Multi-person pose prediction method based on fuzzy control and differentiated double-flow modeling

Abstract

The application discloses a multi-person posture prediction method based on fuzzy control and differential double-flow modeling, and belongs to the technical field of computer vision and human-computer interaction. The method comprises the following steps: S1, a data preprocessing stage; S2, independent individual data exclusive processing; S3, interactive individual data signal purification processing; S4, a data modeling optimization stage; and S5, processing result scene adaptation and quantitative verification. The multi-person posture prediction method based on fuzzy control and differential double-flow modeling is adopted, interactive and independent individuals are accurately divided by double fuzzy rules, noise is isolated by exclusive Transformer encoder-decoder, independent individual trajectory deviation is avoided, the signal is purified by an interactive module to improve accuracy, a weighted loss is used to reduce data dependence and enhance stability, and after scene adaptation and quantitative verification, the method can stably cope with complex scenes and improve the average prediction accuracy, and the effect is reliable.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and human-computer interaction technology, specifically involving a multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling. Background Technology

[0002] Human motion prediction technology has gradually expanded from single-person scenarios to more challenging multi-person interactive scenarios. Unlike single-person prediction, the core of multi-person prediction lies in effectively modeling the social interactions between individuals and using these interactions to improve prediction accuracy.

[0003] However, existing technologies have the following shortcomings: independent individuals are forcibly introduced into irrelevant social contexts, causing trajectory prediction to deviate; interacting individuals are interfered with by non-interacting individuals, diluting the real interaction signal; some technologies use only a single feature to identify interaction, ignoring the consistency of movement direction; and most are "hard classifications" (non-interaction / interaction), which cannot adapt to real-world scenarios such as "weak interaction" and "transitional state"; the use of a unified modeling module for interaction and independent individuals leads to independent individuals being interfered with by false social noise, and the capture of the dependency relationship of interacting individuals is inaccurate.

[0004] Therefore, a new method is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling. This method accurately distinguishes between interaction and independent individuals through dual fuzzy rules, isolates noise and avoids trajectory deviation of independent individuals with a dedicated Transformer codec, improves accuracy by purifying signals with an interaction module, and reduces data dependence and enhances stability by weighted loss. After scene adaptation and quantization verification, it stably copes with complex scenes and improves the average prediction accuracy, with reliable results.

[0006] To achieve the above objectives, this invention provides a multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling, comprising the following steps:

[0007] S1. Obtain the historical motion sequence of multiple people, and based on the behavior pattern discrimination mechanism of fuzzy control, divide the individuals corresponding to the historical motion sequence into interactive behavior pattern individuals and corresponding interactive information and independent behavior pattern individuals through preset directional alignment rules and spatial proximity rules.

[0008] The specific formula for fuzzy AND operation is as follows:

[0009] ;

[0010] In the formula, Soft membership degree;

[0011] like If they are determined to be an interactive pair, Determined to be an independent entity;

[0012] according to The value splits the original joint velocity sequence into an interactive individual dataset. Compared with independent individual datasets It also carries a membership matrix that records the interaction relationships between individuals. ;

[0013] S2. Input the dataset of independent individuals passed in S1 into a dedicated encoder and decoder; the encoder extracts temporal features from the historical joint velocity sequence of the independent individuals, and the decoder generates preliminary prediction results for the independent individuals based solely on the temporal features of the individual itself;

[0014] S3 receives the interactive individual dataset transmitted by S1, first extracts the dynamic interaction information between individuals through temporal difference operation, and then inputs the information into the cross-modal multi-head cross-attention module. Using the motion features of the interactive individual itself as the query vector and the motion features of other real interactive individuals as the key vector and value vector, feature aggregation is performed to generate preliminary prediction results of the interactive individuals.

[0015] S4. Receive the preliminary prediction results of independent individuals from S2 and the preliminary prediction results of interactive individuals from S3, and simultaneously import the true values ​​of historical sequences and future sequences; calculate the restoration error of the preliminary prediction results to the historical sequences based on the reconstruction loss, calculate the deviation between the preliminary prediction results and the future true sequences based on the prediction loss, fuse the two losses through a weighted formula, and adjust the model parameters in reverse according to the loss values ​​to obtain the optimized prediction results.

[0016] The optimized prediction results include the optimized prediction results for independent individuals and the optimized prediction results for interacting individuals.

[0017] The specific process of loss calculation and parameter optimization includes:

[0018] Reconstruction loss Represented as:

[0019] ;

[0020] In the formula, This represents the model's reconstruction results of historical motion sequences; This is a true historical sequence of events; It is an L2 norm;

[0021] Predicting losses Represented as:

[0022] ;

[0023] In the formula, This is the initial predicted motion sequence output by the model. For real future motion sequences;

[0024] Weighted total loss Represented as:

[0025] ;

[0026] In the formula, The weighting factor is set at 0.6-0.8 to balance the reconstruction and prediction objectives.

[0027] Using the Adam optimizer, the learning rate Weight decay ,based on The parameters of the Transformer codec in S2 and the cross-attention module in S3 are iteratively optimized until... convergence;

[0028] S5. Receive the independent individual optimization prediction results and interactive individual optimization prediction results transmitted by S4. For the scenarios of "large-scale group", "dynamic interaction" and "weak interaction transition state", call the corresponding scenario adaptation algorithm to adjust the prediction results respectively.

[0029] For large-scale groups, a batch processing logic is adopted, based on the membership matrix. The group was divided into small groups of 5-10 people, and the results were spliced ​​together after parallel computation.

[0030] For dynamic interactive scenarios, the membership matrix is ​​updated every 5 frames. Re-execute the double fuzzy rule judgment of S1 and adjust the interaction / independent individual classification in real time.

[0031] For weakly interactive transition states, a smoothing factor is introduced. Calculate the final prediction results for individuals in the weakly interactive transition state. The formula is:

[0032] ;

[0033] Always Predicting outcomes for independent individuals; Predict results for interacting individuals;

[0034] Calculate the joint position error, average position error, and final position error between the predicted and actual results; integrate the predicted and actual results to optimize the prediction results, generate the final future motion posture sequence of all individuals, and generate a quantitative verification report containing joint position error, average position error, and final position error.

[0035] Joint position error Represented as:

[0036] ;

[0037] In the formula, For the number of joints, ; To predict joint coordinates, by The result is obtained by integration; Joint dimensions; The actual label sequence; This is the actual data of the positions of each joint in the human body corresponding to the actual motion sequence;

[0038] Average position error Represented as:

[0039] ;

[0040] In the formula, The optimized predicted motion sequence; For the true sequence and the predicted sequence A frame at the same time point; For the real motion sequence The actual human posture data corresponding to the frame;

[0041] Final position error Represented as:

[0042] ;

[0043] In the formula, This is the prediction result for the last frame.

[0044] Preferably, the execution process of the orientation alignment rule in S1 includes:

[0045] Extract the historical joint velocity sequence trend for each individual to generate a motion orientation vector. Calculate the cosine of the angle between the motion orientation vectors corresponding to any two individuals, and determine whether the cosine of the angle satisfies the following conditions:

[0046] ;

[0047] In the formula, This is a reasonable deviation; The value is the cosine of the included angle.

[0048] Preferably, the execution process of the spatial proximity rule in S1 includes:

[0049] Taking the average coordinates of the key joints of each individual as the body center, the three-dimensional Euclidean distance between the body centers of any two individuals is calculated using the following formula:

[0050] ;

[0051] ;

[0052] In the formula, This is the spatial proximity threshold; , , For the first The three-dimensional coordinates of the body center of each individual; , , For the first The three-dimensional coordinates of the body center of each individual; The distance between the centers of the two bodies is the three-dimensional Euclidean distance.

[0053] Preferably, the codec described in S2 is constructed using the Transformer model, specifically including:

[0054] The encoder employs a 3-layer Transformer encoder, with each layer containing a 4-head self-attention mechanism and a feedforward network. Perform temporal feature extraction and output the encoded features, represented as follows:

[0055] ;

[0056] ;

[0057] In the formula, For the first The high-dimensional temporal features obtained by processing the historical sequence of individuals with no interaction or association with each other through a Transformer encoder; This represents the high-dimensional feature dimension of the encoder output, with a value of 256.

[0058] The decoder uses a 2-layer Transformer decoder, in order to Given the input, predict the joint velocity sequence of future frames and output the preliminary prediction result, represented as:

[0059] ;

[0060] ;

[0061] In the formula, For the first Preliminary predicted motion sequences of individuals with no interactive associations; The preset number of prediction frames, with a value range of [value range missing]. frame.

[0062] Preferably, the specific execution process of S3 includes:

[0063] S301, to Perform temporal difference operations to calculate joint velocity changes between adjacent frames and capture interactive dynamic information. The formula is as follows:

[0064] ;

[0065] ;

[0066] In the formula, For individuals with interactive behavior patterns, the temporal difference sequence is used. For individuals with interaction behavior patterns in the first Historical motion data of frames; For individuals with interaction behavior patterns in the first Historical motion data of frames;

[0067] S302, will Decomposed into its own features With interaction features The differential features of the current individual are represented as:

[0068] ;

[0069] In the formula, This is the feature dimension of the query vector, with a value of 128. For query vector; membership matrix middle The corresponding other individual difference features are represented as follows:

[0070] ;

[0071] In the formula, The key vector; It is a value vector;

[0072] S303. Calculate the output of a single attention head, and then concatenate all the head outputs to obtain the final interaction features. The formula is:

[0073] ;

[0074] In the formula, For the first The output of each attention head; For the first The query vector corresponding to each attention head; For the first The key vector corresponding to each attention head; The normalized activation function; As the key dimension, , The number of attention heads is set to 6. For the first The value vector corresponding to each attention head;

[0075] S304, will Input 2 layers The hidden layer has a dimension of 256, and the output is the preliminary prediction result for the interacting individuals, represented as:

[0076] ;

[0077] ;

[0078] In the formula, Preliminary predicted motion sequences for individuals with interactive behavior patterns.

[0079] Therefore, the present invention employs the aforementioned multi-person pose prediction method based on fuzzy control and differentiated dual-stream modeling. Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0080] (1) Adopting the dual fuzzy rules of "directional alignment and spatial proximity", and combining the output of soft membership degree by fuzzy "AND" operation, we overcome the problem of "misjudgment of interaction state" caused by the default "all individuals are in an interactive state" in the existing technology without distinguishing the actual interaction relationship, and realize the accurate division of interactive individuals and independent individuals, so as to provide accurate state basis for subsequent differentiated modeling;

[0081] (2) Design a dedicated “Transformer codec” for each individual, modeling only based on its own historical sequence, overcoming the problem that existing technologies force individual individuals to introduce irrelevant social context, leading to deviation in trajectory prediction, and achieving isolation of irrelevant social noise, thereby preventing the predicted trajectory of individual individuals from deviating from the real path;

[0082] (3) Design a "temporal difference and cross-modal multi-head cross attention" module for interactive individuals, which aggregates only the features of real interactive individuals, overcomes the problem that existing technologies mix non-interactive individuals as "noise" into interaction modeling, dilute the real interaction signal, and cause inaccurate capture of group dependency, thereby strengthening the capture of group dependency relationship between interactive individuals and improving the prediction accuracy in interactive scenarios.

[0083] (4) Introducing a weighted loss of “reconstruction loss and prediction loss” overcomes the problem that existing technologies only use prediction loss, which leads to the encoder easily losing original spatiotemporal features, poor generalization, and high dependence on labeled data. This reduces the dependence on labeled data and enhances the model’s generalization and prediction stability.

[0084] (5) Design scenario adaptation logic for “large-scale groups”, “dynamic interaction” and “weak interaction transition state”, and combine JPE, APE and FDE core indicators for quantitative verification. Overcome the problem that existing technologies can only adapt to simple scenarios such as small groups and indoors, and have low credibility of technical effects without quantitative indicators for verification. This will enable stable response to complex multi-person scenarios while improving the average prediction accuracy.

[0085] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0086] Figure 1 This is an architecture diagram of an embodiment of the multi-person pose prediction method based on fuzzy control and differentiated dual-stream modeling of the present invention;

[0087] Figure 2 This is a schematic diagram illustrating the interaction discrimination of an embodiment of the multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling of the present invention.

[0088] Figure 3 This is a cross-modal multi-head cross-attention structure diagram of an embodiment of the multi-person pose prediction method based on fuzzy control and differentiated dual-stream modeling of the present invention. Detailed Implementation

[0089] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0090] Example 1

[0091] like Figures 1-3 As shown, the multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling of the present invention includes the following steps:

[0092] S1. Perform a difference operation on the three-dimensional joint coordinates of the human body's historical posture sequence to obtain the joint velocity sequence, represented as:

[0093] ;

[0094] ;

[0095] In the formula, This represents the total number of individuals in the current scenario. For historical frame counts; Joint dimensions; For the first The joint velocity sequence of each individual; A set of joint velocity sequences;

[0096] like frame, This yields joint velocity data;

[0097] The correlation degree of individual motion direction is calculated based on directional alignment, and the correlation degree of individual physical position is calculated based on spatial proximity. Then, the two correlation degrees are fused through fuzzy AND operation to output the soft membership degree of each individual. Based on the soft membership degree threshold, the joint velocity data is divided into interactive individual data and independent individual data.

[0098] Orientation alignment, that is, for any two volumes Calculate the angle between the motion orientation vectors ( for The overall direction vector angle of the joint velocity sequence satisfies the formula:

[0099] ;

[0100] In the formula, To allow for a reasonable deviation, a range of 0.1-0.2 is used, which is then dynamically adjusted based on the specific scenario.

[0101] If satisfied, then the orientation is aligned with the membership degree. ,otherwise ;

[0102] Spatial proximity, i.e., computation The Euclidean distance between the body centers satisfies the formula:

[0103] ;

[0104] ;

[0105] In the formula, The spatial proximity threshold is set at 1.5-2m for public spaces; , , For the first The three-dimensional coordinates of the body center of each individual; , , For the first The three-dimensional coordinates of the body center of each individual; The three-dimensional Euclidean distance between the centers of the two bodies;

[0106] If the condition is met, then the spatial proximity membership degree is... ,otherwise ;

[0107] The comprehensive membership degree is calculated using fuzzy logic operations, and the formula is as follows:

[0108] ;

[0109] In the formula, Soft membership degree;

[0110] like If they are determined to be an interactive pair, Determined to be an independent entity;

[0111] according to The value splits the original joint velocity sequence into an interactive individual dataset. ( (Corresponding individuals) and independent individual datasets ( (corresponding individuals), and carries a membership matrix. (Record the interaction relationships between individuals) and pass them to subsequent steps respectively;

[0112] S2. Input the dataset of independent individuals passed in S1 into a dedicated encoder and decoder; the encoder extracts temporal features from the historical joint velocity sequence of the independent individuals, and the decoder generates preliminary prediction results for the independent individuals based solely on the temporal features of the individual itself;

[0113] In this step, the receiver S1 transmits... To filter out individual data with no interactive associations, the formula is:

[0114] ;

[0115] In the formula, Assign a unique number to each individual; For the first Historical movement sequence data of individuals with no interactive associations;

[0116] The codec is constructed using a Transformer model. The encoder is a 3-layer Transformer encoder, with each layer containing a 4-head self-attention mechanism and a feedforward network (FFN). Perform temporal feature extraction and output the encoded features, represented as follows:

[0117] ;

[0118] ;

[0119] In the formula, For the first The high-dimensional temporal features obtained by processing the historical sequence of individuals with no interaction or association with each other through a Transformer encoder; The high-dimensional feature dimension of the encoder output is set to 256.

[0120] The decoder uses a 2-layer Transformer decoder, in order to As input, predict the future frame( The joint velocity sequence of (frames) is used to output the preliminary prediction result, represented as follows;

[0121] ;

[0122] ;

[0123] In the formula, For the first Preliminary predicted motion sequences of individuals with no interactive associations; The preset number of prediction frames;

[0124] S3 receives the interactive individual dataset from S1, first extracts the dynamic interaction information between individuals through temporal difference operation, and then inputs this information into the cross-modal multi-head cross-attention module, using the motion features of the interactive individual itself as the query vector. The motion characteristics of other real-world interacting individuals are key vectors. AND value vector Feature aggregation is performed to generate preliminary prediction results for interacting individuals; this includes the following steps:

[0125] S301, Receive the data transmitted by S1 With membership matrix ,right Perform temporal difference operations to calculate joint velocity changes between adjacent frames and capture interactive dynamic information. The formula is as follows:

[0126] ;

[0127] ;

[0128] In the formula, For individuals with interactive behavior patterns, the temporal difference sequence is used. For individuals with interaction behavior patterns in the first Historical motion data of frames; For individuals with interaction behavior patterns in the first Historical motion data of frames;

[0129] S302, will Decomposed into its own features With interaction features The differential features of the current individual are represented as:

[0130] ;

[0131] In the formula, This is the feature dimension of the query vector, with a value of 128. For query vector;

[0132] Membership matrix middle The corresponding other individual difference features are represented as follows:

[0133] ;

[0134] In the formula, The key vector; It is a value vector;

[0135] S303. Calculate the output of a single attention head, and then concatenate all the head outputs to obtain the final interaction features. The formula is:

[0136] ;

[0137] In the formula, For the first The output of each attention head; For the first The query vector corresponding to each attention head; For the first The key vector corresponding to each attention head; The normalized activation function; As the key dimension, , The number of attention heads is set to 6. For the first The value vector corresponding to each attention head;

[0138] S304, will Input 2 layers (Hidden layer dimension 256), output the preliminary prediction results for the interacting individuals, represented as:

[0139] ;

[0140] ;

[0141] In the formula, Preliminary predicted motion sequences for individuals with interactive behavior patterns;

[0142] S4. Receive the preliminary prediction results of independent individuals from S2 and the preliminary prediction results of interactive individuals from S3, and simultaneously import the true values ​​of historical sequences and future sequences; calculate the restoration error of the preliminary prediction results to the historical sequences based on the reconstruction loss, calculate the deviation between the preliminary prediction results and the future true sequences based on the prediction loss, fuse the two losses through a weighted formula, and adjust the model parameters in reverse according to the loss values ​​to obtain the optimized prediction results.

[0143] The optimized prediction results include the optimized prediction results for independent individuals and the optimized prediction results for interacting individuals.

[0144] In this step, the receiver of S2 S3 Simultaneously loading real historical sequences (and (corresponding original historical joint velocity), true future sequence (The actual joint velocity to be predicted);

[0145] Reconstruction loss Represented as:

[0146] ;

[0147] In the formula, This represents the model's reconstruction results of historical motion sequences; This is a true historical sequence of events; It is an L2 norm;

[0148] Predicting losses Represented as:

[0149] ;

[0150] In the formula, This is the initial predicted motion sequence output by the model. For real future motion sequences;

[0151] Weighted total loss Represented as:

[0152] ;

[0153] In the formula, The weighting factor is set at 0.6-0.8 to balance the reconstruction and prediction objectives.

[0154] Using the Adam optimizer (learning rate) Weight decay ),based on The parameters of the Transformer codec in S2 and the cross-attention module in S3 are iteratively optimized until... Convergence (e.g., after 500 iterations, the loss decreases to a stable value);

[0155] The optimized prediction results for individual entities ( (Correction value after parameter optimization) and interactive individual prediction results ( (For the corresponding correction amount), it is passed to subsequent steps;

[0156] S5. Receive the independent individual optimization prediction results and interactive individual optimization prediction results transmitted by S4. For the scenarios of "large-scale group", "dynamic interaction" and "weak interaction transition state", call the corresponding scenario adaptation algorithm to adjust the prediction results respectively.

[0157] Calculate the joint position error, average position error, and final position error between the predicted and actual results; integrate the predicted and actual results to optimize the prediction results, generate the final future motion posture sequence of all individuals, and generate a quantitative verification report containing joint position error, average position error, and final position error.

[0158] In this step, the receiver of S4 and The individual sequences are integrated into a total prediction sequence based on individual IDs. ;

[0159] For large groups It adopts a batch processing logic, based on the membership matrix. The group was divided into small groups of 5-10 people, and the results were spliced ​​together after parallel computation.

[0160] For dynamic interactive scenarios, the membership matrix is ​​updated every 5 frames. Re-execute the double fuzzy rule judgment of S1 and adjust the interaction / independent individual classification in real time.

[0161] For weakly interactive transition states Introducing a smoothing factor Calculate the final prediction results for individuals in the weakly interactive transition state. The formula is:

[0162] ;

[0163] Joint position error Represented as:

[0164] ;

[0165] In the formula, For the number of joints, ; To predict joint coordinates, by The result is obtained by integration; Joint dimensions; The actual label sequence; This is the actual data of the positions of each joint in the human body corresponding to the actual motion sequence;

[0166] Average position error Represented as:

[0167] ;

[0168] In the formula, The optimized predicted motion sequence; For the true sequence and the predicted sequence A frame at the same time point; For the real motion sequence The actual human posture data corresponding to the frame;

[0169] Final position error Represented as:

[0170] ;

[0171] In the formula, This is the prediction result for the last frame.

[0172] Therefore, this invention adopts the above-mentioned multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling. This method accurately distinguishes between interaction and independent individuals through dual fuzzy rules, isolates noise and avoids deviation of independent individual trajectories with a dedicated Transformer codec, improves accuracy by purifying signals with an interaction module, and reduces data dependence and enhances stability by weighted loss. After scene adaptation and quantization verification, it stably copes with complex scenes and improves the average prediction accuracy, with reliable results.

[0173] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-person attitude prediction method based on fuzzy control and differentiated dual-stream modeling, characterized in that, Includes the following steps: S1. Obtain the historical motion sequence of multiple people, and use a behavior pattern discrimination mechanism based on fuzzy control, with the following formula: ； In the formula, Soft membership degree; This represents the membership degree for orientation alignment; it is 1 if the orientation alignment rule is satisfied, and 0 otherwise. This represents the spatial proximity membership degree, which is 1 if the spatial proximity rule is satisfied, and 0 otherwise. like If they are determined to be an interactive pair, Determined to be an independent entity; The execution process of the orientation alignment rule, based on the preset orientation alignment rule and spatial proximity rule, is as follows: Extract the historical joint velocity sequence trend for each individual to generate a motion orientation vector. Calculate the cosine of the angle between the motion orientation vectors corresponding to any two individuals, and determine whether the cosine of the angle satisfies the following conditions: ； In the formula, This is a reasonable deviation; The value is the cosine of the included angle; The individuals corresponding to the historical motion sequence are divided into individuals with interactive behavior patterns and corresponding interactive information and individuals with independent behavior patterns. according to The value splits the original joint velocity sequence into an interactive individual dataset. Compared with independent individual datasets It also carries a membership matrix that records the interaction relationships between individuals. ; S2. Input the dataset of independent individuals passed in S1 into a dedicated encoder and decoder; the encoder extracts temporal features from the historical joint velocity sequence of the independent individuals, and the decoder generates preliminary prediction results for the independent individuals based solely on the temporal features of the individual itself; S3 receives the interactive individual dataset transmitted by S1, first extracts the dynamic interaction information between individuals through temporal difference operation, and then inputs the information into the cross-modal multi-head cross-attention module. Using the motion features of the interactive individual itself as the query vector and the motion features of other real interactive individuals as the key vector and value vector, feature aggregation is performed to generate preliminary prediction results of the interactive individuals. S4. Receive the preliminary prediction results of independent individuals from S2 and the preliminary prediction results of interactive individuals from S3, and import the true values ​​of historical sequences and the true values ​​of future sequences. Based on the reconstruction loss, the error of the initial prediction result in restoring the historical sequence is calculated. Combined with the prediction loss, the deviation between the initial prediction result and the future real sequence is calculated. The two losses are fused by a weighted formula. The model parameters are adjusted in reverse according to the loss value to obtain the optimized prediction result. S5. Receive the independent individual optimization prediction results and interactive individual optimization prediction results transmitted by S4. For scenarios of "large-scale groups," "dynamic interactions," and "weak interaction transition states," respectively, call the corresponding scenario adaptation algorithms to adjust the prediction results, specifically as follows: For large-scale groups, a batch processing logic is adopted, based on the membership matrix. The group was divided into groups of 5-10 people, and the results were spliced ​​together after parallel computation. For dynamic interactive scenarios, the membership matrix is ​​updated every 5 frames. Re-execute the double fuzzy rule judgment of S1 and adjust the interactive / independent individual classification in real time; For weakly interactive transition states, a smoothing factor is introduced. Calculate the final prediction results for individuals in the weakly interactive transition state. The formula is: ； Always Predicting outcomes for independent individuals; Predict results for interacting individuals; Calculate the joint position error, average position error, and final position error between the predicted and actual results; integrate the predicted and actual results to generate the final future motion posture sequence for all individuals, and generate a quantitative verification report including joint position error, average position error, and final position error.

2. The multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling according to claim 1, characterized in that, The execution process of the spatial proximity rule described in S1 includes: Taking the average coordinates of the key joints of each individual as the body center, the three-dimensional Euclidean distance between the body centers of any two individuals is calculated using the following formula: ； ； In the formula, This is the spatial proximity threshold; , , For the first The three-dimensional coordinates of the body center of each individual; , , For the first The three-dimensional coordinates of the body center of each individual; The distance between the centers of the two bodies is the three-dimensional Euclidean distance.

3. The multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling according to claim 2, characterized in that, The codec described in S2 is constructed using the Transformer model, specifically including: The encoder employs a 3-layer Transformer encoder, with each layer containing a 4-head self-attention mechanism and a feedforward network. Perform temporal feature extraction and output the encoded features, represented as follows: ； ； In the formula, For the first The high-dimensional temporal features obtained by processing the historical sequence of individuals with no interaction or association with each other through a Transformer encoder; This represents the high-dimensional feature dimension of the encoder output, with a value of 256. For the first Historical motion sequence data of individuals with no interactive associations; For historical frame counts; It is a 3-layer Transformer encoder; The decoder uses a 2-layer Transformer decoder, in order to Given the input, predict the joint velocity sequence of future frames and output the preliminary prediction result, represented as: ； ； In the formula, For the first Preliminary predicted motion sequences of individuals with no interactive associations; The preset number of prediction frames, with a value range of [value range missing]. frame; It is a 2-layer Transformer decoder; For joint dimensions.

4. The multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling according to claim 3, characterized in that, The specific execution process of S3 includes: S301. Perform temporal difference operations on the interactive individual dataset to calculate the joint velocity changes between adjacent frames and capture interactive dynamic information. The formula is as follows: ； ； In the formula, For individuals with interactive behavior patterns, the temporal difference sequence is used. For individuals with interaction behavior patterns in the first Historical motion data of frames; For individuals with interaction behavior patterns in the first Historical motion data of frames; S302, will Decomposed into its own features With interaction features The differential features of the current individual are represented as: ； In the formula, This is the feature dimension of the query vector, with a value of 128. For query vector; membership matrix middle The corresponding other individual difference features are represented as follows: ； In the formula, The key vector; It is a value vector; S303. Calculate the output of a single attention head, and then concatenate all the head outputs to obtain the final interaction features. The formula is: ； In the formula, For the first The output of each attention head; For the first The query vector corresponding to each attention head; For the first The key vector corresponding to each attention head; The normalized activation function; As the key dimension, , The number of attention heads is set to 6. For the first The value vector corresponding to each attention head; S304, will Input 2 layers The hidden layer has a dimension of 256, and the output is the preliminary prediction result for the interacting individuals, represented as: ； ； In the formula, Preliminary predicted motion sequences for individuals with interactive behavior patterns.

5. The multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling according to claim 4, characterized in that, In S4, the optimized prediction results include independent individual optimized prediction results and interactive individual optimized prediction results; The specific process of loss calculation and parameter optimization includes: Reconstruction loss Represented as: ； In the formula, This represents the model's reconstruction results of historical motion sequences; This is a true historical sequence of events; It is an L2 norm; Predicting losses Represented as: ； In the formula, This is the initial predicted motion sequence output by the model. For real future motion sequences; Weighted total loss Represented as: ； In the formula, The weighting factor is set at 0.6-0.8 to balance the reconstruction and prediction objectives. Using the Adam optimizer, the learning rate Weight decay ,based on The parameters of the Transformer codec in S2 and the cross-attention module in S3 are iteratively optimized until... convergence.

6. The multi-person posture prediction method based on fuzzy control and differentiated dual-stream modeling according to claim 5, characterized in that, In S5, the joint position error, average position error, and final position error are expressed as follows: Joint position error Represented as: ； In the formula, For the number of joints, ; To predict joint coordinates, by The integral is obtained. Predict sequences for all individuals; Joint dimensions; The actual label sequence; This is the actual data of the positions of each joint in the human body corresponding to the actual motion sequence; Average position error Represented as: ； In the formula, The optimized predicted motion sequence; For the true sequence and the predicted sequence A frame at the same time point; For the real motion sequence The actual human posture data corresponding to the frame; Final position error Represented as: ； In the formula, The prediction result for the last frame; This is the human pose data for the last frame of the actual motion sequence.

7. A computer device, characterized in that, include: A processor configured to be coupled to memory, read and execute instructions and / or program code in the memory to perform the method as described in any one of claims 1-6.

8. A computer-readable medium, characterized in that, The computer-readable medium stores computer program code that, when executed on a computer, causes the computer to perform the method as described in any one of claims 1-6.

Images