Synergistic Application of Modern Dance Education and Creation in Cross-Cultural Perspective

Motion retrieval can be divided into two parts: offline and real-time, after the establishment of the motion database, the choreographer can establish the motion database from a large number of real motion fragments directed graph offline part, the offline part is mainly to retrieve the “difficult action” from the real motion library, calculate its appearance, end of the time point, and then the choreographer through subjective judgment of the “difficulty action” before and after the action fragment can be connected with it, also known as “connected action”. The real-time search part only needs to search the depth or breadth of the established movement “graph” according to the connection relationship between dots and lines. The real-time part is used to select a series of vertices and edges from the established directed graph according to the difficulty movement chosen by the choreographer, and add these calibrated difficulty movements and connected movements to the directed graph structure.

Only after defining the similarity between the motion frames can a “difficult action” be automatically localized from the “original motion sequences”. In this paper, a quaternion-based method is used to define the inter-frame distance: 1D(t1,t2)=‖m(t1)−m(t2)‖=∑i=1nα1d(qi(t1),qi(t2))$D\left( {{t_1},{t_2}} \right) = \left\| {m\left( {{t_1}} \right) - m\left( {{t_2}} \right)} \right\| = \sum\limits_{i = 1}^n {{\alpha_1}} d\left( {{q_i}\left( {{t_1}} \right),{q_i}\left( {{t_2}} \right)} \right)$

where α_i, i = 1,2⋯n represents the importance of the ird joint of the human body, which is a positive number, qi(t1)${q_i}\left( {{t_1}} \right)$ and qi(t2)${q_i}\left( {{t_2}} \right)$ are two quaternions, d(qi(t1),qi(t2))$d\left( {{q_i}\left( {{t_1}} \right),{q_i}\left( {{t_2}} \right)} \right)$ represents its distance, the distance between frame t₁ and frame t₂ can be obtained by equation (1), the distance value indicates the degree of similarity between the two frames, the smaller the value, the higher the similarity, by calculating the distance between the “difficulty action” frame and the “original motion” frame, the position of the “difficulty action” in the “original motion” can be located.

After the choreographer has done the above work, he/she determines the “difficult movements” and “connecting movements” according to the vertices and edges on the specified diagram, and connects all the movements involved in the path formed by connecting the vertices and edges in turn, that is, the final movement that meets the choreographer’s requirements. This is the final movement that meets the requirements of the dance designer. When the traversed path is transformed into a continuous smooth motion, it is necessary to carry out the corresponding coordinate transformation. Since the (x, z)-plane coordinates of the root node of the human body in the original motion segment record the 3D coordinates of the spatial position of the performer, it is necessary to multiply the motion represented by the outgoing edges of a vertex and the incoming edges of the node by a transformation matrix M when the motion represented by the outgoing edges of the vertex is joined together. M are initially unit matrices, and when the action ends on one edge and begins on the next, M are multiplied by the transformation matrices T_θ, x₀, z₀ corresponding to aligning two adjacent edges and updating the iteration. T_θ, x₀, z₀ denote the matrices formed by rotating θ° around the y-axis and translating (x0,z0)$\left( {{x_0},{z_0}} \right)$ in the xz-plane.

Motion editing techniques are varied and cover a wide range of areas. The main motion editing operations we explore are motion path editing, motion blending and motion editing constraints. Motion blending technique can connect multiple motion segments into a new motion segment, and in this paper, we use linear interpolation to compute the blended motion of the overlapping parts. According to 2Pp=α(p)PA1+p+[1−α(p)]PBj−t+1+p${P_p} = \alpha (p){P_{{A_{1 + p}}}} + \left[ {1 - \alpha (p)} \right]{P_{{B_{j - t + 1 + p}}}}$

where P represents the frame, p is an integer that regulates the frame number, and α is a weighting factor (α∈(0,1))$\left( {\alpha \in (0,1)} \right)$, the motion mixing of one segment of motion A_i frames to A_l+k−1 frames and another segment of motion B_j−k+1 frames to B_j frames can be obtained as the translation position of the root node of the human body in the resultant motion in the p th frame P_p, and according to the 3qpi=slerp(qλi+pi,qBj−k+1+pi,α(p))$q_p^i = slerp\left( {q_{{\lambda_{i + p}}}^i,q_{{B_{j - k + 1 + p}}}^i,\alpha (p)} \right)$

It can be seen that the quaternion qpi$q_p^i$ can characterise the degree of rotation of the human joint i in the synthetic motion at frame p, 0 ≤ p ≤ k − 1, and the slerp function in Eq. (3) is the quaternion spherical linear interpolation function.

The simple linear interpolation method does not take into account the constraints of the original motion, which will make the final synthetic motion in the ‘slippery step’ phenomenon, so the linear interpolation method can be obtained although the transition motion is very smooth, but ignores the constraints of the original motion. In order to overcome the defects of the original algorithm, the algorithm in this paper effectively controls the slip-step phenomenon.

In this paper, the motion path extraction and editing algorithm based on multilevel B-splines is used to edit the motion paths, how to locate the motion paths from the oscillating motion trajectory curves is the core problem of this algorithm, and the extraction of the smooth path curves can be regarded as a low-pass filtering or a low-precision approximation of the motion trajectory curves.

Maya software is a new type of animation software under Autodesk, which is widely used in domestic and international animation production, very powerful, and occupies a leading position in many fields such as game development and animation production! In this paper, the character model is selected Maya modelling production.

Polygon modelling technology in Maya is a relatively common modelling method. Currently more mature modelling methods are polygonal modelling methods and NURBS modelling methods. The object of polygonal modelling is fundamentally different from the object of NURBS modelling, which is a parametric surface with only four possible surfaces and a strict UV direction, while the object of polygonal modelling is a topology composed of a series of discrete points in three-dimensional space, with better editing operability. Therefore, in this paper, polygonal modelling method is chosen to construct the character model.

In order to simplify the process of skeleton construction and skin weight binding adjustment, and to generate high-quality character models while efficiently utilising motion capture data, a neural network that predicts and generates skeletons that can be highly matched to the character model and accurately binds the weights, i.e., the neural fusion shaping technique, is used to generate skeletons with the specified structure and accurately bind the skin weights of the skeletons, so as to simplify the deformation process of the character model. The deformation process of the model is simplified. The method not only predicts the skeleton of the mannequin, but also binds and sets up the mannequin structure with the specified skeleton, generating neural mixture shapes, residual displacements and pose-related displacements. The technique mainly contains envelope deformation branch and residual deformation branch. 1)

Envelope deformation branch

The envelope deformation branch learns the parameters of a specific skeleton hierarchy consisting of offsets through indirect supervision, and finally predicts the skeleton, skinning and weight binding from the input character model. Following the typical workflow for binding and skinning, starting with data represented by a mesh with vertices V∈ℝV×3$V \in {\mathbb{R}^{V \times 3}}$, the envelope deformation mesh predicts the offsets of the skeleton, i.e., the offsets of each joint from its parent joint. O∈ℝJ×3$O \in {\mathbb{R}^{J \times 3}}$ using a specified frame hierarchy that contains J joint and skin weight matrices W∈ℝV×J$W \in {\mathbb{R}^{V \times J}}$. The network learns to match the skeleton topology from the 3D data to any character model using indirect supervision. During training, the network is supervised only by the vertex positions of the joints and the corresponding joint rotations. The network learns the relationship between joint rotations and joint features by embedding the estimated binding and skinning parameters in the network. Thus, the exact skeletal binding and skinning parameters of the input character model can be inferred from the training network.

In order to learn binding and skinning parameters that are not provided in the training, a pose R={Ri}whereRi∈ℝ3×3$R = \left\{ {{R_i}} \right\}where\ {R_i} \in {\mathbb{R}^{3 \times 3}}$ represented by local joint rotations is injected in each iteration, which can guide the deformation of the input data and the predicted binding and skinning. Two steps are used to transform the local joint rotations and offsets to globally map each joint transformation Ti∈ℝ4×4${T_i} \in {\mathbb{R}^{4 \times 4}}$. First, for each joint, the locally mapped transformations are accumulated by positive kinematics along its kinematic chain (starting from Root) {Ri,Oi}$\left\{ {{R_i},{O_i}} \right\}$, and then a differential Linear Blending Skinning (LBS) layer is applied to compute the global transformations for each vertex based on the skinning matrix as shown in Eq. (4): 4TRj=∑iWjiTi${T_{Rj}} = \sum\limits_i {{W_{ji}}} {T_i}$

Once the transformation has been computed, the vertex-by-vertex mapping transformation TR={TRj}${T_R} = \left\{ {{T_{{R_j}}}} \right\}$ is applied to the input roles: 5V˜R=TR⊙V${\tilde V_R} = {T_R} \odot V$

where ⊙ denotes a vertex-by-vertex operation that performs a global mapping transformation T_R on the input vertices.

The envelope deformation branch learns the skeletal parameters of a particular skeleton level consisting of offsets through indirect supervision, and finally predicts the skeleton, skinning and weight binding from the input character model. However, the predicted skinning effect is still slightly biased, which leads to skin depression when the skin muscles are flexed and rotated at the joints. Therefore, on this basis, the skinning effect is processed by residual deformation branching. 2)

Residual Deformation Branch

The residual deformation branch can predict the corresponding blend shape according to the input model mesh connection. Taking the blended shape concept as the theoretical basis, the blended shapes and blending coefficients are predicted based on the connectivity of the input meshes and the rotation of the input joints, and a fixed set of residual shapes are predicted, which are interpolated by the coefficients related to the poses and added to the input characters in order to improve the quality of the deformation.

In residual hybrid shapes, given the input vertex position V and connectivity F, the residual branch starts from a pre-trained skin with fixed weights in the envelope deformation branch. The output mask W is then connected to deep vertices V′ along the channel dimension ({V′,W}∈ℝV×(K+J))$\left( {\left\{ {V^\prime ,W} \right\} \in {\mathbb{R}^{V \times (K + J)}}} \right)$ and then fed into the network. The combination of deep vertices and masks provides the hybrid shape network with rich information about the vertex-skeleton relationship, which is crucial for hybrid shape generation. Similar to the envelope deformation branch, a set of N remaining shapes {Bi}i=1N$\left\{ {{B_i}} \right\}_{i = 1}^N$, Bi∈ℝV×3${B_i} \in {\mathbb{R}^{V \times 3}}$ are generated using edge feature representations, via mesh convolution blocks. At the same time, a small neural network containing J MLP blocks, each of which is conditioned by a single joint rotation, is input and outputs a series of pose-dependent coefficients for each joint J that are used to interpolate the residual shapes accrued and added to the input vertices {αij}i=1N$\left\{ {{\alpha_{ij}}} \right\}_{i = 1}^N$: 6V˜=V+∑j=1J∑i=1NαijMjBi$\tilde V = V + \sum\limits_{j = 1}^J {\sum\limits_{i = 1}^N {{\alpha_{ij}}} } {M_j}{B_i}$

where M_j is a binary mask that specifies the vertex associated with joint j. This computation is done by selecting all non-zero entries in the mask matrix associated with the two bones corresponding to the joint. This operation allows us to enhance the localisation in the structure of the mixed shapes and to avoid unwanted deformations of the vertices associated with the static joints. The loss function Lv${\mathcal{L}_v}$ calculates the difference between the character roles and the corresponding real values: 7Lv=‖V˜R−VR‖2${\mathcal{L}_v} = {\left\| {{{\tilde V}_R} - {V_R}} \right\|^2}$

The residual deformation branch can be used to predict the corresponding fusion shape based on the input mesh connections. Meanwhile, the fusion coefficients are predicted based on the joint rotations, and then the compensating deformations are obtained based on this interpolation.

The neural fusion shape technique with the addition of envelope deformation branches and residual deformation branches ultimately generates a character model in *.FBX format, which significantly reduces the manual intervention required in the existing methods, and achieves automatic generation of high-quality 3D character models, which greatly improves the quality of the generated models. The model has a more realistic body shape, more natural folds and fuller muscles.

The number of vertices and edges before and after the elimination of transitivity in the prior modification relation graph is shown in Figure 1.

Figure 1.

Eliminates the number of vertices and edges before and after transitivity

From the figure, it can be seen that for various different dance movements in the listed modern dances, the number of edges in the prior relational graphs decreases significantly after transmissibility elimination of their prior relational graphs. For example, in the hip-hop dance genre, when the number of vertices is 20, the number of edges in the original graph is 53, and the number of edges is reduced to 32 after transitive elimination; in the Latin dance when the number of vertices is 33, the number of edges in the original graph is 76, and the number of edges is reduced to 40 after transitive elimination. This demonstrates that by applying our methodology it is possible to transform the prerequisite relationship graph into a more concise representation.

In dance movement data, learning units can be categorised as multi-attribute time series data. When calculating the movement complexity, two main aspects are considered: node non-correlation as well as motion non-smoothness. The former considers the correlation between nodes, and the latter measures the kinetic parameters of the motion such as the change of velocity direction, or the change of rate. If the correlation between nodes is smaller, the higher the non-correlation is, the more complex the movement is considered to be. The more abrupt changes the action contains, the less smooth and more complex the action is. In this paper, we use the correlation coefficient to measure the correlation between nodes. As far as non-smoothness is concerned, the cubic spline interpolation (CSI) algorithm is used to interpolate the dance movements according to a certain accuracy criterion, where the number of interpolated points is used to measure the non-smoothness of the movements. Finally, the complexity of the movement can then be expressed as a weighted sum of the node non-correlation and the movement non-smoothness.

In this chapter, data sets were collected for the three scenarios above. A student was invited to perform twenty different actions, twice each, and then the complexity of these actions was calculated using the methods presented in this chapter. The twenty actions performed by that student the first time were grouped into sets, and the twenty actions performed by that student the second time were grouped into sets. The comparison of the motion complexity of the motions in A, B is shown in Fig. 2.

Figure 2.

Comparison of motion complexity of motion in A and B

From the circles “●” in the figure, it can be seen that for different subperformances of the same action, their complexity is very approximate, since all the circles “●” are near the y=x line. Then, two more students (a boy and a girl) were invited to perform the same set of actions, and the actions performed by the boy were categorized into sets and the actions performed by the girl were categorized into sets, and we also calculated the complexity of their actions. From the squares “■” in Fig. 2, we can see that the same moves performed by different people also have similar complexity, since all the squares “■” are also near the y = x line. On the other hand, the study has chosen a number of pairs of motions that look different in complexity and calculated their complexity. The results are shown in the figure with the triangles “▲”, and it can be seen that all these triangles “▲” are located far away from the y=x line, which illustrates that for motions that appear to be of different complexity, the complexity computed by our method is also different.

On the other hand, the study applied the approximate entropy approach instead of the complexity method proposed in this chapter repeats the above experiment, and the comparison of the approximate entropy of the motions in A, and B is shown in Fig. 3.

Figure 3.

Comparison of approximate entropy of motion in A and B

By looking at Figure 3, we can see that for the same action, whether performed by the same person (circle “●”) or by a different person (square “■”), their proximity cannons are not quite the same as we can see that both circle “●” and square “■” are both far from the y=x line. Also, the approximate entropy approach is not able to distinguish between actions of different complexity, as shown in the triangle “▲” in Fig. We can see that relative to the comparison of movement complexity in Fig. 2, the triangle “▲” in the comparison of approximate entropy in Fig. 3 is closer to the line y=x some.

1)

Purpose of the experiment

This study applies the designed teaching model of applying neural fusion technology to generate a modern dance course from a cross-cultural perspective to the teaching of modern dance elective courses in case colleges and universities, and compares it with the traditional teaching model to highlight the application effect. Thus, the teaching model proposed in this paper is verified to have an impact on modern dance beginners in terms of dance skill level, physical quality, sports independent learning ability and learning interest, and suitable solutions and suggestions are proposed for the problems existing in the actual teaching of modern dance courses in colleges and universities.

1)

Analysis of the difference in body shape between the two groups of students before the experiment

Before the official start of the teaching experiment, all the students participating in this teaching experiment will be unified measurement of the relevant indexes, after the recovery of the data were counted out, analyzed by the SPSS test to obtain the following results, the analysis of the students’ body morphology before the experiment is shown in Figure 4.

Figure 4.

Analysis of students’ body shape before experiment

As shown in Figure 4, the P-values for height and weight of girls in the experimental and control groups are 0.634 and 0.587 respectively (P>0.05), and the P-values of height and weight of the experimental and control groups of male students are 0.667 and 0763 (P>0.05), respectively, and these data scores are sufficient to fully prove that there is no significant difference in the level of physical form possessed by the two groups of students participating in the teaching experiment in this case. Therefore, all the students participating in this experiment meet the requirements of this teaching experiment in terms of physical form. 2)

Analysis of the differences in physical fitness between the two groups of students before the experiment

There are often differences between individual students, and differences in physical quality may cause differences in the acceptability of knowledge, speed of mastery of skills and accuracy of mastery of physical dance. In order to ensure a high degree of rigor in the experiment, it is very necessary to test the physical quality of the students before the experiment formally begins, and according to the characteristics of the physical dance program, under the guidance of the tutor and a number of experts, to According to the characteristics of the sports dance program, under the guidance of the tutor and several experts, the appropriate physical quality items are selected and tested, and the following results are obtained, and the differences in the physical quality of the two groups of students before the experiment are analyzed as shown in Table 1.

As shown in Table 1, the physical quality data measured before the beginning of the teaching experiment were imported into SPSS, and independent samples t-tests were conducted on the scores of different groups of the same item and the same gender, respectively, and these physical quality test items can clearly reflect the differences in the physical quality of all students in the two groups before the experiment. From the above results, we can know that the P-value of the four physical fitness items is greater than 0.05. These data scores are sufficient to prove that the physical fitness level of all students in the two groups participating in the teaching experiment in this case is not very different and will not affect the results of the teaching experiment. 3)

Analysis of differences in students’ interest in learning and independent learning ability between the two groups before the experiment

Cultivating students’ interest in learning and independent learning ability can effectively improve the efficiency of classroom learning, before conducting the experiment, questionnaires about interest in learning and independent learning were distributed to students in the experimental group and the control group, and the data were statistically analyzed. The results of the survey on students’ interest in learning modern dance and independent learning ability in the two groups before the experiment are shown in Figure 5.

Figure 5.

Study interest and independent learning ability results before the experiment

According to the experimental data in Figure 5, the mean value of girls’ interest in learning modern dance in the experimental group is 34.91, and the mean value of boys’ interest in learning modern dance is 31.52, while the mean values of male and female students in the control group are 35.12 and 32.02 respectively, and the difference between the two groups of data is not statistically significant (P>0.05), and the P value of the two groups of students’ independent learning ability is 0.671 (girls ) and 0.869 (boys), the P-value is greater than 0.05, there is no significant difference. It means that the learning interest and independent learning ability of the two groups of students basically remain at the same level before the experiment, which is in line with the experimental conditions of homogeneous samples.

1)

Analysis of the difference in the performance of physical dance skills between the two groups of students after the experiment

The final performance evaluation of this modern dance teaching experiment contains five dimensions: five aspects of dance movement fluency, basic technique - middle, basic technique - lower limbs, body posture, rhythmic accuracy, 20 points for each dimension, totaling 100 points. The final results of the experimental group and the control group were analyzed in terms of each component, and Figure 6 shows the statistical results of the students’ performance in various physical dance skills after the experiment, and Figure 7 shows the results of the comparison of the total results.

Figure 6.

Statistical results of dance skills after the experiment

Figure 7.

Comparison of total performance of dance skills after experiment

As can be seen from Figure 6 and Figure 7, after the teaching experiment of this semester, there is no significant difference between the scores of the two groups of students in the two items of basic technique - middle section and basic technique - lower limbs, and the P values of the two are 0.097 and 0.114 respectively, which are both less than 0.05, indicating that the impact of the two teaching modes on the middle section technique and lower limbs technique of the beginners of modern dance is not different from each other; and after the experiment, the experimental group has significantly higher than the control group in the three items of scores of movement fluency After the experiment, the experimental group’s movement fluency, body posture and rhythmic accuracy scores were significantly higher than those of the control group, and there was a significant difference (P<0.01). Reflected in the total score, the experimental group’s average score of modern dance at the end of the term was 79.07±2.35, and the control group’s average score of physical education dance at the end of the term was 74.17±2.63, obviously the experimental group’s score was higher, and the results of the t-test of the independent samples showed that the p-value was 0.000 (P<0.01). Therefore, it is concluded that the teaching mode of generating modern dance courses by applying neural fusion technology under the cross-cultural perspective has a more obvious effect on students’ modern dance technique compared with the traditional mode, and it is more conducive to improving the dance technique level of beginners in modern dance. 2)

Analysis of the difference in physical fitness between the two groups of students after the experiment

Enhancing students’ physical fitness is the primary task of physical education. After 15 weeks of physical education dance teaching experiment, the physical quality of the two groups of students was measured again, and the results of the physical quality analysis of the two groups of students after the experiment are shown in Table 2.

As can be seen from Table 2, there is no significant difference between the scores of students in the experimental group and the control group in the qualities of standing long jump (lower limb strength) and 10s in situ leg raise (speed of movement) after the experiment, and the P-values of the male and female students are 0.166, 0.874 and 0.343, 0.173, respectively (P>0.05), which proves that there is no significant difference between the two teaching modes in terms of the lower limb strength and the speed of movement of the beginners of physical dance. The difference in the effects is not great. At the same time, it can be seen that there is a significant difference between the experimental group and the control group in the scores of the two physical qualities of the experimental group in the Baselift (comprehensive balance) and the 30s cross quadrant jump (sensitivity and coordination), with P-values of 0.038, 0.003 and 0.025, 0.008, respectively (P<0.05), and that the girls are better than the boys in terms of the enhancement effect. It is proved that the teaching mode of modern dance course based on the application of neural fusion technology under the cross-cultural perspective can promote the development of the comprehensive balance, agility and coordination qualities of beginners in physical dance more significantly than the traditional mode. By applying the teaching model proposed in this paper in modern dance course teaching, the efficiency of learning new knowledge in class can be improved by enriching students’ after-class practice, so that more time can be used for students’ communication and discussion and practice demonstration, and the practice can reach the accumulation of quantity, which can ultimately improve the physical quality corresponding to it. 3)

Analysis of the differences in learning interest and independent learning ability between the two groups of students after the experiment

Learning interest is an important driving force for students’ growth and learning. Students who are full of love and strong interest in the technical movements they want to learn will be more motivated to learn, thus improving learning efficiency and promoting the level of modern dance. And the cultivation of independent learning ability helps to cultivate students’ innovative ability and lifelong learning consciousness. The results of the investigation of the interest in learning modern dance and independent learning ability of the two groups of students after the experiment are shown in Figure 8.

Figure 8.

Study interest and independent learning ability results after the experiment

According to the analysis of the experimental data in Figure 8, it can be observed that after the teaching experiment, the results of the interest in learning modern dance of the experimental group of female students increased to 38.64, and the male students’ increased to 35.95, while the results of the control group of students were 35.28 and 33.04, and the students of the two groups have significant differences in learning interest (P < 0.05). The test results of independent learning ability of male and female students in the experimental group were 39.96 and 34.61 respectively, while the test results of independent learning ability of male and female students in the control group were 34.17 and 33.58 respectively. There is a significant difference between the two groups of students in terms of independent learning ability (P < 0.05), and the improvement of independent learning ability of female students is significantly greater than that of male students. It shows that compared with the traditional teaching students’ learning interest and independent learning ability, the application of neural fusion technology under the cross-cultural perspective proposed in this paper to generate modern dance course assisted teaching is more conducive to enhancing their learning level.