Dance choreography intelligent recommendation method and system

Abstract

The application discloses a choreography intelligent recommendation method and a matching system, which first extracts the mel-frequency cepstral coefficient and the beat signal of audio to generate music context features, and constructs a dancer skill vector in combination with a dancer body parameter; inputs the music features, the dancer features and an existing primitive sequence into a Transformer decoder to output a first probability distribution of action primitives; then, with the aid of a compatibility encoder, the compatibility of action conversion is calculated, the probability distribution is corrected, a new action primitive is sampled, and a complete primitive sequence is generated through cyclic iteration. Finally, a fluency network is used to score multiple sequences, an optimal sequence is selected, and the optimal sequence is converted into skeletal motion data by relying on an action primitive dictionary. The application effectively solves the problems of harsh action connection and action not meeting the body conditions of dancers in the existing choreography scheme.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology for dance choreography, specifically to intelligent recommendation methods and systems for dance choreography. Background Technology

[0002] With the development of computer vision and artificial intelligence technologies, the automatic generation of dance choreography sequences has gradually become a research hotspot, with wide application needs in film and television production, game animation, virtual reality, and dance-assisted teaching. The automation of dance choreography requires the system to generate a complete, coherent, and expressive sequence of dance movements based on given music and dancer characteristics, which is essentially a complex problem of temporal data generation.

[0003] Early methods were primarily rule-based and template-driven, requiring manual definition of the mapping relationship between movements and musical rhythms and melodies. These methods relied on expert experience, had weak generalization ability, and struggled to handle diverse music and dance styles. Subsequently, retrieval and splicing-based solutions emerged, searching for movement segments matching music clips from pre-collected motion capture databases and then splicing them using motion graphs or transition algorithms. While this approach utilizes real human motion data, abrupt transitions often occur at the splicing points, resulting in unnatural movement transitions. Furthermore, the splicing-based synthesis limits the creativity and overall consistency of choreography.

[0004] In recent years, deep generative models, especially sequence modeling networks such as recurrent neural networks, variational autoencoders, and Transformers, have been used to directly generate continuous sequences of human poses from music features. These models typically predict the position or angle of skeletal joints frame-by-frame or segment-by-segment in an autoregressive manner, with the training objective of maximizing the similarity probability between the generated sequence and the real dance sequence. However, because the model architecture focuses on learning the statistical correlation between music and movement, it lacks explicit modeling of the physical transformation laws of the movement itself. This results in frequent abrupt changes in the generated sequences that do not conform to the laws of human kinematics, such as instantaneous joint displacement, violent swaying of the center of gravity, and jumps between poses in adjacent frames that cannot be completed by the human skeletal muscle system. These abnormal movements not only make the choreography visually unusable, but may even be impossible for dancers to actually perform.

[0005] To alleviate this problem, some works introduce kinematic constraints or smoothing post-processing after generation, such as inverse kinematics optimization or spline smoothing of the generated poses. However, these post-processing methods are retrospective corrections and do not fundamentally change the sampling tendency of the generated model, still easily leading to problems such as disjointed overall motion sequences or broken choreography concepts. Another approach is to divide the motion into discrete primitives and constrain the generation by constructing motion transition graphs. Traditional motion graph methods require manually defining states and transition conditions, which are difficult to adapt to large-scale, high-degree-of-freedom dance choreography scenarios.

[0006] Given the current situation, how to integrate the rationality of physical transitions between movements into the deep generation process in a learnable and differentiable way, so that the next movement generated in each step can be naturally connected to the current movement, thereby obtaining a dance choreography sequence that combines musical expressiveness and motion continuity from end to end, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide an intelligent recommendation method and system for dance choreography, which solves the problems of disjointed and physically impractical dance movements generated by existing methods. This invention generates a first probability distribution of candidate motion primitives by inputting musical features and dancer physical parameters into a Transformer decoder; simultaneously, it uses a pre-trained graph convolutional network transformation compatibility encoder to evaluate the transformation compatibility scores between the current primitive and each candidate primitive, thereby correcting the probability distribution and sampling to gradually generate motion primitive sequences; finally, it uses a temporal convolutional network fluency scoring network to score multiple complete primitive sequences, selecting the sequence with the highest score as continuous skeletal motion data for output. This method uses the kinematic connection knowledge inherent in the motion primitive transformation graph as an explicit constraint, acting on each step of the generation process, thereby ensuring that the resulting choreography motion sequence is smooth, fluid, and conforms to the laws of human movement.

[0008] This invention provides an intelligent recommendation method for dance choreography, including:

[0009] S1: Extract the Mel frequency cepstral coefficient sequence signal and the beat point sequence signal of the input music audio, and generate the music context feature signal;

[0010] S2: Generate dancer skill vector signals based on dancer's physical parameters;

[0011] S3: Input the music context feature signal, the dancer skill vector signal and the current primitive index sequence signal into the Transformer decoder, and output the first probability distribution signal for the set of motion primitives;

[0012] S4: Input the last primitive signal in the current primitive index sequence signal and each primitive signal in the action primitive set into the conversion compatibility encoder in pairs to obtain the compatibility score signal; use the compatibility score signal to correct the first probability distribution signal to obtain the second probability distribution signal, and sample from the second probability distribution signal to determine the next primitive signal, and add the next primitive signal to the current primitive index sequence signal;

[0013] S5: Repeat S3 and S4 until the termination condition is met to obtain the complete primitive index sequence signal; input the complete primitive index sequence signal into the fluency scoring network to obtain the fluency score, select the complete primitive index sequence signal with the highest score, and map it into the skeletal motion data signal through the action primitive dictionary for output.

[0014] In one embodiment of the present invention, generating a music context feature signal includes: performing one-hot encoding on a beat point sequence signal to obtain a beat encoding sequence; concatenating a Mel frequency cepstral coefficient sequence signal with the beat encoding sequence in the time dimension to obtain a music feature matrix; inputting the music feature matrix into a pre-trained music encoder, the music encoder comprising stacked bidirectional gated recurrent unit layers and self-attention pooling layers, the bidirectional gated recurrent unit layers performing temporal modeling on the music feature matrix and outputting a hidden state sequence, the self-attention pooling layers calculating the attention weights at each time step in the hidden state sequence and performing a weighted summation of the hidden state sequence to obtain a fixed-dimensional music context feature signal; the music encoder is pre-trained using a large number of dance music samples with the training objective of predicting music beats or audio reconstruction, and the calculation formula for the self-attention pooling layer is as follows:

[0015]

[0016] Where ht is the hidden state at time step t, Watt is the attention projection matrix, batt is the attention bias term, T is the total number of time steps of the music feature matrix, αt is the attention weight at time step t, and exp is the exponential operation.

[0017] In one embodiment of the present invention, the dancer's physical parameters include height, limb length, and joint flexibility range; generating the dancer's skill vector signal includes: normalizing the height, limb length, and joint flexibility range respectively, mapping the normalized values ​​to a preset interval, and concatenating them into a feature vector, inputting the feature vector into a fully connected layer for linear projection, and outputting the dancer's skill vector signal; the dimension of the dancer's skill vector signal is consistent with the dimension of the music context feature signal, so as to perform feature fusion in the Transformer decoder.

[0018] In one embodiment of the present invention, the generation method of the motion primitive set includes: collecting dance motion capture data of multiple dance styles to obtain a three-dimensional coordinate sequence of human skeletal joints; dividing the three-dimensional coordinate sequence into fixed time windows and extracting the angular velocity or bone direction vector of each joint in each window as a motion descriptor; clustering all motion descriptors, with the number of cluster centers set according to the diversity of motion; storing each cluster center as a motion primitive and assigning it an integer index to form a motion primitive dictionary; each entry in the motion primitive dictionary contains a primitive index and corresponding skeletal joint offset data, used to decode the primitive index into continuous skeletal motion.

[0019] In one embodiment of the present invention, the transformation compatibility encoder is an encoder based on a graph convolutional network. Its training process includes: constructing an action primitive transformation graph based on an action primitive set, treating each action primitive as a graph node, and determining the weight of directed edges based on the co-occurrence frequency of adjacent primitive pairs statistically analyzed in the acquired dance sequence; a higher weight indicates a more natural transition between two action primitives in the dance; constructing training samples, each sample consisting of a source primitive index, a target primitive index, and a compatibility label as a supervision signal, the compatibility label representing the transition probability from the source primitive to the target primitive in the transformation graph; and building a graph convolutional network, the graph convolutional network containing... The algorithm contains multiple graph convolutional layers. Each graph convolutional layer aggregates the embeddings of neighboring nodes based on the adjacency relationships of the action primitive transformation graph and updates the node embedding representations. The embedding vectors of the nodes corresponding to the indices of the source and target primitives are extracted, and their dot product or cosine similarity is calculated as the predicted compatibility score. A mean squared error loss function is used to train the graph convolutional network parameters to obtain the transformation compatibility encoder. The transformation compatibility encoder receives a pair of primitive signals and outputs a compatibility score signal representing the naturalness of the transformation from the first primitive to the second primitive. The embedding formula for the convolutional layer is as follows:

[0020]

[0021] Where A is the normalized adjacency matrix of the action primitive transformation graph, H(l) is the embedding of the convolution node of the l-th layer graph, W(l) is the convolution weight matrix of the l-th layer graph, b(l) is the convolution bias vector of the l-th layer graph, and σ is the ReLU activation function.

[0022] In one embodiment of the present invention, a second probability distribution signal is obtained by correcting a first probability distribution signal using a compatibility score signal. This includes: the action primitive set contains multiple candidate primitives; the first probability distribution signal is a vector with a dimension equal to the total number of candidate primitives; the compatibility score signal is a vector composed of the compatibility scores calculated for the current last primitive and each candidate primitive; an exponential function is applied to each element of the compatibility score vector to obtain a positive weight vector; the first probability distribution signal is multiplied element-wise by the positive weight vector to obtain a weighted vector; the weighted vector is normalized by dividing it by the sum of all its elements to obtain the second probability distribution signal; when the compatibility score of a candidate primitive with the current last primitive is lower than that of other candidate primitives, its corresponding weight decreases, thereby reducing the probability of that candidate primitive in the second probability distribution signal. The formula for normalizing the weighted vector is as follows:

[0023]

[0024] Where P1(i) is the first probability distribution value of the i-th action primitive, si is the compatibility score of the i-th action primitive, N is the total number of action primitives, and P2(i) is the second probability distribution value of the i-th action primitive.

[0025] In one embodiment of the present invention, the next primitive signal is determined by sampling from the second probability distribution signal using a temperature-based dynamic sampling method. Specifically, a temperature parameter is set, the value of which changes with the number of primitives generated during sequence generation. In the early stage of sequence generation, the temperature parameter is set to a higher value, which reduces the probability difference between primitives in the second probability distribution signal and promotes action diversity. As the number of generation steps increases, the temperature parameter gradually decreases, making the advantages of high-probability primitives in the second probability distribution signal more prominent and ensuring the stability of action transitions. The logarithm of the probability of each primitive in the second probability distribution signal is taken using the temperature parameter, divided by the temperature parameter, and then the probability distribution after temperature adjustment is recalculated using a normalized exponential function. The next primitive signal is obtained by random sampling from this distribution. The temperature parameter change curve adopts linear decay, and its initial value and decay rate are set according to the characteristics of dance style or music rhythm. Before sampling, the number of consecutive occurrences of the corresponding primitive in the second probability distribution signal in the generated sequence is checked. If the number of consecutive occurrences exceeds a preset upper limit, the probability of the primitive is set to zero.

[0026] In one embodiment of the present invention, the fluency scoring network is a temporal convolutional network, comprising an input embedding layer and multiple stacked residual modules. Each residual module includes a dilated convolutional layer, a weight normalization layer, and a modified linear unit activation function. The dilation factor of the dilated convolutional layer increases in powers of two according to the module depth. The input of the fluency scoring network is a time series of bone displacement vectors obtained by mapping the complete primitive index sequence signal through an action primitive dictionary, and the output is a scalar fluency score. The fluency scoring network is trained using labeled fluency dance sequences and non-fluency dance sequences for a ranking loss function, so that the score of the fluency sequence is higher than that of the non-fluency sequence.

[0027] In one embodiment of the present invention, the termination condition is met when the length of the current primitive index sequence reaches the target sequence length determined according to the length of the beat point sequence signal; repeating S3 and S4 includes retaining multiple candidate sequences at each sampling step, using a beam search algorithm to maintain multiple local sequences simultaneously, and finally generating multiple complete primitive index sequence signals; inputting each complete primitive index sequence signal into a fluency scoring network to obtain a fluency score, comparing the fluency scores of each candidate sequence, and selecting the complete primitive index sequence signal with the highest score for the final mapping output.

[0028] This invention also provides an intelligent recommendation system for dance choreography, comprising:

[0029] The music feature extraction module extracts the Mel frequency cepstral coefficient sequence signal and the beat point sequence signal of the input music audio, and generates music context feature signals;

[0030] The dancer parameter acquisition module generates dancer skill vector signals based on the dancer's physical parameters.

[0031] The primitive probability generation module inputs the music context feature signal, the dancer skill vector signal, and the current primitive index sequence signal into the Transformer decoder and outputs the first probability distribution signal for the set of motion primitives.

[0032] The conversion compatibility correction module pairs the last primitive signal in the current primitive index sequence signal with each primitive signal in the action primitive set and inputs them into the conversion compatibility encoder to obtain a compatibility score signal. The first probability distribution signal is corrected using the compatibility score signal to obtain a second probability distribution signal. The next primitive signal is determined by sampling from the second probability distribution signal and added to the current primitive index sequence signal.

[0033] The sequence generation and output module repeatedly calls the primitive probability generation module and the conversion compatibility correction module until the termination condition is met, to obtain the complete primitive index sequence signal. The complete primitive index sequence signal is then input into the fluency scoring network to obtain a fluency score. The complete primitive index sequence signal with the highest score is selected and mapped into a skeletal motion data signal through the action primitive dictionary for output.

[0034] The intelligent recommendation method and system for dance choreography provided by this invention generates a first probability distribution of candidate motion primitives by inputting music features and dancer physical parameters into a Transformer decoder. Simultaneously, a pre-trained graph convolutional network transformation compatibility encoder is used to evaluate the transformation compatibility scores between the current primitive and each candidate primitive, thereby correcting the probability distribution and sampling to gradually generate motion primitive sequences. Finally, a temporal convolutional network fluency scoring network scores multiple complete primitive sequences, selecting the sequence with the highest score and mapping it to continuous skeletal motion data for output. This method uses the kinematic connection knowledge inherent in the motion primitive transformation graph as an explicit constraint, acting on each step of the generation process, thereby ensuring that the resulting motion sequence is smooth, fluid, and conforms to the laws of human movement. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1Flowchart of intelligent recommendation methods for dance choreography;

[0037] Figure 2 Generate an architecture diagram for music context features;

[0038] Figure 3 Diagram of the sampling architecture for primitive probability generation and compatibility correction;

[0039] Figure 4 Generate an architecture diagram for the complete orchestration process.

[0040] Figure 5 System architecture diagram for an intelligent recommendation system for dance choreography. Detailed Implementation

[0041] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0042] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0043] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0044] Please see Figure 1-5The figure shows the intelligent recommendation method and system for dance choreography of the present invention. The intelligent recommendation method for dance choreography of the present invention includes: S1: extracting the Mel frequency cepstral coefficient sequence signal and the beat point sequence signal of the input music audio, and generating a music context feature signal; S2: generating a dancer skill vector signal based on the dancer's physical parameters; S3: inputting the music context feature signal, the dancer skill vector signal, and the current primitive index sequence signal into a Transformer decoder, and outputting a first probability distribution signal for the motion primitive set; S4: pairing the last primitive signal in the current primitive index sequence signal with each primitive signal in the motion primitive set and inputting them into a conversion compatibility encoder to obtain a compatibility score signal; using the compatibility score signal to correct the first probability distribution signal to obtain a second probability distribution signal, and sampling from the second probability distribution signal to determine the next primitive signal, and adding the next primitive signal to the current primitive index sequence signal; S5: repeating S3 and S4 until the termination condition is met to obtain a complete primitive index sequence signal; inputting the complete primitive index sequence signal into a fluency scoring network to obtain a fluency score, selecting the complete primitive index sequence signal with the highest score, and mapping it to a skeletal motion data signal through a motion primitive dictionary for output.

[0045] like Figure 1As shown, the intelligent dance choreography recommendation method of this invention is a technical solution for automatically generating a coherent dance movement sequence based on given music and dancer characteristics. Its core lies in using movement primitives as generation units, predicting the primitive distribution through a deep learning model, and introducing explicit movement transition compatibility constraints to correct the distribution, thereby ensuring that the generated dance sequence not only matches the music but also possesses movement coherence that conforms to the laws of human motion. This method first processes the input music audio, extracting two types of key acoustic feature signals: one is the Mel-frequency cepstral coefficient sequence signal, and the other is the beat point sequence signal. The Mel-frequency cepstral coefficient is an acoustic feature designed based on the characteristics of human auditory perception, capable of characterizing the short-time spectral envelope of the music signal and reflecting musical content information such as timbre and harmony. The beat point sequence signal identifies the starting time of the rhythmic downbeat in the music and is an important basis for aligning dance movements with the musical rhythm. These two types of signals are fed into a pre-trained music encoder. The role of a music encoder is to perform temporal context modeling on the input acoustic feature sequence, compressing and abstracting it into a fixed-dimensional vector representation, which is the music context feature signal. A typical implementation of a music encoder can use stacked bidirectional gated recurrent units to aggregate forward and backward information from the input temporal sequence, and then use a self-attention pooling mechanism to weight and integrate the hidden states of all time steps, thereby extracting the music context feature signal that is most instructive for generating dance movements. This signal encapsulates high-level semantic information such as the overall style, rhythmic pattern, dynamics, and paragraph structure of the music, and will continue to guide the generation of movements in subsequent steps.

[0046] like Figure 2As shown, this architecture fully realizes the entire process of transformation from raw music audio to fixed-dimensional music context features. It also enriches the algorithm logic with pre-training branches and parameter backpropagation loop structures, overcoming the limitations of single linear operations. The input audio module, as the starting module of the entire process, is responsible for accessing the raw audio materials. Various styles and sampling rates of dance background music can undergo unified format preprocessing in this module, removing environmental noise and invalid redundant segments, laying the raw data foundation for the subsequent two feature extraction steps. The MFCC sequence extraction module focuses on audio spectral feature analysis, relying on mature acoustic feature algorithms to decompose audio frequency domain information, transforming continuously changing sound wave signals into discrete Mel-frequency cepstral coefficient time-series arrays. This array can accurately record implicit musical information such as melody fluctuations and timbre changes, serving as the core data carrier for depicting musical details. The beat sequence extraction module specifically locates all rhythm nodes within the audio, capturing key beat positions such as drum beats and downbeats based on temporal energy changes, marking all beat placements with time coordinates, and forming separate beat time-series data. The beat one-hot encoding module performs digital mapping on discrete beat time-series data, transforming beat information at different time positions into standardized one-hot encoded vectors. This converts beat information from coordinate symbols into computable numerical features, facilitating subsequent fusion with spectral features. The time-dimensional concatenation → music feature matrix module builds upon the previous two feature results, aligning the MFCC sequence with the encoded beat sequence on a unified time axis. It integrates spectral details and rhythmic information through dimensional concatenation, generating a well-structured two-dimensional music feature matrix. Each row of the matrix corresponds to a single time step, and the columns carry both audio acoustic and beat-related attribute data.

[0047] The bidirectional GRU temporal modeling module receives the pre-formed feature matrix and utilizes the forward and backward temporal traversal capability of the bidirectional gated recurrent unit to bidirectionally mine the temporal relationships between preceding and following sequences of music playback, outputting the corresponding hidden state vectors step by step. This completes the deep extraction of global temporal information, upgrading shallow matrix features into high-dimensional hidden features with temporal correlation. The self-attention pooling weight calculation module, relying on a dedicated calculation formula, calculates temporal attention weights for each hidden state output by the GRU, assigning weight values ​​according to the importance of the music content at each time step. Temporal segments corresponding to downbeats and melodic climaxes are assigned higher weight proportions. The weighted summation → music context feature module uses the calculated attention weights to perform weighted operations on all hidden vectors, compressing temporal features of variable length to a fixed vector dimension, ultimately obtaining music context features that can be used for subsequent cross-module fusion. The pre-training branch decision module, as a branch node in the process, divides the formal inference and model pre-training into two different running paths. The branch direction directly determines the computational logic of subsequent modules. When selecting a pre-training path, the module for calculating the beat / audio reconstruction loss is entered. This module uses beat prediction accuracy and original audio signal reconstruction error as dual loss indicators to quantify the goodness of fit of the current network parameters. The module for backdating GRU and attention parameters performs gradient backpropagation based on the loss value, correcting all learnable parameters within the bidirectional GRU and self-attention layers layer by layer. After the parameters are updated, the data is fed back into the bidirectional GRU temporal modeling module, forming a closed-loop iterative training cycle. After multiple rounds of iterative optimization, the feature extraction accuracy of the network model continues to improve. Non-pre-trained inference paths directly jump to the final feature output stage. The entire architecture relies on a combination of parallel feature extraction, branch selection, and closed-loop iteration to ensure the refinement and adaptive optimization capabilities of music feature extraction.

[0048] Parallel to music feature extraction is the modeling of individual dancer characteristics. Dancer physical parameters are collected and used to generate dancer skill vector signals. These parameters include height, limb length, and joint flexibility range, which directly determine the dancer's range of motion and athletic ability. Specifically, continuous values ​​such as height, limb length, and joint flexibility range are normalized, mapped to a unified numerical space, and then concatenated into a feature vector. This vector is then linearly projected through a fully connected layer to obtain a dancer skill vector signal whose dimension matches the music context feature signal. This design allows individual dancer conditions to be integrated into the representation space of the generative model through vector addition or concatenation, ensuring that the final generated dance sequence is adapted to the specific dancer's physical condition and avoiding the generation of movements beyond the dancer's capabilities.

[0049] After obtaining the music context feature signal and the dancer skill vector signal, this method enters the iterative generation stage, generating one motion primitive at each step. A motion primitive is the basic motion unit defined in this method; it is not a continuous frame sequence, but a discrete symbol obtained by clustering and discretizing representative instantaneous postures or short-term motion segments in the dance motion space. The generation method of motion primitives is as follows: First, a large amount of human motion capture data covering various dance styles is collected to obtain a three-dimensional coordinate sequence of multiple joints of the human skeleton changing over time. These coordinate sequences are divided into time windows of fixed length, and feature descriptors describing the local morphology of the motion are extracted from each window, such as the angular velocity of each joint or the bone direction vector. Then, a clustering algorithm is used to cluster the descriptors of all windows. The number of cluster centers is pre-set according to the required motion primitive scale and motion space coverage. Each cluster center is a motion primitive and is assigned a unique integer index. All motion primitives and their corresponding skeletal joint displacement data together constitute a motion primitive dictionary. This dictionary establishes a bidirectional mapping relationship between discrete indices and continuous skeletal motion data, providing a foundation for subsequently reconstructing continuous motion from the primitive sequence.

[0050] like Figure 3 The architecture shown is built around the selection of single-step dance movement primitives and the iterative generation of sequences. The process internally incorporates branch decision logic and a closed-loop sequence update loop, establishing a fusion computation link between three types of data: dancer individual attributes, musical characteristics, and movement connection patterns. The dancer physical parameter input module is responsible for inputting the dancer's original body parameters, including measured data such as height, limb lengths, and joint flexibility. All original parameters are uniformly aggregated within the module for standardization processing, serving as the data source for personalized dance choreography customization. The parameter normalization and splicing module receives scattered, multi-type physical parameters, normalizes and scales them according to preset value ranges, eliminating computational interference caused by differences in the dimensions of different parameters. The processed single-dimensional data is then spliced ​​end-to-end to generate a complete original dancer feature vector. The fully connected projection → dancer skill vector module uses a multi-layer fully connected network to perform linear spatial mapping on the spliced ​​feature vector, adjusting the feature dimensions to be consistent with the musical context features, and finally outputting a standardized dancer skill vector, ensuring that features from two different sources can successfully complete feature fusion computation. The music context feature module, as an external data access port, directly receives fixed-dimensional music features generated by the previous music processing link. These features completely preserve the implicit information of music rhythm and melody, becoming a key constraint for action selection to fit the background music.

[0051] The existing primitive index sequence module stores all action primitive numbers generated in the current step. The sequence content is continuously updated with the addition of each new primitive, serving as both a storage medium for historical choreography results and an important reference for the generation of subsequent actions. The Transformer decoder → First Probability P1 module simultaneously receives three input data: music features, dancer skill vectors, and existing primitive sequences. Relying on the decoder's temporal modeling capabilities, it integrates these three types of information and outputs the initial probability distribution result corresponding to the full set of action primitives. This probability only performs preliminary selection of candidate actions from the perspective of music compatibility and dancer's physical limitations, without considering the smoothness of transitions between adjacent actions. The End Primitive Extraction module extracts the last primitive number from the currently stored existing primitive index sequence, using it as the starting point for calculating the transformation and adaptation effect of subsequent candidate actions. The full action primitive set module contains all standard action units generated through clustering, summarizing all available basic dance action indices and corresponding skeletal data, providing all candidate target samples for compatibility calculation. The GCN compatibility encoder module takes the initial primitive and all candidate primitives as paired inputs. Relying on the topological association rules built into the graph convolutional network and combining the motion co-occurrence patterns obtained from massive amounts of real dance statistics, it calculates the natural fit score for the transition between two consecutive motions. The output module summarizes all the fit values ​​output by the encoder, forming a score vector that matches the number of primitives. This vector is used to correct the initial probability distribution output by the decoder. The exponentially weighted normalization → second probability P2 module inputs the compatibility scores to perform probability weighting operations. Through exponential transformation and global normalization, it merges the initial probability and the transition fit weights to generate a corrected probability distribution that balances music matching and motion continuity.

[0052] The primitive continuous over-limit judgment module, as a key branch node in the process, detects the number of times a candidate primitive appears consecutively in the previous generated sequence and splits it into two different processing paths according to a preset threshold. The over-limit path enters the violation primitive probability zeroing module, which directly sets the probability value of the consecutively over-limit actions to zero, preventing the problem of long-term repetition of a single action in the dance choreography. The data from the non-over-limit path and the data after probability zeroing are uniformly merged into the temperature dynamic sampling module. The module adjusts the dispersion of the probability distribution based on the temperature parameter that decays linearly with the number of generation steps, generating sampling probabilities that meet the needs of the adaptation stage. The next action primitive determination module randomly samples the action units to be added in this round based on the adjusted probabilities. The new primitive number is added to the existing primitive index sequence module. After the data is fed back, a new round of probability calculation loop is started. The entire architecture relies on branch constraints and closed-loop iteration to continuously generate coherent action primitives that fit the conditions of the dancer and the music.

[0053] The core of the iterative generation process is a Transformer decoder that operates in an autoregressive manner, receiving three types of input signals at each step. The first type of input is the music context feature signal generated in the previous steps, providing the decoder with global guidance information within the current music context. The second type of input is the dancer's skill vector signal, enabling the decoder to consider the dancer's physical constraints during prediction. The third type of input is the currently generated primitive index sequence signal, i.e., the sequence of action primitive indices determined by all previous steps. The Transformer decoder internally consists of multiple stacked decoder layers, each containing a multi-head self-attention sublayer with causal masking and a cross-attention sublayer. In the self-attention sublayer, the current primitive index sequence signal, after embedding and positional encoding, calculates the correlation between positions within the sequence using a causal mask self-attention mechanism. The causal mask ensures that each generation time can only observe its own primitives and those from previous times, maintaining the causality of the autoregressive generation. In the cross-attention sublayer, the output of the self-attention sublayer serves as the query vector, and the music context feature signal serves as the key and value vectors. Dot-product attention calculations guide the decoder to focus on the parts of the music features relevant to the current generation time. The dancer's skill vector signal can be concatenated with the primitive embedding representation or added as an additional bias term to the attention weight calculation, thereby influencing the prediction based on the dancer's individual conditions. Each decoder layer is followed by a feedforward network consisting of linear transformations and modified linear unit activation functions to enhance nonlinear expressiveness. The output representation of the last decoder layer is mapped to a vector of the action primitive set dimension through a linear projection matrix, and then passed through a normalized exponent function, i.e., taking the exponent of each element of the vector and dividing by the sum of the exponents of all elements, transforming it into a probability distribution, which is the first probability distribution signal. Each element in the first probability distribution signal represents the initial probability estimate of selecting the corresponding action primitive as the next action under the current music context, dancer conditions, and the generated sequence history. The Transformer decoder is trained using a teacher-forced strategy, i.e., the decoder input at each step during training is a prefix of the real primitive sequence rather than the model's own prediction result from the previous step, and the loss function uses the cross-entropy between the predicted distribution and the one-hot encoding of the real primitive index.

[0054] However, relying solely on the Transformer decoder to select the next motion primitive based on the first probability distribution generated by the music and dancer conditions often results in disjointed motion. This is because the decoder primarily learns the statistical co-occurrence relationship between music and motion during training, lacking explicit modeling of whether the transition between motions can be physically smooth. Even if two motion primitives are statistically highly correlated with the current music segment, the instantaneous switch between them may be physically impossible for the human body to complete. To fundamentally address this technical problem, this method adds a transformation compatibility correction step after the decoder outputs the first probability distribution signal. In this step, an independent, pre-trained motion primitive transformation compatibility encoder is introduced, which is used to evaluate the naturalness of the transition between any two motion primitives.

[0055] like Figure 4 As shown, this architecture represents the top-level operational flow of the entire intelligent dance choreography system. It relies on a cyclic decision structure and multi-sequence optimization branches to achieve a complete chain of batch candidate choreography generation and automatic optimization output, integrating four core business functions: music processing, dancer parameter analysis, movement sequence iteration, and fluency filtering. The music feature extraction module, as the starting unit for music data processing in the entire system, integrates pre-processing logic such as audio import, multi-dimensional acoustic feature extraction, temporal encoding, and feature compression. It centrally completes the conversion of raw audio into standardized feature data. The processed data is then sent to the music context feature generation module, which summarizes all intermediate processing results and outputs a feature vector with unified dimensions, containing music rhythm and melody information, serving as the music benchmark for subsequent movement generation. The dancer parameter acquisition module is specifically responsible for the data collection and input of dancer body indicators, uniformly collecting personalized physical information such as height, limb dimensions, and joint flexibility. After internal normalization and conversion, the raw data flows to the dancer skill vector generation module. This module completes the conversion of individual parameters into standard feature vectors through normalization, feature splicing, and fully connected projection, realizing the digital expression of dancer body conditions.

[0056] The primitive probability generation & compatibility correction sub-module is the core computing unit of the entire system. It integrates multiple sub-steps, including initial motion probability calculation, motion connection compatibility correction, dynamic sampling and primitive selection, and real-time sequence updates. Each call adds a new motion primitive to the existing sequence, continuously expanding the overall length of the primitive index sequence. The sequence length verification module is responsible for determining the start and stop of the loop. It compares the total length of the currently generated sequence with the target sequence length calculated in advance based on the audio beat. If the preset length standard is not met, the process data flows back to the primitive probability generation & compatibility correction sub-module to start a new round of motion primitive generation calculation, relying on iterative looping to complete all the required motion units. When the sequence length meets the set standard, the process switches to the bundle search → multiple complete primitive sequence module. This module uses the bundle search algorithm to simultaneously and in parallel maintain multiple differentiated complete motion index sequences, breaking away from the limitations of single-path generation and producing multiple independent dance choreography options at once, enriching the subsequent selection space. Each sequence is fed into the TCN fluency scoring network module to receive a batch of candidate sequences. All primitive indices are mapped to continuous skeletal displacement temporal data through the built-in action dictionary. The data is then input into the scoring model built on a temporal convolutional network to calculate the fluency quantification score for each choreography scheme.

[0057] The system compares the scores of all candidate sequences across all smoothness scores, horizontally comparing the scores to accurately identify the choreography with the best smoothness performance within the entire candidate sequence set. The highest-scoring primitive sequence selection module locks the target index sequence with the highest score, eliminating other candidate schemes that do not meet the smoothness standard, and simplifying to obtain a unique optimal primitive index result. The primitive dictionary mapping skeletal motion data module retrieves a preset motion primitive dictionary and, according to the one-to-one correspondence between the index and skeletal data, decodes the discrete primitive number sequence into a continuous three-dimensional skeletal coordinate temporal signal of human joints, completing the conversion from symbolic indexes to concrete motion data. The final dance skeletal data output module serves as the final unit of the entire process, exporting standardized skeletal motion files that can be directly used for dance previews, motion replication, and other practical applications. This top-level architecture, through a combination of iterative generation, parallel selection of multiple schemes, and quantitative scoring for optimal selection, ensures that the movements fit the music and the dancer's conditions while selecting the dance choreography with the best smoothness, fully realizing the business goal of intelligent and automatic choreography.

[0058] The construction and training of the motion primitive transformation compatibility encoder are independent of the Transformer decoder. First, based on the aforementioned clustered set of motion primitives, the frequency of adjacent primitive pairs is statistically analyzed using collected dance motion capture data to construct a motion primitive transformation graph. This transformation graph is a directed weighted graph, where each node corresponds to a motion primitive. The weight of a directed edge from primitive A to primitive B represents the frequency or probability of primitive B immediately following primitive A in a real dance sequence. Higher weights indicate a more natural and common transformation from A to B in human motion. Then, this motion primitive transformation graph is used as the adjacency structure input to a graph convolutional network (GCNN) to train the GCNN to learn the embeddings of each node in the graph. The GCNN contains multiple graph convolutional layers. Each GCNN layer performs message passing and aggregation of node features based on the adjacency relationships of the transformation graph, ensuring that the node's embedding representation not only reflects its own motion form but also encodes its structural position in the transformation graph and its transformation relationships with other motion primitives. During training, the embedding vectors of the nodes corresponding to any pair of source and target primitives are extracted, and the similarity between them is calculated as the predicted transition compatibility. The supervision signal uses the transition probabilities of that pair of primitives obtained from the transition graph. By minimizing the mean squared error between the predicted compatibility and the true transition probabilities, the graph convolutional network learns an encoder that can accurately evaluate the naturalness of transitions between action primitives. After training, given any pair of action primitive indices, the transition compatibility encoder can output a scalar value, namely the compatibility score signal, which quantifies the smoothness of the transition from the first primitive to the second primitive.

[0059] In each generation step, the last primitive signal in the currently generated primitive index sequence is obtained. This primitive signal is paired with each candidate primitive signal in the action primitive set, and the pairs are input into the transformation compatibility encoder for inference. For each candidate primitive in the action primitive set, the encoder outputs a compatibility score. The compatibility scores of all candidate primitives constitute a compatibility score signal vector with the same dimension as the first probability distribution signal. Next, the first probability distribution signal is corrected using this compatibility score signal. One specific correction method is to apply an exponential function transformation to each element in the compatibility score signal vector to obtain a weight vector with all positive values. The weight vector is then multiplied element-wise with the first probability distribution signal. The resulting vector is then normalized by dividing by the sum of all elements to finally form the second probability distribution signal. Since candidate primitives with low compatibility scores have smaller weights after the exponential transformation, their product with the corresponding elements of the first probability distribution signal is also reduced, thus giving them a lower selection probability in the second probability distribution signal. Conversely, candidate primitives that have a natural physical connection with the current last primitive and a high compatibility score receive a weight amplification, increasing their probability of being selected in the second probability distribution signal. This mechanism implements explicit kinematic constraints on the generation process at the probabilistic level, ensuring that each primitive selection not only considers musical fit but also prioritizes actions that can be smoothly connected physically.

[0060] After determining the second probability distribution signal, sampling is needed from this probability distribution to determine the next primitive signal. A temperature parameter can be introduced during the sampling process to control the balance between the diversity and determinism of motion generation. The temperature parameter is a mechanism for adjusting the sharpness of the probability distribution. In the early stages of sequence generation, a higher temperature value can be set to reduce the probability differences among primitives in the second probability distribution signal, making the distribution more uniform and encouraging the model to explore more diverse motion combinations. As the number of generation steps increases, the temperature parameter gradually decays to a lower value, amplifying the advantage of the originally high-probability primitives in the second probability distribution signal, making the distribution sharper, and resulting in more stable and coherent generated motions. The decay curve of the temperature parameter can adopt a linear decay scheme. Furthermore, deduplication constraints can be added during sampling, i.e., checking the number of consecutive occurrences of candidate primitives in the generated local sequences. If a primitive has appeared consecutively for more than a preset upper limit, its corresponding probability in the second probability distribution signal is forcibly set to zero to avoid generating monotonous and repetitive dance segments. Finally, random sampling is performed from the second probability distribution signal after temperature adjustment and deduplication based on the probability magnitude to obtain the determined next primitive index signal, and this index signal is appended to the end of the current primitive index sequence signal.

[0061] The method repeats the steps of generating a first probability distribution signal using a decoder, correcting for compatibility to obtain a second probability distribution signal, and sampling to determine the next primitive signal. Each iteration generates one primitive, and the primitive index sequence continuously grows until a preset termination condition is met. The termination condition can be that the sequence length reaches a target length determined by the length of the music beat sequence signal, thus ensuring that the time span of the dance movement sequence matches the music duration. During the repeated generation process, this method can also employ a cluster search strategy, where multiple current primitive sequences are retained as candidates in each generation step. Each candidate sequence is independently used for generating subsequent primitives and performing compatibility correction, ultimately resulting in multiple complete candidate primitive index sequence signals.

[0062] After all candidate primitive index sequences have been generated, their quality needs to be evaluated to select the best choreography result. To this end, this method pre-trains a fluency scoring network. This fluency scoring network preferably adopts a temporal convolutional network architecture, containing stacked residual modules. Each residual module consists of dilated convolutional layers, weight normalization layers, and modified linear unit activation functions. The dilation factor increases exponentially with module depth, enabling the network to obtain a sufficiently large receptive field with fewer layers, capturing long-range coherence patterns in motion sequences. During evaluation, the complete primitive index sequence signal of each candidate is first mapped back to continuous skeletal motion data through an action primitive dictionary, resulting in a time series composed of joint displacement vectors or joint position coordinates. This motion data time series is fed into the fluency scoring network, which outputs a scalar value as the fluency score. The fluency scoring network is trained using a ranking loss function on labeled fluency and non-fluency dance sequence pairs, ensuring that the scores of fluency sequences are systematically higher than those of non-fluency sequences. The fluency scores of all candidate sequences are compared, and the complete primitive index sequence signal with the highest score is selected as the final choreography result.

[0063] Finally, the selected complete primitive index sequence signal is decoded and mapped using a motion primitive dictionary. Specifically, for each primitive index in the sequence, the corresponding relative displacement data of the skeletal joints is retrieved from the dictionary. The displacement is accumulated from the joint position of the previous frame to obtain the joint position of the current frame, and the sequence of skeletal joint positions is reconstructed frame by frame. For minor positional jumps that may occur at the transition between adjacent primitives, a cubic spline interpolation algorithm is used to smooth the skeletal position trajectory, eliminating boundary discontinuities. The final result is a skeletal motion data signal that is continuous over time, smooth in space, and contains the three-dimensional coordinates of multiple major joints of the human body. This signal is the recommended result for dance choreography. This complete process, starting from the music and dancer conditions, uses a combination of deep generation and explicit transformation constraints to generate, end-to-end, a dance movement sequence that conforms to the musical rhythm and style while ensuring the physical feasibility and overall coherence of the movements.

[0064] like Figure 5 This invention also provides an intelligent recommendation system for dance choreography, comprising: a music feature extraction module, which extracts the Mel frequency cepstral coefficient sequence signal and beat point sequence signal of the input music audio and generates a music context feature signal; a dancer parameter acquisition module, which generates a dancer skill vector signal based on the dancer's physical parameters; a primitive probability generation module, which inputs the music context feature signal, the dancer skill vector signal, and the current primitive index sequence signal into a Transformer decoder and outputs a first probability distribution signal for the action primitive set; and a transformation compatibility correction module, which compares the last primitive signal in the current primitive index sequence signal with the primitive signals of each primitive in the action primitive set. The input signals are paired and converted to a compatibility encoder to obtain a compatibility score signal. The first probability distribution signal is corrected using the compatibility score signal to obtain a second probability distribution signal. The next primitive signal is determined by sampling from the second probability distribution signal and added to the current primitive index sequence signal. The sequence generation and output module repeatedly calls the primitive probability generation module and the conversion compatibility correction module until the termination condition is met to obtain a complete primitive index sequence signal. The complete primitive index sequence signal is input into a fluency scoring network to obtain a fluency score. The complete primitive index sequence signal with the highest score is selected and mapped to a skeletal motion data signal through an action primitive dictionary for output.

[0065] The intelligent recommendation method and system for dance choreography of this invention generates a first probability distribution of candidate motion primitives by inputting music features and dancer physical parameters into a Transformer decoder. Simultaneously, a pre-trained graph convolutional network transformation compatibility encoder is used to evaluate the transformation compatibility scores between the current primitive and each candidate primitive, thereby correcting the probability distribution and sampling to gradually generate motion primitive sequences. Finally, a temporal convolutional network fluency scoring network scores multiple complete primitive sequences, selecting the sequence with the highest score and mapping it to continuous skeletal motion data for output. This method uses the kinematic connection knowledge inherent in the motion primitive transformation graph as an explicit constraint, acting on each step of the generation process, thereby ensuring that the resulting motion sequence is smooth, fluid, and conforms to the laws of human movement.

[0066] Therefore, the intelligent dance choreography recommendation method and system of the present invention can solve the problems of disjointed and physically impractical dance movements generated by existing methods.

[0067] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for intelligent recommendation of dance choreography, characterized in that, include: S1: Extract the Mel frequency cepstral coefficient sequence signal and the beat point sequence signal of the input music audio, and generate the music context feature signal; S2: Generate dancer skill vector signals based on dancer's physical parameters; S3: Input the music context feature signal, the dancer skill vector signal and the current primitive index sequence signal into the Transformer decoder, and output the first probability distribution signal for the action primitive set; S4: Input the last primitive signal in the current primitive index sequence signal and each primitive signal in the action primitive set into the conversion compatibility encoder in pairs to obtain a compatibility score signal; use the compatibility score signal to correct the first probability distribution signal to obtain a second probability distribution signal, and sample from the second probability distribution signal to determine the next primitive signal, and add the next primitive signal to the current primitive index sequence signal; S5: Repeat S3 and S4 until the termination condition is met to obtain a complete primitive index sequence signal; input the complete primitive index sequence signal into the fluency scoring network to obtain a fluency score, select the complete primitive index sequence signal with the highest score, and map it into a skeletal motion data signal through the action primitive dictionary for output.

2. The method according to claim 1, characterized in that, The generation of music context feature signals includes: performing one-hot encoding on the beat point sequence signal to obtain a beat encoding sequence; concatenating the Mel frequency cepstral coefficient sequence signal with the beat encoding sequence in the time dimension to obtain a music feature matrix; inputting the music feature matrix into a pre-trained music encoder, which includes stacked bidirectional gated recurrent unit layers and self-attention pooling layers. The bidirectional gated recurrent unit layers perform temporal modeling on the music feature matrix and output a hidden state sequence. The self-attention pooling layers calculate the attention weights at each time step in the hidden state sequence and perform a weighted summation of the hidden state sequence to obtain a fixed-dimensional music context feature signal. The music encoder is pre-trained using a large number of dance music samples with the training objective of predicting music beats or audio reconstruction. The calculation formula for the self-attention pooling layer is as follows: in, The hidden state at time step t, For attention projection matrix, Here, T represents the attention bias term, and T is the total number of time steps in the music feature matrix. Let t be the attention weight at time step t. This is for exponential operations.

3. The method according to claim 1, characterized in that, The dancer's physical parameters include height, limb length, and joint flexibility range. Generating the dancer's skill vector signal involves: normalizing the height, limb length, and joint flexibility range; mapping the normalized values ​​to a preset interval; concatenating them into a feature vector; inputting this feature vector into a fully connected layer for linear projection; and outputting the dancer's skill vector signal. The dimension of the dancer's skill vector signal is consistent with the dimension of the music context feature signal, allowing for feature fusion in the Transformer decoder.

4. The method according to claim 1, characterized in that, The generation method of the motion primitive set includes: collecting dance motion capture data of multiple dance styles to obtain a three-dimensional coordinate sequence of human skeletal joints; dividing the three-dimensional coordinate sequence into fixed time windows and extracting the angular velocity or bone direction vector of each joint in each window as a motion descriptor; clustering all motion descriptors, with the number of cluster centers set according to the diversity of motions; storing each cluster center as a motion primitive and assigning it an integer index to form a motion primitive dictionary; each entry in the motion primitive dictionary contains a primitive index and corresponding skeletal joint offset data, used to decode the primitive index into continuous skeletal motion.

5. The method according to claim 1, characterized in that, The transformation compatibility encoder is a graph convolutional network-based encoder. Its training process includes: constructing an action primitive transformation graph based on the action primitive set, treating each action primitive as a graph node, and determining the weight of directed edges based on the co-occurrence frequency of adjacent primitive pairs statistically analyzed in the acquired dance sequence; higher weights indicate a more natural transition between two action primitives in the dance; constructing training samples, each sample consisting of a source primitive index, a target primitive index, and a compatibility label as a supervision signal, where the compatibility label represents the transition probability from the source primitive to the target primitive in the transformation graph; and building a graph convolutional network containing multiple graph convolutional layers. Each graph convolutional layer aggregates the embeddings of neighboring nodes based on the adjacency relationships of the action primitive transformation graph, and updates the node embedding representation. The embedding vectors of the nodes corresponding to the indices of the source and target primitives are extracted, and their dot product or cosine similarity is calculated as the predicted compatibility score. The mean squared error loss function is used to train the graph convolutional network parameters to predict the difference between the compatibility score and the compatibility label, resulting in the transformation compatibility encoder. The transformation compatibility encoder receives a pair of primitive signals and outputs a compatibility score signal representing the naturalness of the transformation from the first primitive to the second primitive. The embedding formula for the convolutional layer is as follows: in, Normalize the adjacency matrix of the action primitive transformation graph. Embedding of convolutional nodes in the l-th layer graph. Let be the convolution weight matrix of the l-th layer graph. Let l be the graph convolution bias vector of the l-th layer. This is the ReLU activation function.

6. The method according to claim 1, characterized in that, The step of correcting the first probability distribution signal using the compatibility score signal to obtain the second probability distribution signal includes: the action primitive set contains multiple candidate primitives; the first probability distribution signal is a vector with a dimension equal to the total number of candidate primitives; the compatibility score signal is a vector composed of the compatibility scores calculated for the current last primitive and each candidate primitive; an exponential function is applied to each element in the compatibility score vector to obtain a positive weight vector; the first probability distribution signal is multiplied element-wise by the positive weight vector to obtain a weighted vector; the weighted vector is normalized by dividing it by the sum of all its elements to obtain the second probability distribution signal; when the compatibility score of a candidate primitive with the current last primitive is lower than that of other candidate primitives, its corresponding weight decreases, thereby reducing the probability of that candidate primitive in the second probability distribution signal. The formula for normalizing the weighted vector is as follows: in, Let i be the first probability distribution value of the i-th action primitive. Let N be the compatibility score of the i-th action primitive, and N be the total number of action primitives in the set. Let be the second probability distribution value of the i-th action primitive.

7. The method according to claim 1, characterized in that, The step of sampling the next primitive signal from the second probability distribution signal employs a temperature-based dynamic sampling method. Specifically, a temperature parameter is set, the value of which varies with the number of primitives generated during sequence generation. In the early stages of sequence generation, the temperature parameter is set to a higher value, reducing the probability differences among primitives in the second probability distribution signal and promoting action diversity. As the number of generation steps increases, the temperature parameter gradually decreases, making the advantages of high-probability primitives in the second probability distribution signal more prominent and ensuring the stability of action transitions. The logarithm of the probability of each primitive in the second probability distribution signal is taken using the temperature parameter, divided by the temperature parameter, and then the probability distribution after temperature adjustment is recalculated using a normalized exponential function. The next primitive signal is obtained by randomly sampling from this distribution. The temperature parameter change curve adopts linear decay, and its initial value and decay rate are set according to the characteristics of the dance style or music rhythm. Before sampling, check the number of times the corresponding primitive in the second probability distribution signal appears consecutively in the generated sequence. If the number of consecutive occurrences exceeds the preset upper limit, the probability of the primitive is set to zero.

8. The method according to claim 1, characterized in that, The smoothness scoring network is a temporal convolutional network, comprising an input embedding layer and multiple stacked residual modules. Each residual module includes a dilated convolutional layer, a weight normalization layer, and a modified linear unit activation function. The dilation factor of the dilated convolutional layer increases in powers of two with respect to the module depth. The input to the smoothness scoring network is a time series of bone displacement vectors obtained by mapping the complete primitive index sequence signal through an action primitive dictionary. The output is a scalar smoothness score. The smoothness scoring network is trained using labeled smooth dance sequences and non-smooth dance sequences for a ranking loss function, so that the score of the smooth sequence is higher than that of the non-smooth sequence.

9. The method according to claim 1, characterized in that, The termination condition is met when the length of the current primitive index sequence reaches the target sequence length determined by the length of the beat point sequence signal. The repetition of S3 and S4 includes retaining multiple candidate sequences at each sampling step, using a beam search algorithm to maintain multiple local sequences simultaneously, and finally generating multiple complete primitive index sequence signals. Each complete primitive index sequence signal is input into a fluency scoring network to obtain a fluency score. The fluency scores of each candidate sequence are compared, and the complete primitive index sequence signal with the highest score is selected for the final mapping output.

10. A system using the intelligent recommendation method for dance choreography according to any one of claims 1-9, characterized in that, include: The music feature extraction module extracts the Mel frequency cepstral coefficient sequence signal and the beat point sequence signal of the input music audio, and generates music context feature signals; The dancer parameter acquisition module generates dancer skill vector signals based on the dancer's physical parameters. The primitive probability generation module inputs the music context feature signal, the dancer skill vector signal, and the current primitive index sequence signal into the Transformer decoder and outputs a first probability distribution signal for the action primitive set. The conversion compatibility correction module pairs the last primitive signal in the current primitive index sequence signal with each primitive signal in the action primitive set and inputs them into the conversion compatibility encoder to obtain a compatibility score signal. The first probability distribution signal is corrected using the compatibility score signal to obtain a second probability distribution signal. The next primitive signal is determined by sampling from the second probability distribution signal and added to the current primitive index sequence signal. The sequence generation and output module repeatedly calls the primitive probability generation module and the conversion compatibility correction module until the termination condition is met, to obtain a complete primitive index sequence signal. The complete primitive index sequence signal is then input into the fluency scoring network to obtain a fluency score. The complete primitive index sequence signal with the highest score is selected and mapped into a skeletal motion data signal through the action primitive dictionary for output.

Images