dynamic systems. These methods have shown certain potential in capturing simple reciprocating motions such as walking. However, they face the challenge of insufficient accuracy when processing complex movements; 2) Recurrent Neural Networks (RNNs): RNN-based methods have attracted attention due to their advantages in processing sequence data. These methods embed human posture into a latent space through an encoding-recursive decoding framework to predict future motion. However, RNNs have limitations in spatial perception, resulting in significant cumulative errors; 3) Graph Convolutional Networks (GCNs): Utilizing the trajectory, time and spatial information of human posture, high-precision prediction is achieved by constructing a spatiotemporal architecture. However, GCN-based methods usually rely on manually designed skeleton relationship graphs, which not only increases the complexity of the network, but also limits the versatility and portability of the model; 4) Lightweight networks: There are often some problems in prediction accuracy and prediction range, especially in application scenarios that require fast response and high prediction accuracy.

[0004] Therefore, the existing technology has at least the following limitations: it is insufficient in capturing the strong spatial and temporal correlation of human motion, it is difficult to run on resource-constrained mobile devices due to the large amount of computing resources required, the specifically designed skeleton relationship graph limits the versatility and adaptability of the model, and the accuracy and reliability of long-term prediction remains a challenge, especially in rapidly changing environments. Summary of the Invention

[0005] The purpose of the present invention is to address the above problems and propose a highly compact long-term human motion prediction method based on a full MLP architecture, which achieves a lightweight design, greatly reduces the number of model parameters, thereby reducing computational complexity and resource consumption, and improving the accuracy and reliability of the prediction, with good versatility.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] The present invention proposes a highly compact long-term human motion prediction method based on a full MLP architecture, comprising the following steps:

[0008] S1. Obtaining original time series human skeleton data and performing preprocessing. The preprocessing is to transform the original time series human skeleton data from the time domain to the frequency domain and then map it to a high-dimensional space to form high-dimensional space features;

[0009] S2. Establish a full MLP architecture model. The full MLP architecture model includes N sequentially connected dynamic spatiotemporal feature perception modules. The dynamic spatiotemporal feature perception modules perform the following operations:

[0010] S21. Use a spatial multilayer perceptron and a temporal multilayer perceptron to extract features from the input features, respectively, to obtain spatial features and temporal features. The spatial multilayer perceptron is a multilayer perceptron that acts on the spatial dimension, and the temporal multilayer perceptron is a multilayer perceptron that acts on the temporal dimension.

[0011] S22, respectively adding the spatial features and the temporal features to the input features to obtain spatial fusion features and temporal fusion features;

[0012] S23, using dynamic aggregation mechanism to aggregate spatial fusion features and temporal fusion features into spatiotemporal features;

[0013] S24, using a channel multilayer perceptron to enhance the spatiotemporal features to obtain channel enhanced features, where the channel multilayer perceptron is a multilayer perceptron that acts on the channel dimension;

[0014] S25, adding the channel enhancement feature and the spatiotemporal feature to obtain an output feature;

[0015] S3. Input the high-dimensional spatial features into the full MLP architecture model to obtain the first extracted features, which are the output features of the last dynamic spatiotemporal feature perception module;

[0016] S4, restoring the first extracted features from the high-dimensional space to the three-dimensional space through layer normalization and the first linear layer in sequence;

[0017] S5, transforming the restored first extracted features from the frequency domain to the time domain to form second extracted features;

[0018] S6. Input the second extracted features into a multi-period joint prediction module to obtain a final prediction result. The multi-period joint prediction module includes multiple parallel second linear layers with different time periods and performs the following operations:

[0019] After adding the second linear layers of different time periods to the last frame of human skeleton data in the original time series human skeleton data, they are spliced ​​in sequence in the time direction to form the final prediction result.

[0020] Preferably, the high-dimensional space feature is expressed by the following formula:

[0021]

[0022] Where, Represents the original time series human skeleton data, T o Represents the number of frames in the original temporal human skeleton data, Represents the three-dimensional coordinate data of each node of the human skeleton in the mth frame, m∈[-T o ,-1], K represents the number of nodes of each frame of human skeleton data, DCT represents discrete cosine transform, W 0 represents the weight of the third linear layer, b 0 represents the bias of the third linear layer, Y 0 Represents high-dimensional space features.

[0023] Preferably, there are eight dynamic spatiotemporal feature perception modules.

[0024] Preferably, the multilayer perceptron includes layer normalization, GELU activation function, fourth linear layer and fifth linear layer connected in sequence, which is expressed by the following formula:

[0025] MLP(Y in )=W 2 (GELU(W 1 LN(Y in )+b 1 ))+b 2 ,

[0026] Where MLP(·) represents a multi-layer perceptron, Y in Represents the input features of the multilayer perceptron, W 1 represents the weight of the fourth linear layer, b 1 represents the bias of the fourth linear layer, W 2 represents the weight of the fifth linear layer, b 2 represents the bias of the fifth linear layer, LN represents layer normalization, and GELU represents the GELU activation function.

[0027] Preferably, the dynamic aggregation mechanism is represented by the following formula:

[0028]

[0029] Where, represents the spatial fusion feature of the i-th dynamic spatiotemporal feature perception module, represents the temporal fusion feature of the i-th dynamic spatiotemporal feature perception module, i = 1 to N, Concat represents the splicing of the channel dimension, represents the weight of the sixth linear layer in the i-th dynamic spatiotemporal feature perception module, represents the bias of the sixth linear layer in the i-th dynamic spatiotemporal feature perception module, σ represents the sigmoid activation function, represents the spatiotemporal feature dynamic coefficient learned by the i-th dynamic spatiotemporal feature perception module, represents the spatial dynamic coefficient learned by the i-th dynamic spatiotemporal feature perception module, represents the temporal dynamic coefficient learned by the i-th dynamic spatiotemporal feature perception module, Represents the output feature of the dynamic aggregation mechanism in the i-th dynamic spatiotemporal feature perception module.

[0030] Preferably, the second extracted feature is expressed by the following formula:

[0031] Y p =DCT -1 (W p (LN(Y N ))+b p ),

[0032] Where Y N represents the output feature of the Nth dynamic spatiotemporal feature perception module, i.e., the first extracted feature, LN represents layer normalization, and W p represents the weight of the seventh linear layer, b p Represents the bias of the seventh linear layer, DCT -1 represents the inverse discrete cosine transform, Y p represents the second extracted feature.

[0033] Preferably, the multi-period joint prediction module includes four parallel second linear layers corresponding to four different time periods: short time, long time, super long time, and extremely long time.

[0034] Preferably, the final prediction result is expressed by the following formula:

[0035]

[0036] Where, Represents the final prediction result, that is, the predicted sequence human skeleton data, Represents the three-dimensional coordinate data of each node of the human skeleton in the nth frame, n∈[0,T p -1], T p Indicates the number of predicted frames, X st represents the prediction sequence of the second linear layer in the short time period corresponding to the 0th frame to the n1-1th frame, X lt represents the predicted sequence of the second linear layer in the long time period corresponding to the 0th frame to the n2-1th frame, X eltrepresents the prediction sequence of the second linear layer in the ultra-long time period corresponding to the 0th frame to the n3-1th frame, X eelt The second linear layer corresponds to the 0th frame to the Tth frame. p -1 frame prediction sequence, represents the weight of the second linear layer in the short time period, represents the bias of the second linear layer in the short time period, represents the weight of the second linear layer in the long term, represents the bias of the second linear layer in the long time period, represents the weight of the second linear layer in the ultra-long time period, represents the bias of the second linear layer in the ultra-long time period, represents the weight of the second linear layer at very long time intervals, represents the bias of the second linear layer for a very long time period, T o represents the number of frames in the original time series human skeleton data, K represents the number of nodes in each frame of human skeleton data, and Y p Represents the second extracted feature, R st (x -1 ) represents the short-term residual sequence, R lt (x -1 ) represents the long-term residual sequence, R elt (x -1 ) represents the residual sequence of the ultra-long time period, R eelt (x -1 ) represents the extremely long time period residual sequence, the residual sequence is the last frame of human skeleton data in the original time series human skeleton data with the corresponding time period length, Concat T represents the splicing of the time dimension, Represents the prediction sequence from frame 0 to frame n1-1 in a short time period, Represents the predicted sequence from the n1th frame to the n2-1th frame in the long time period, Represents the predicted sequence from the n2th frame to the n3-1th frame in the ultra-long time period, Indicates the extremely long time period from frame n3 to frame T p -1 frame prediction sequence, n1 represents the first frame of a long time period, n2 represents the first frame of an ultra-long time period, and n3 represents the first frame of an extremely long time period. The range of n1, n2, and n3 is (0, T p -1).

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] This application proposes a highly compact long-term human motion prediction method based on a full MLP architecture. Based on the full MLP architecture model, it predicts the continuous motion of the human body over a period of time by analyzing the original sequence of human skeleton data. This method not only achieves high-precision and high-efficiency prediction in a lightweight manner, but also achieves a long-term prediction range while maintaining simplicity and understandability. It is suitable for application scenarios that require rapid response and extended prediction range, such as autonomous driving, human-computer interaction, motion monitoring, and public safety. Specifically, the dynamic spatiotemporal feature perception module (STDP) improves computational efficiency and maintains interpretability. Different but parallel spatial multi-layer perceptrons (SMLPs) and temporal multi-layer perceptrons (TMLPs) simultaneously capture the spatial correlation between human skeleton pose joints and the temporal dynamics of each joint. Subsequently, the dynamic aggregation mechanism (DA) combined with the channel multi-layer perceptron (CMLP) dynamically integrates and refines spatial and temporal features, thereby improving prediction accuracy. Secondly, the multi-period joint prediction module (MTUP) enables accurate prediction of a preset time period (e.g., 0-4000ms), avoiding repeated short-term prediction iterations and achieving accurate prediction of complex human motion. This method has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Flowchart of the highly compact long-term human motion prediction method based on the full MLP architecture of the present invention;

[0040] Figure 2 Schematic diagram of the overall network architecture of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0042] It should be noted that when a component is referred to as being "connected" to another component, it may be directly connected to the other component or there may be an intermediate component. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art in the art of this application. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0043] In view of the limitations of the existing technology, the present invention aims to provide an excellent human motion prediction network architecture, which can not only achieve high-precision prediction of complex human motion in a lightweight manner, but also achieve a longer prediction range while maintaining simplicity and understandability, and can adapt to different application scenarios, especially application scenarios such as autonomous driving, human-computer interaction, motion monitoring, and public safety that have extremely high requirements for real-time and accuracy.

[0044] like Figure 1-2 As shown in FIG, a highly compact long-term human motion prediction method based on a full MLP architecture includes the following steps:

[0045] S1. Obtain the original time series human skeleton data and perform preprocessing. The preprocessing is to transform the original time series human skeleton data from the time domain to the frequency domain and then map it to a high-dimensional space to form high-dimensional space features.

[0046] In one embodiment, the high-dimensional spatial feature is expressed by the following formula:

[0047]

[0048] Where, Represents the original time series human skeleton data, T o Represents the number of frames in the original temporal human skeleton data, Represents the three-dimensional coordinate data of each node of the human skeleton in the mth frame, m∈[-T o ,-1], K represents the number of nodes of each frame of human skeleton data, DCT represents discrete cosine transform, W 0 represents the weight of the third linear layer, b 0 represents the bias of the third linear layer, Y 0 Represents high-dimensional space features.

[0049] Specifically, such as Figure 2 As shown, for the input original temporal human skeleton data Use discrete cosine transform (DCT) and a linear layer to convert it into a high-dimensional spatial feature.

[0050] S2. Establish a full MLP architecture model. The full MLP architecture model includes N sequentially connected dynamic spatiotemporal feature perception modules. The dynamic spatiotemporal feature perception modules perform the following operations:

[0051] S21. Use a spatial multilayer perceptron and a temporal multilayer perceptron to extract features from the input features, respectively, to obtain spatial features and temporal features. The spatial multilayer perceptron is a multilayer perceptron that acts on the spatial dimension, and the temporal multilayer perceptron is a multilayer perceptron that acts on the temporal dimension.

[0052] S22, respectively adding the spatial features and the temporal features to the input features to obtain spatial fusion features and temporal fusion features;

[0053] S23, using dynamic aggregation mechanism to aggregate spatial fusion features and temporal fusion features into spatiotemporal features;

[0054] S24, using a channel multilayer perceptron to enhance the spatiotemporal features to obtain channel enhanced features, where the channel multilayer perceptron is a multilayer perceptron that acts on the channel dimension;

[0055] S25. Add the channel enhancement feature and the spatiotemporal feature to obtain the output feature.

[0056] In one embodiment, there are eight dynamic spatiotemporal feature perception modules.

[0057] In one embodiment, a multilayer perceptron includes layer normalization, a GELU activation function, a fourth linear layer, and a fifth linear layer connected in sequence, and is represented by the following formula:

[0058] MLP(Y in )=W 2 (GELU(W 1 LN(Y in )+b 1 ))+b 2 ,

[0059] Where MLP(·) represents a multi-layer perceptron, Y in Represents the input features of the multilayer perceptron, W 1 represents the weight of the fourth linear layer, b 1 represents the bias of the fourth linear layer, W 2 represents the weight of the fifth linear layer, b 2 represents the bias of the fifth linear layer, LN represents layer normalization, and GELU represents the GELU activation function.

[0060] In one embodiment, the dynamic aggregation mechanism is represented by the following formula:

[0061]

[0062] Where, represents the spatial fusion feature of the i-th dynamic spatiotemporal feature perception module, represents the temporal fusion feature of the i-th dynamic spatiotemporal feature perception module, i = 1 to N, Concat represents the splicing of the channel dimension, represents the weight of the sixth linear layer in the i-th dynamic spatiotemporal feature perception module, represents the bias of the sixth linear layer in the i-th dynamic spatiotemporal feature perception module, σ represents the sigmoid activation function, represents the spatiotemporal feature dynamic coefficient learned by the i-th dynamic spatiotemporal feature perception module, represents the spatial dynamic coefficient learned by the i-th dynamic spatiotemporal feature perception module, represents the temporal dynamic coefficient learned by the i-th dynamic spatiotemporal feature perception module, Represents the output feature of the dynamic aggregation mechanism in the i-th dynamic spatiotemporal feature perception module.

[0063] Specifically, if Figure 2 As shown, the full MLP architecture model of this embodiment includes N=8 dynamic spatiotemporal feature perception modules (STDP) connected in sequence. The dynamic spatiotemporal feature perception module extracts the spatial features and temporal features of the human skeleton joints by a spatial multilayer perceptron (SMLP) and a temporal multilayer perceptron (TMLP), respectively, and adds the spatial features and temporal features to the input features to obtain spatial fusion features and temporal fusion features respectively. The dynamic aggregation mechanism (DA) is used to aggregate the spatial fusion features and the temporal fusion features into spatiotemporal features, which are further enhanced using a channel multilayer perceptron (CMLP).

[0064] The SMLP, TMLP, and CMLP involved in the dynamic spatiotemporal feature perception module are all implemented by the basic MLP acting on the spatial, temporal, and channel dimensions respectively. During this process, non-current working dimensions must be hidden. SMLP and TMLP act on the spatial and temporal dimensions of the input features, respectively, to obtain spatial features and temporal features. The spatial features and temporal features are added to the input features to obtain spatial fusion features and temporal fusion features, respectively. They are expressed using the following formula:

[0065]

[0066] in, represents the spatial multilayer perceptron (SMLP), represents the temporal multilayer perceptron (TMLP), Y i-1 is the output feature of the i-1th dynamic spatiotemporal feature perception module (STDP) and serves as the input feature of the current i-th dynamic spatiotemporal feature perception module (STDP). The input of the first dynamic spatiotemporal feature perception module (STDP) is the high-dimensional spatial feature. represents the spatial fusion feature of the i-th dynamic spatiotemporal feature perception module, Represents the temporal fusion features of the i-th dynamic spatiotemporal feature perception module, i = 1~N.

[0067] The dynamic aggregation mechanism (DA) is used to dynamically fuse the spatial fusion features and temporal fusion features of the current dynamic spatiotemporal feature perception module to balance spatial information and temporal information.

[0068] The channel multilayer perceptron (CMLP) acts on the channel dimension of the feature to further aggregate and enhance the learned spatiotemporal features, which can be expressed as follows:

[0069]

[0070] Where, represents the channel multilayer perceptron (CMLP), Y i Represents the output features of the dynamic spatiotemporal feature perception module (STDP).

[0071] In summary, the output features of the dynamic spatiotemporal feature perception module (STDP) can reflect the spatiotemporal correlation of each learned node.

[0072] S3. Input the high-dimensional spatial features into the full MLP architecture model to obtain the first extracted features. The first extracted features are the output features of the last dynamic spatiotemporal feature perception module.

[0073] S4. The first extracted features are restored from the high-dimensional space to the three-dimensional space through layer normalization and the first linear layer in sequence.

[0074] S5. Transform the restored first extracted features from the frequency domain to the time domain to form second extracted features.

[0075] In one embodiment, the second extracted feature is expressed by the following formula:

[0076] Y p =DCT -1 (W p (LN(Y N ))+b p ),

[0077] Where Y N represents the output feature of the Nth dynamic spatiotemporal feature perception module, i.e., the first extracted feature, LN represents layer normalization, and W p represents the weight of the seventh linear layer, b p Represents the bias of the seventh linear layer, DCT -1 represents the inverse discrete cosine transform, Y p represents the second extracted feature.

[0078] Among them, such as Figure 2 As shown, the output feature of the full MLP architecture model, that is, the output feature Y of the last dynamic spatiotemporal feature perception module N , use layer normalization and linear layer to restore from high-dimensional space to three-dimensional space, and use inverse discrete cosine transform (IDCT) to transform from frequency domain to time domain.

[0079] S6. Input the second extracted features into a multi-period joint prediction module to obtain a final prediction result. The multi-period joint prediction module includes multiple parallel second linear layers with different time periods and performs the following operations:

[0080] After adding the second linear layers of different time periods to the last frame of human skeleton data in the original time series human skeleton data, they are spliced ​​in sequence in the time direction to form the final prediction result.

[0081] In one embodiment, the multi-period joint prediction module includes four parallel second linear layers, one for each of the four different time periods: short, long, very long, and extremely long. If the short, long, very long, and extremely long time periods are divided sequentially, the divisions follow a rule such that each frame in each time period is a future frame of the previous time period.

[0082] In one embodiment, the final prediction result is expressed by the following formula:

[0083]

[0084] Where, Represents the final prediction result, that is, the predicted sequence human skeleton data, Represents the three-dimensional coordinate data of each node of the human skeleton in the nth frame, n∈[0,T p -1], T p Indicates the number of predicted frames, X st represents the prediction sequence of the second linear layer in the short time period corresponding to the 0th frame to the n1-1th frame, X lt represents the predicted sequence of the second linear layer in the long time period corresponding to the 0th frame to the n2-1th frame, X elt represents the prediction sequence of the second linear layer in the ultra-long time period corresponding to the 0th frame to the n3-1th frame, X eelt The second linear layer corresponds to the 0th frame to the Tth frame. p -1 frame prediction sequence, represents the weight of the second linear layer in the short time period, represents the bias of the second linear layer in the short time period, represents the weight of the second linear layer in the long term, represents the bias of the second linear layer in the long time period, represents the weight of the second linear layer in the ultra-long time period, represents the bias of the second linear layer in the ultra-long time period, represents the weight of the second linear layer at very long time intervals, represents the bias of the second linear layer for a very long time period, T o represents the number of frames in the original time series human skeleton data, K represents the number of nodes in each frame of human skeleton data, and Y p Represents the second extracted feature, Rst (x -1 ) represents the short-term residual sequence, R lt (x -1 ) represents the long-term residual sequence, R elt (x -1 ) represents the residual sequence of the ultra-long time period, R eelt (x -1 ) represents the extremely long time period residual sequence, the residual sequence is the last frame of human skeleton data in the original time series human skeleton data with the corresponding time period length, Concat T represents the splicing of the time dimension, Represents the prediction sequence from frame 0 to frame n1-1 in a short time period, Represents the predicted sequence from the n1th frame to the n2-1th frame in the long time period, Represents the predicted sequence from the n2th frame to the n3-1th frame in the ultra-long time period, Indicates the extremely long time period from frame n3 to frame T p -1 frame prediction sequence, n1 represents the first frame of a long time period, n2 represents the first frame of an ultra-long time period, and n3 represents the first frame of an extremely long time period. The range of n1, n2, and n3 is (0, T p -1).

[0085] Specifically, the second extracted feature Y p Input the multi-time period joint prediction module (MTUP), pass through four linear layers representing different time periods of short time, long time, ultra-long time, and extremely long time, add the residuals, and splice them in the time direction (i.e. Figure 2 The final prediction result is formed by the time series splicing in .

[0086] In this embodiment, the predicted frame number T p =100, corresponding to a predicted duration of 0-4s, which can be expressed by the following formula:

[0087]

[0088] in, are four learnable parameters representing the linear layer of different time periods: short, long, very long, and extremely long. st (x -1 ), R lt (x -1 ), R elt (x -1 ), R eelt (x -1 ) represents the residuals corresponding to different time periods of short time, long time, super long time and extremely long time. Finally, the predicted sequence human skeleton data is obtained Here, the short time period corresponds to frames 0 to 9, the long time period corresponds to frames 10 to 24, the very long time period corresponds to frames 25 to 49, and the extremely long time period corresponds to frames 50 to 99. It is easy to understand that the specific number of linear layers and the division of different time periods can be adjusted according to actual needs.

[0089] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0090] The above-described embodiments merely represent specific and detailed examples of the present application and should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A highly compact method for long-term human motion prediction based on a full MLP architecture, characterized by: The highly compact long-term human motion prediction method based on the full MLP architecture comprises the following steps: S1. Obtaining original time series human skeleton data and performing preprocessing, wherein the preprocessing is to transform the original time series human skeleton data from the time domain to the frequency domain and then map it to a high-dimensional space to form a high-dimensional space feature; S2. Establish a full MLP architecture model, wherein the full MLP architecture model includes N sequentially connected dynamic spatiotemporal feature perception modules, and the dynamic spatiotemporal feature perception modules perform the following operations: S21. Use a spatial multilayer perceptron and a temporal multilayer perceptron to extract features from the input features, respectively, to obtain spatial features and temporal features, respectively. The spatial multilayer perceptron is a multilayer perceptron that acts on the spatial dimension, and the temporal multilayer perceptron is a multilayer perceptron that acts on the temporal dimension. S22, respectively adding the spatial features and the temporal features to the input features to obtain spatial fusion features and temporal fusion features; S23. A dynamic aggregation mechanism is used to aggregate the spatial fusion features and the temporal fusion features into spatiotemporal features. The dynamic aggregation mechanism is expressed by the following formula: ， ， Where, represents the spatial fusion feature of the i-th dynamic spatiotemporal feature perception module, Represents the temporal fusion feature of the i-th dynamic spatiotemporal feature perception module, i=1~N, Concat represents the splicing of the channel dimension, represents the weight of the sixth linear layer in the i-th dynamic spatiotemporal feature perception module, represents the bias of the sixth linear layer in the i-th dynamic spatiotemporal feature perception module, represents the sigmoid activation function, represents the spatiotemporal feature dynamic coefficient learned by the i-th dynamic spatiotemporal feature perception module, represents the spatial dynamic coefficient learned by the i-th dynamic spatiotemporal feature perception module, represents the temporal dynamic coefficient learned by the i-th dynamic spatiotemporal feature perception module, represents the output feature of the dynamic aggregation mechanism in the i-th dynamic spatiotemporal feature perception module; S24. Using a channel multilayer perceptron to enhance the spatiotemporal features to obtain channel enhancement features, wherein the channel multilayer perceptron is a multilayer perceptron that acts on the channel dimension; S25, adding the channel enhancement feature and the spatiotemporal feature to obtain an output feature; S3. Input the high-dimensional spatial features into the full MLP architecture model to obtain a first extracted feature, which is the output feature of the last dynamic spatiotemporal feature perception module; S4, restoring the first extracted features from the high-dimensional space to the three-dimensional space through layer normalization and the first linear layer in sequence; S5, transforming the restored first extracted features from the frequency domain to the time domain to form second extracted features; S6. Input the second extracted features into a multi-period joint prediction module to obtain a final prediction result. The multi-period joint prediction module includes multiple parallel second linear layers with different time periods, that is, four parallel second linear layers corresponding to four different time periods: short time, long time, very long time, and extremely long time, and performs the following operations: After adding the second linear layers of different time periods to the last frame of human skeleton data in the original time series human skeleton data, they are spliced ​​in sequence in the time direction to form the final prediction result.

2. The highly compact long-term human motion prediction method based on a full MLP architecture as claimed in claim 1, characterized in that: The high-dimensional space feature is expressed by the following formula: ， Where, Represents the original time series human skeleton data, T o Represents the number of frames in the original temporal human skeleton data, Represents the three-dimensional coordinate data of each node of the human skeleton in the mth frame, m∈[ ], K represents the number of nodes of each frame of human skeleton data, DCT represents discrete cosine transform, W 0 represents the weight of the third linear layer, b 0 represents the bias of the third linear layer, Y 0 Represents high-dimensional space features.

3. The highly compact long-term human motion prediction method based on a full MLP architecture as claimed in claim 1, characterized in that: There are eight dynamic spatiotemporal feature perception modules.

4. The highly compact long-term human motion prediction method based on a full MLP architecture as claimed in claim 1, characterized in that: The multilayer perceptron includes layer normalization, GELU activation function, fourth linear layer and fifth linear layer connected in sequence, which is expressed by the following formula: ， in, represents a multilayer perceptron, Represents the input features of the multilayer perceptron, W 1 represents the weight of the fourth linear layer, b 1 represents the bias of the fourth linear layer, W 2 represents the weight of the fifth linear layer, b 2 represents the bias of the fifth linear layer, LN represents layer normalization, and GELU represents the GELU activation function.

5. The highly compact long-term human motion prediction method based on a full MLP architecture as claimed in claim 1, characterized in that: The second extracted feature is expressed by the following formula: ， Where, represents the output feature of the Nth dynamic spatiotemporal feature perception module, that is, the first extracted feature, LN represents layer normalization, represents the weight of the seventh linear layer, represents the bias of the seventh linear layer, represents the inverse discrete cosine transform, represents the second extracted feature.

6. The highly compact long-term human motion prediction method based on a full MLP architecture as claimed in claim 1, characterized in that: The final prediction result is expressed by the following formula: ， ， Where, Represents the final prediction result, that is, the predicted sequence human skeleton data, Represents the three-dimensional coordinate data of each node of the human skeleton in the nth frame, n∈[ ], Indicates the number of predicted frames, Represents the prediction sequence of the second linear layer in the short time period corresponding to the 0th frame to the n1-1th frame, Represents the predicted sequence of the second linear layer in the long time period corresponding to the 0th frame to the n2-1th frame, Represents the prediction sequence of the second linear layer of the ultra-long time period corresponding to the 0th frame to the n3-1th frame, The second linear layer corresponds to the 0th frame to the Tth frame. p -1 frame prediction sequence, represents the weight of the second linear layer in the short time period, represents the bias of the second linear layer in the short time period, represents the weight of the second linear layer in the long term, represents the bias of the second linear layer in the long time period, represents the weight of the second linear layer in the ultra-long time period, represents the bias of the second linear layer in the ultra-long time period, represents the weight of the second linear layer for very long time periods, represents the bias of the second linear layer for a very long time period, T o Represents the number of frames in the original time series human skeleton data, K represents the number of nodes in each frame of human skeleton data, represents the second extracted feature, represents the short-term residual sequence, represents the long-term residual series, represents the residual sequence of the ultra-long time period, Represents a very long time period residual sequence, which is the last frame of human skeleton data in the original time series human skeleton data with the corresponding time period length. represents the splicing of the time dimension, Represents the prediction sequence from frame 0 to frame n1-1 in a short time period, Represents the predicted sequence from the n1th frame to the n2-1th frame in the long time period, Represents the predicted sequence from the n2th frame to the n3-1th frame in the ultra-long time period, Indicates the extremely long time period from frame n3 to frame T p -1 frame prediction sequence, n1 represents the first frame of a long time period, n2 represents the first frame of an ultra-long time period, and n3 represents the first frame of an extremely long time period. The range of n1, n2, and n3 is ( ).

Images