Design and Implementation of Digital Dance Teaching Platform Based on Kinect

In the era of digitalization, the field of education is also facing unprecedented challenges and opportunities [1]. Dance education in colleges and universities, as an important part of art education [2], has also been profoundly affected by the wave of digitization. The rapid development of digital technology has not only changed people’s way of life and learning, but also brought brand-new possibilities and needs for the curriculum design and teaching methods of dance majors [3-4]. In the digital era, dance education is facing multiple challenges, and the traditional dance teaching mode has been difficult to meet the diverse needs of students and changes in learning styles [5]. At the same time, the rapid development of digital technology also brings new opportunities for teaching dance majors [6]. The introduction of virtual reality, augmented reality, online platforms and other technological tools provide innovative teaching means and resources for dance education, enriching students’ learning content and methods [7]. Therefore it is particularly important to explore the curriculum design and teaching methods for dance majors in colleges and universities [8]. It is necessary to deeply study the teaching mode and tools in the digital era, and explore how to effectively integrate digital technology and the connotation of dance education in order to improve the quality of teaching and students’ learning experience [9-11]. The digital era provides us with more possibilities, how to maintain the core value of art while facing the development of technology is an important issue we need to face and solve [12-13].

For students in higher vocational colleges and universities, the teaching of dance professional courses should optimize the content of theoretical courses, but also need to strengthen the quality of practical teaching [14]. Only from a practical point of view, to improve the quality level of dance teaching in higher vocational colleges and universities, and synchronously strengthen the teaching of both theory and practice, can we meet the requirements of dance teaching in vocational colleges and universities in the new era, improve the quality of the online platform design and application, and meet the differentiated learning needs of students [15-17].

The construction of a functional and structurally complete online learning platform is also conducive to helping students introduce more online learning resources and improve the quality of dance professional talent training [18-19]. Analysis of the Service Group Profile of Online Teaching Platform for Dance Majors in Higher Vocational Colleges and Universities In the process of teaching dance in higher vocational colleges and universities, the construction of an online teaching platform is mainly aimed at improving better online teaching services [20]. The advantages of the convenience and diversity of information interaction of the online platform are fully utilized to provide students with a better online learning experience [21].

Based on Kinect and combining human movement recognition technology and human movement comparison technology, this paper builds and designs a digital dance teaching platform. In the human movement recognition technology, a fixed-axis-based joint point angle representation method is proposed, combined with the pose recognition of mean Hausdorff distance and the pose recognition method of Hidden Markov Model, which ensures that the to-be-measured line and the baseline are relatively stable, and accurately recognizes the human body poses. In the human movement comparison technique, the introduction of a preprocessing process for extracting key frames from the raw skeletal data greatly reduces the amount of computation required for analysis. For the problem that the dynamic time regularization algorithm is less efficient for longer action sequences, i.e., when the amount of data is large, an improved longest common subsequence algorithm in dynamic programming is proposed to solve the problem. Finally, the digital dance teaching platform constructed in this paper is applied in practice and analyzed from two aspects, namely, platform performance and dance teaching effect.

2

Digital Dance Teaching Platform Design

With the development of information technology, the emergence of depth camera represented by Kinect makes it easier for people to obtain the human body movement related features such as human skeleton coordinates position, which also provides a new data source for human movement comparison analysis.

Combined with Kinect, this paper proposes a human movement recognition algorithm and a human movement comparison algorithm, and builds a digital dance teaching platform based on them.

2.1

Human Movement Recognition Algorithm

Through human pose recognition, people can communicate with computers through gestures. However, the complexity of human movement and the effect of occluded light make markerless pose recognition very difficult.

2.1.1

Pose Recognition Based on Mean Hausdorff Distance

Hausdorff distance is a measure used to describe the similarity between two given point sets A = {a₁,a₂,…a_m} and B = {b₁,b₂,…b_n}. The Hausdorff distance between A and B can be expressed as: (1) $H (A, B) = \max (h (A, B), h (B, A))$ (2) $h (A, B) = \max_{a \in A} \min_{b \in B} ‖ a - b ‖, h (B, A) = \max_{b \in B} \min_{a \in A} ‖ b - a ‖$

║~║ denotes the Euclidean distance between A and B . Function h (A,B) is the directed Hausdorff distance from A to B . The size of the Hausdorff distance is proportional to the similarity of the two sets of A, B, i.e., the smaller the Hausdorff distance is, the more similar the two sets of A, B are.

Noise will have a large impact on the Hausdorff distance, which directly leads to the bias of the recognition results, Dubuission et al. proposed the concept of partial Hausdorff distance, i.e.: (3) $H_{L K} (A, B) = \max {h_{L} (A, B), h_{K} (B, A)}$ (4) $h_{L} (A, B) = L_{a \in = A}^{t h} d_{B} (a), h_{K} (B, A) = \underset{b \in B}{K^{t h}} d_{A} (b)$

where the meaning of $L_{a = A}^{t h}$ is the L rd value in the A–B sequence, the meaning of $K_{b = B}^{t h}$ is the K th value in the B–A sequence, 0≤L≤m, 0≤K≤n. Partial Hausdorff distances are mismatched if the target in the image is occluded or interfered by strong noise. Therefore the Mean Hausdorff Distance (MHD) is proposed, defined as: (5) $H (A, B) = \max (h (A, B), h (B, A))$ (6) $h (A, B) = \frac{1}{r} \sum_{a \in A} \min_{b \in B} ‖ a - b ‖ h (B, A) = \frac{1}{s} \sum_{b \in B} \min_{a \in A} ‖ b - a ‖$

where r and s denote the number of elements in A and B.

Using the mean Hausdorff distance to determine the poses, the sample sequence and the sequence to be recognized are downscaled to three dimensions to obtain the sample sequence S_i and the sequence to be recognized S, and the mean Hausdorff distances of S_j. and S are obtained respectively, and the discriminative criteria are: (7) $B = \arg \min_{i} (H (S, S_{i})), i = 1, 2, 3, 4, 5$

2.1.2

Hidden Markov Model-based Pose Recognition Approach

1)

Markov model

Assuming that the system has N state S = {S₁,S₂,…,S_n}, when the state of the system shifts to another state over time, assuming that the state at t is Q^t = {q₁,q₂,…,q_t}, the probability of occurrence of the system’s state S_j at t is directly related to the previous state, and the relationship can be expressed as: (8) $P (q_{t} = S_{j} | q_{t - 1} = S_{i}, q_{t - 2} = S_{k}, \dots)$

If the state of the system at time t is only related to the state at time t–1, then this constitutes a discrete first-order Markov chain.

(9)

P (q_{t} = S_{j} | q_{t - 1} = S_{i}, q_{t - 2} = S_{k}, \dots) = P (q_{t} = S_{j} | q_{t - 1} = S_{i})

If only stochastic processes with independent time t are considered: (10) $P (q_{t} = S_{j} | q_{i - 1} = S_{i}) = a_{i j}, 1 \leq i, j \leq n$

where the state transfer probability a_ij must satisfy a_ij≥0 and $\sum_{j = 1}^{n} a_{i j} = 1$ , then the process is called a Markov model.

2)

Hidden Markov Models

In a Markov model, states and observations correspond to each other, and the observer can directly observe the states. Hidden Markov models are composed of two stochastic processes, one of which has a Markov chain of finite states, and the other stochastic process is a random probability function, which relates each state to an observation. From the observer’s point of view, only the observations can be seen, and their existence and properties can only be perceived through a stochastic process, hence the name Hidden Markov Model.

In a Hidden Markov Model, each random event has a sequence of observations corresponding to it as O₁,O₂,…,O_t, and there is a hidden time series q₁,q₂,…,q_r. There are three preconditions for the establishment of a Hidden Markov Model, and each of the three assumptions is as follows.

Assumption I, Markovianity Assumption: (11) $P (q_{i} | q_{i - 1}, \dots, q_{1}) = P (q_{i} | q_{i - 1})$

Assumption two, the immobility assumption: (12) $P (q_{i + 1} | q_{i}) = P (q_{j + 1} | q_{j})$

Hypothesis three, the output independence hypothesis: (13) $P (O_{1}, O_{2}, \dots, O_{T} | q_{1}, q_{2}, \dots, q_{T}) = \prod_{t = 1}^{T} P (O_{t} | q_{t})$

A Hidden Markov Model can be defined as a model λ consisting of 5 parameters, which can be denoted as: (14) $λ = (N, M, A, B, π)$

Or it can be abbreviated as: (15) $λ = (π, A, B)$

Where N = {q₁,q₂,…,q_n} finite set of states.

M = [v₁, v₂,…,v_M} Finite set of observations.

A = {a_ij}, a_ij = P(q_t = s_i|q_t–1 = s_i) State transfer probability matrix.

B = {b_jk}, b_jk = P(O_t = v_k |q_t = s_i) Observation probability distribution matrix.

π = {π_i}, π_i = P(q₁ = s_i) Initial state probability distribution.

In practice, solving the following three basic problems is a prerequisite for using Hidden Markov Models.

Valuation problem. Assuming that a Hidden Markov Model is given and its transfer probabilities a_ij and b_jk are known, the probability that this model produces a sequence of visible states is: (16) $P (O) = \sum_{r = 1}^{r_{\max}} {P (O | Q_{r})} P (Q_{r})$

Decoding Problem. First given the model and the observation sequence O, the hidden state sequence max_Q{P(Q|O)} of this observation sequence is generated by finding its maximum likelihood, i.e.

Learning problem. If the general structure of a Hidden Markov Model is known, but a_ij. and b_jk are not given, how to obtain these parameters from training with a set of known symbols.

The recognition process is divided into the learning process and the valuation process, five sets of Markov model parameters can be obtained after the training of the corresponding discrete training data λ_i= (π_i, A_i, B_i), i = 1,2,…,5, and the discrete sequence probability to be recognized P(O|λ_i), i = 1,2,…,5, and then select the one with the largest probability as the result of the pose recognition. Assuming that the initial model is $π_{i} \approx \frac{1}{N}$ , $\sum_{i} π_{i} = 1$ , $a_{i j} \approx \frac{1}{N}$ , $\sum_{j} a_{i j} = 1$ , $b_{j k} \approx \frac{1}{M}$ , $\sum_{k} b_{j k} = 1$ , the discriminative criteria are: (17) $B = \arg \max P (O | λ_{i})$

2.1.3

Pose Recognition Based on Joint Point Angle Measurement

Let the coordinates of two points in space be X(x₁,x₂,x₃), Y(y₁,y₂,y₃), then the Euclidean distance between these two points can be obtained: (18) $D (X, Y) = \sqrt{{(x_{1} - y_{1})}^{2} + {(x_{2} - y_{2})}^{2} + {(x_{3} - y_{3})}^{2}}$

In order to get the angle of entrapment of any joint of the body, we can get it from the coordinates of three points of the joint. The cosine theorem is utilized to calculate to get the pinch angle of that joint: (19) ${\begin{array}{l} a = D (B, C) \\ b = D (A, C) \\ c = D (A, B) \end{array}$ (20) $θ = \cos^{- 1} \frac{(a^{2} - b^{2} + c^{2})}{2 a c}$

The above joint coordinate-based angle calculation method is theoretically feasible, but in practical applications because the joint points are unstable with each other, the error of the obtained results is large, and cannot be directly used for pose recognition. Therefore, this paper proposes a fixed-axis-based angle representation method, i.e., the positive direction of the and axis is used as the reference line, the line between the two joints is used as the line to be measured, the line to be measured is positively oriented in the outward direction of the human body’s central axis, and the transverse axis of the shoulder is positively oriented in the outward direction of the body’s central axis as the center, and the angle between the line to be measured and the reference line is obtained in a counterclockwise order, and this angle is defined as the angle of these two joint points. Using this method can ensure that the line to be measured and the reference line are relatively stable, ensuring the accuracy of the angle measurement.

This paper defines the joint point angle as: (21) $P = {P_{1}, P_{2}, θ, τ}$

In Eq. (21), P₁ and P₂ are the two joints, θ is the angle between the line connecting the two joints and the baseline, and τ is the angle adjustment threshold. The definition of different poses can be translated into the definition of the angle between the joints, and the angle adjustment threshold can be used to control different accuracy requirements. Assuming θ₁ = (Left shoulder, Left elbow), θ₂ = (Left elbow, Left wrist), θ₃ = (Right shoulder, Right elbow), θ₄ = (Right elbow, Right wrist), then the pose of the upper body can be defined as: (22) $Δ = (θ_{1}, θ_{2}, θ_{3}, θ_{4}, τ)$

After setting the queue range of angles corresponding to the pose, all the angle values are first read and it is judged whether the values of θ₁, θ₂, θ₃ and θ₄ are within the pre-set threshold range. If all the angles are within the threshold range, then the pose matching is recognized as successful, and if any one of the angles is not within the threshold range, then the pose matching is recognized as a failure and needs to be re-matched, and the matching requirements are shown in the following equation (23): (23) $\max_{i \in N} | θ_{i} - α_{i} | < τ$

Where, θ_i is the actual measurement angle, α_i is the target angle and τ is the pre-set threshold.

When all the angles are within the previously set thresholds, then the body postures are correctly recognized, such as lowering the hands, π posture, raising the hands, raising the left hand and other actions.

2.2

Human Motion Comparison Algorithm

2.2.1

Pre-processing of skeletal data

In this paper, the idea of dimensionality reduction is introduced to extract key frames (poses) in skeletal frame data. In the algorithm of image processing, the practice of dimensionality reduction is to decompose the video sequence into multiple image frames and then extract and select the key images, this idea can also be directly applied to the dimensionality reduction of the time series, that is, to select and extract some of the representative data frames from the long action sequence to represent the whole time series, so as to reduce the high-dimensional time series to the low-dimensional in another perspective. Keyframe extraction is also widely used in human movement analysis, action data retrieval, and other applications. Unlike the keyframes in the traditional 2-dimensional time series, the keyframes in the action domain refer to the 3-dimensional skeletal data frames that can represent the entire action sequence.

Assuming that all the data involved in the operation are B = {b₁,b₂,b₃…b_m}.

1)

Select the number of clusters K

2)

Select the initial centers of each of these K classes {c₁, c₂…c_K}

3)

Calculate the Euclidean distance d² = (b_i–c_j)² iϵ[1,m] of the remaining points to the initial centers, and jϵ[1,K] group the data points b_i to the centers that are close in distance. Then calculate the centers ${{c^{'}}_{1}, {c^{'}}_{2} \dots {c^{'}}_{K}}$ of different classes separately as the new clustering center points.

4)

Repeat the steps 2) 3) until the clustering centers reach the required convergence conditions.

2.2.2

Characterization

An action sequence can be viewed as a collection of skeletal data, and the coordinates of skeletal joint point positions captured by Kinect can also be considered as a type of feature representation. If we want to analyze and compare the differences between two action sequences, the essence is to compare the differences of features in the action sequences, and some of the methods of feature representation are described in the following.

1)

Angular features

Angular features are extracted in 20 skeletal joint points, if the angular features are calculated for a particular frame of the action sequence, the angle γ is used to denote the angle of pinch in the elbow joint point p₉, and the two ends are composed of shoulder joint point p₂ and wrist joint point p₁₁. Assuming that the coordinates of p₉ are (x₉, y₉, z₉), and the coordinates of p₂ and p₁₁ are (x₂,y₂,z₂) and (x₁₁,y₁₁,z₁₁), respectively, the calculation of angle γ can be obtained by the angle relationship in mathematics, vector ${\vec{p}}_{9} p_{11} = (x_{11} - x_{9}, y_{11} - y_{9}, z_{11} - z_{9})$ , vector $\vec{p_{9} p_{2}} = (x_{2} - x_{9}, y_{2} - y_{9}, z_{2} - z_{9})$ , then angle γ is equal to: (24) $γ = \arccos (| \frac{\vec{p_{9} p_{11}} \cdot \vec{p_{9} p_{2}}}{| \vec{p_{9} p_{11}} | | \vec{p_{9} p_{2}} |} |)$

By this calculation method it is possible to derive the angular features of different joints, and according to the needs of action analysis can choose different (all or part of the important nodes of the angle) features to represent the action information of a certain frame.

2)

Spatial position difference of skeletal nodes

Due to the differences in human height, limb length, etc., the skeletal data obtained by using Kinect will be quite different. 20 skeletal joint points of the human body are captured by Kinect, and the construction of a reasonable coordinate normalization system can attenuate the influence of the above factors. Choosing the appropriate reference joints is the first thing to consider in the whole normalization system. In general, the reference joints are selected from the points that move less during the motion process, so the selection of reference joints can be different for different motion processes. The 20 joints are connected to form 19 joints. Assuming that p₄ is chosen as the reference joint, the remaining 19 points can form a spatial position difference centered on p₄, or called “virtual joints” centered on p₄ . Similarly, the spatial coordinates of 20 joints are (x,y,z), and the position difference with respect to the reference center is: (25) $Δ x_{i} = x_{i} - x_{4} 1 \leq i \leq 20$ (26) $Δ y_{j} = y_{j} - y_{4} 1 \leq j \leq 20$ (27) $Δ z_{k} = z_{k} - z_{4} 1 \leq k \leq 20$

Based on the characteristics of bone joint nodes, the different characteristics and their differences of human gestures can be fully expressed, and the traditional characteristics representation method has a complete advantage.

2.2.3

DTW-based comparison analysis algorithm

The DTW algorithm uses the idea of dynamic programming to combine the computation of time regularization and distance measure, which is a typical optimization problem. The regularization function can describe the similarity matching relationship between two data. Suppose two action data sequences are J and K, corresponding to lengths m and n respectively: (28) $J = {J_{1}, J_{2} \dots J_{i} \dots J_{m}}$ (29) $K = {K_{1}, K_{2} \dots K_{j} \dots K_{n}}$

In the field of action recognition one of the action sequences can be used as a template sequence, and the action comparison analysis can be viewed as a standard action sequence. Sequences J and K have m and n frames of action, respectively, and J_m and K_n are the feature representations in each frame. m and n can be equal or unequal in length, and in the case of inequality, calculating the shortest distance between the two sequences requires the creation of a n×m matrix to align the two sequences, and the distance between points in the matrix is denoted as d(J_i,K_j) = (J_i–K_j)². The next step is to search for the shortest path, and the regularized path P defines the mapping relation between the sequences J and K, if the l th element in P is P_l = (i,j)_l: (30) $P = {P_{1}, P_{2} \dots P_{l} \dots P_{L}} \max (m, n) \leq L < m + n - 1$

The regularized path P needs to satisfy the following three constraints: 1)

Boundary condition constraints, the start and end points of P correspond to P₁ =(1,1) and P_L = (m, n) respectively, i.e., the start and end points of the regular path correspond to the two end points on the diagonal of the matrix.

2)

Continuity constraints, P_I = (a,b), P_I+1= (a′,b′) is satisfied by a′–a ≤ 1, b′–b ≤ 1, i.e., the paths are found one after another until all the points are found at the end.

3)

The monotonicity condition constraint, P_i =(a,b), P_i+1 =(a′,b′) also satisfies a′–a ≥0, b′–b ≥ 0, i.e., the regularized path is monotonically increasing in time.

There may be several paths from the start point till the end point, but the aim here is to find the shortest path. DTW path distance is: (31) $D T W (x, y) = \min {\sqrt{\sum_{k = 1}^{K} P_{k} / K}}$

Satisfying the above conditions, there are only three ways to regularize the path P_i to the next point, incrementing to the right in the matrix, incrementing upwards and along the diagonal. The final cumulative distance is: (32) $β (i, j) = \min [\begin{matrix} D (i, j - 1) \\ D (i - 1, j) \\ D (i - 1, j - 1) \end{matrix}] + d (J_{i}, K_{j})$

2.2.4

Improved LCSS based comparison analysis algorithm

The LCSS algorithm can also be used as a similarity measure function that allows time series to vary in length and cope well with data noise.The LCSS algorithm is to find the longest of all subsequences, which are derived from a number of data in the original sequence.The longest common subsequence of two sequences is denoted as 4. If the last number in the initial sequence is 5, then 6, 7 is the LCS8 of 8 and 9. Suppose two sequences are A = (a₁,a₂⋯a_i.…a_x) and B = {b₁,b₂…b_j…b_y}, and their longest common subsequence LCS(A,B) is denoted as C = {c₁,c₂…c_k}. If the last number in the initial sequence is a_x= b_y, then c_k = a_x = b_y and c_k–1 are the LCSs of b_y–1 and a_x–1, and if a_x ≠ b_y, then c_k is either the LCS of a_x and b_y–1 or the LCS of a_x–1 and b_y. Eventually, solving for LCS(A, B) turns out to be a two-subproblem problem solved by recursion, which is inefficient, and is usually done using DP’s method. method. Using a two-dimensional array d[i][j]to hold the lengths of the common subsequences of the two sequences A, B respectively, the recursive transfer equation is: (33) $d [i, j] = {\begin{array}{l} 0 & i = 0 O r j = 0 \\ d [i - 1, j - 1] + 1 & i, j > 0 A n d a_{i} = b_{j} \\ \max (d [i, j - 1], d [i - 1, j]) & i, j > 0 A n d a_{i} \neq b_{j} \end{array}$

It has been mentioned in the previous paper that the similarity metric of action is fuzzy matching, after the feature representation with LCSS algorithm can not directly get the similarity of two action sequences, here we still need to introduce the concept of distance metric to determine whether the two action frames data are similar or not. The key frames (poses) of the skeletal data are obtained after preprocessing of the skeletal data in the previous section, and in this paper, least squares is introduced as the distance metric here. The purpose is to determine whether the key frames are similar after preprocessing, assuming that the two frames of skeletal data are G = {g₁, g₂…g_i…g_m} and H = {h₁, h₂…h_i…h_m}, if the value of m is 20 for the skeletal coordinate point features, the data in G and H correspond to the joints and the value of (x,y,z) is a matrix with 3 rows and 1 column, and the correspondence between G and H is: (34) $h_{i} = P g_{i} + Q + n_{i} 1 \leq i \leq m$

P refers to a rotation matrix with 3 rows and 3 columns, Q is a translation matrix with 3 rows and 1 column, and P and Q can be calculated using the following method.

1)

Calculate the mean g′, h′ of g and h, which yields G_i = g_i–g′, H_i = h_i–h′, i.e. {G_i,}{H_i}

2)

Compute the matrix of 3 × 3 for $L = \sum_{i = 1}^{m} G_{i}^{T} H_{i}$ , and $G_{i}^{T}$ for the transpose of G_i.

3)

Next is the singular value decomposition, L = JΛK^T, which is computed to give Y = KJ^T.

4)

Compute the determinant det(Y) of Y, and rotate the matrices $\overset{Λ}{P}$ and Y as: (35) $\overset{Λ}{P} = {\begin{array}{l} Y = K J^{T} & \det (Y) = 1 \\ K' J^{T} & \det (Y) = - 1, O r d e r K' = (K_{1}, K_{2}, - K_{3}) \end{array}$

5)

Translation vector, $\hat{Q} = g_{i} - P^{T} h_{i}$ is finally calculated: (36) $\sum_{i = 1}^{m} {‖ n_{i} ‖}^{2} = \sum_{i = 1}^{m} {‖ h_{i} - (P g_{i} + Q) ‖}^{2}$

In this paper, the value of ∑║n_i║² is used to measure the similarity of the corresponding keyframes, and by calculating the value of the metric in the corresponding keyframes, the one with the smallest value is considered as the keyframe with high similarity. Then the similar keyframes in the above LCSS algorithm are used as one of the elements in the common subsequence, and the final evaluation of the similarity of the two action sequences is made.

2.3

Digital Dance Teaching Platform Design

The Digital Dance Teaching Platform, as the name suggests, is designed to provide users with a platform where dance teaching can be accomplished without the need for on-site guidance from teachers. The following are the design objectives for the platform functions.

1)

Introduction of basic knowledge. Including background knowledge of dance, types of dance, history of development and current situation, etc., which is presented in the form of text for users to preview the relevant knowledge of dance. Obtained from the local area where the dance originated.

2)

Video playback demonstration. Dancing by the dance inheritor, we record and edit the dance teaching video, so that users can watch the standard dance teaching video and learn.

3)

Dance practice mode. First, we use optical motion capture to record the standard movement data stream of the inheritor, and then we incorporate the data movement into the character model we built as the teacher avatar. Users can imitate the movements of the teacher’s avatar by watching them, and the system will provide instant scores and total scores to verify the standardization of the movements.

The functions of the digital dance teaching platform are mainly divided into three parts: basic knowledge introduction, video playback demonstration, and dance practice mode. The purpose of basic knowledge introduction is to provide users with dance knowledge through text and illustrations. The function of the video playback demonstration part is similar to the conventional video playback software on the market, with progress bar control playback, volume control, multiplier playback, replay, pause and other conventional functions, the user can control the playback of the dance video according to their own needs. The flow of the dance practice module is shown in Figure 1, users need to follow this flow to complete a dance practice, and finally the system will also give users total feedback to measure the learning effect.

It is important to point out that the digital dance teaching platform in practice does not completely replace traditional offline dance teaching. Combining the two methods will be a better choice.

3

Digital Dance Teaching Platform Application

3.1

Dance Teaching Platform Performance Analysis

Selected Latin dance videos were used to construct a database of standardized movements and a dance trainer was invited to learn the dance, and the information was collected and processed in real time through Kinect and computer.

3.1.1

Coordinate-assisted analysis of dance movement joints

Taking the right wrist joint as an example, a comparison of the joint coordinates between the dance trainee and those in the standard movement database is shown in Table 1. In the first left wrist raising movement, the vertical coordinate of the left elbow was -109, while the vertical coordinate of the joints in the standard dance was -53.3, the height of the raising did not meet the requirements of the standard dance movements, and after the wrist was raised to the highest point, the wrist was withdrawn too quickly, and the vertical coordinate became -181.5, which is a big difference from the standard movement with the vertical coordinate of -62.2. In the second right wrist raising action, the raising speed was too fast, and the longitudinal coordinate was 372.1, which was much higher than the longitudinal coordinate of the standard dance joint point, and once again, there was an obvious action difference. Based on the above analysis, the auxiliary teaching based on the change of joint coordinates can intuitively detect the difference between the training movements and the standard dance movements, which can basically meet the auxiliary teaching requirements of dance training.

Table 1.

Auxiliary teaching experiment based on joint coordinates

Body parts	Node coordinates of standard dance		Node coordinates of trainer
Body parts	X-coordinate	Y-coordinate	X-coordinate	Y-coordinate
Head	-80.9	216.7	-70.4	267.9
Neck	-24.8	30.1	-68.8	73.6
Left shoulder	-150.5	-49.4	-210.7	-2.6
Left elbow	-377.2	-53.3	-408.5	-109
Left hand	-652	-62.2	-636	-181.5
Right hand	95.9	-8.1	308.7	372.1
Right elbow	79.5	222	235.9	154.1
Right shoulder	47.2	507.2	90.2	-1.6
Right knee	-108.3	-866.1	-128.1	-8292.7
Right foot	-68.3	-1160.5	-162.6	-995.5
Left knee	-237.8	-938	-24.9	-857.8
Left foot	-314.3	-1170.3	-13.8	-1000.9

3.1.2

Dance Movement Joint Angle Auxiliary Analysis

The joint angle 1 composed of the left shoulder, left elbow, and left wrist, and the joint angle 2 composed of the right hip, right knee, and right foot are selected as the observation targets to test the joint angle-based assisted teaching effect of the system proposed in this paper, and the comparison of the training effect is specifically shown in Fig. 2. From Fig. a, it can be seen that within 0-100 frames, the trajectory of the trainer’s joint angle 1 change is roughly consistent with the change of the position of the joint angle 1 of the sample library, but there will be no more than 12° error occurs. From Fig. b, it can also be found that within 0-100 frames, the trajectory of the trainer’s joint angle 2 change is also generally consistent with the change of the joint angle 2 position of the sample library, but near 76 frames, the system detects a large movement error of the trainer. It can be seen that the system proposed in this paper can accurately detect the training movements of the dance trainer, and can also assist the dance trainer to correct his/her movements by comparing with the sample library. This shows that the dance movement detection system designed in this paper can help dance trainers monitor the standardization of their movements during independent training.

3.2

Analysis of teaching effect of dance teaching platform

The teaching object selected in this chapter is the 2023 dance students of a sports college in Wuhan City, Hubei Province, setting up an experimental class and a control class, with the experimental class utilizing the digital dance teaching platform constructed in this paper to carry out dance teaching, while the experimental class still maintains the traditional dance teaching method. The teaching practice was conducted from September 2023 to November 2023, and Latin dance was taught for 12 weeks with 24 class hours. At the end of the experiment, tests and surveys were conducted on the students of the experimental class and the control class in terms of dance teaching performance, independent learning ability and course satisfaction, and interest in dance learning.

3.2.1

Comparative analysis of dance teaching performance

After 12 weeks of dance teaching practice, the dance teaching performance of the experimental class and the control class is shown in Table 2. As we know from the table, the experimental class’s learning and choreography combination scores in the skill mastery module were 89.23 and 83.28, respectively, which were 6.53 and 3.25 higher than those of the control class. p=0.008<0.01 for the learning combination score of the experimental class and the control class, which showed a highly significant difference, and p=0.047<0.05 for the choreography combination score, which showed a significant difference. In the theory learning module, the average theory score of the experimental class was 84.27, while the average score of the control class was 78.8, P=0.038<0.05, showing a significant difference. Obviously, the use of the digital dance teaching platform constructed in this paper is conducive to improving students’ dance skills and effectively deepening their knowledge of the theoretical knowledge of dance in this specialty.

Table 2.

Dance teaching achievement

Module	Grade	Class	Mean value	T	P
Mastery of skills	Learning team synthesis	Experimental class	89.23	1.453	0.008**
	Learning team synthesis	Control class	82.7	1.453	0.008**
	Choreographic grade	Experimental class	83.28	-1.577	0.047*
	Choreographic grade	Control class	80.03	-1.577	0.047*
Theoretical learning	Theoretical achievement	Experimental class	84.27	1.057	0.038*
Theoretical learning	Theoretical achievement	Control class	78.8	1.057	0.038*

3.2.2

Comparative analysis of independent learning ability and course satisfaction

The performance of the experimental and control classes in terms of independent learning ability and course satisfaction is specifically shown in Table 3. Self-directed learning ability consists of three dimensions: motivation to learn, learning strategies, and ability to learn. In the dimension of learning motivation and learning ability, the T-value of the experimental class and the control class is -3.316 and -2.484 respectively, and the P-value is 0.002, 0.003 and below 0.01, which is a highly significant difference. In the learning strategy dimension, the mean value of the experimental class is 0.22 higher than the control class, with a P value of 0.035 and below 0.05, and there is a significant difference. In the course satisfaction dimension, there is a significant difference between the experimental class and the control class in terms of course implementation and course outcome satisfaction P-values of 0.037 and 0.048 respectively, which are below 0.05. The mean score of overall satisfaction of the experimental class is 3.8, while the control class is 3.38, and the difference between the two sides is 0.42, with a P value of 0.006 and below 0.01, which is a highly significant difference. Overall, using the dance teaching platform is conducive to the improvement of students’ independent learning abilities, and it also improves the students’ satisfaction with the course.

Table 3.

Master learning ability and course satisfaction

Module	Dimension	Mean value		T	P
Module	Dimension	Control class	Experimental class	T	P
Autonomous learning ability	Learning motivation	3.53	3.84	-3.316	0.002**
	Learning strategy	3.61	3.83	1.262	0.035*
	Learning ability	3.57	3.82	-2.484	0.003**
Class satisfaction	Curriculum implementation	3.45	3.77	-3.494	0.037*
	Course technology	3.56	3.83	-0.494	0.048*
	Overall course satisfaction	3.38	3.8	-1.494	0.006**

3.2.3

Comparative analysis of dance learning interest analysis

The dance learning interests of the experimental and control classes are shown in Table 4. The movement participation score of the students in the experimental class was 36.26, and the mean value of the movement participation of the students in the control class was 29.74, and the experimental class was higher than the control class by 6.51, P=0.002<0.01, which indicated that there was a highly significant difference between the movement participation of the experimental class and the control class after the experiment. In terms of positive interest in learning, the mean value of students in the experimental class was 28.43, which was 2.78 higher than the control class, P=0.042<0.05, and there was a significant difference between the two classes. As for the dimension of negative interest in learning, the reverse scoring method was used, the lower the negative interest the higher the mean value. The mean value of negative interest in learning of the students in the experimental class is 25.72, and the mean value of negative interest in learning of the students in the control class is 22.91, which is 2.81 higher than that of the control class, P=0.065>0.05, which means that there is no significant difference in the scores of negative interest in learning of the experimental class students and the control class students. In terms of the degree of independent learning, the mean value of independent learning of the control class is 17.34, and the experimental class is 2.64 higher than it, P=0.014<0.05, indicating that there is a significant difference between the degree of independent learning of the experimental class and the control class. To summarize, the overall learning interest of the experimental class has an overall learning interest mean value of 110.39, and the experimental class is higher than the control class by 14.64, showing a highly significant difference. Using the digital dance teaching platform constructed in this paper, students’ interest in learning dance in the dance course becomes stronger.

Table 4.

Dance learning interest

Dimension	Mean value		T	P
Dimension	Experimental class	Control class	T	P
Sports participation	36.26	29.74	-2.741	0.002**
Study positive interest	28.43	25.65	3.744	0.042*
Study negative interest	25.72	22.91	2.756	0.065**
Autonomous learning degree	19.98	17.34	1.571	0.014*
Overall learning interest	110.39	95.65	-2.454	0.001**

4

Conclusion

Based on Kinect, this paper proposes human movement recognition technology and human movement comparison technology, and further designs and builds a digital dance teaching platform. From the perspective of platform performance and teaching effectiveness, the application practice of digital dance teaching platforms is carried out. The study reveals the following conclusions: 1)

in the digital dance teaching platform performance analysis, respectively, based on the joint coordinates, joint angle of the auxiliary teaching experiments, combined with the platform of dance teaching can intuitively perceive the differences between the training movements and the standard dance movements, can help dance trainers in the independent training, supervise the normality of their movements, basically meet the auxiliary teaching requirements of dance training.

2)

In terms of dance teaching performance in the analysis of teaching effect, the learning combination performance and choreography combination performance of the experimental class were 89.23 and 83.28, which were 6.53 and 3.25 higher than that of the control class, respectively.The average theoretical score of the experimental class was 84.27, while that of the control class was 78.8.The P-value of the experimental class and the control class in the various performance modules was less than 0.05, which showed a significant difference. The digital dance teaching platform constructed in this paper can effectively improve students’ dance skills and deepen their knowledge of dance theory.

3)

In terms of independent learning ability, the P-values of the experimental class and the control class in the dimensions of motivation and learning ability are 0.002, 0.003 and below 0.01 respectively, and there is a highly significant difference. The p-value in the learning strategy dimension is 0.035 and below 0.05, and there is a significant difference. In the course satisfaction dimension, the experimental class is 0.42 higher than the control class with a p- value of 0.006 and below 0.01, which is a highly significant difference. The digital dance teaching platform has a positive impact on students’ independent learning ability and satisfaction with dance courses.

4)

In terms of dance learning interest, the overall learning interest of the experimental class has an overall learning interest mean value of 110.39, and the experimental class is higher than the control class by 14.64, showing a highly significant difference. Under the dance teaching with the digital dance teaching platform, students’ interest in dance learning becomes stronger.

Overall, the digital dance teaching platform can play an important and effective teaching aid for dance teaching, and can play a positive role in promoting dance education for users.

Idioma:: Inglés

Calendario de la edición:: 1 veces al año
Temas de la revista:: Ciencias de la vida, Ciencias de la vida, otros, Matemáticas, Matemáticas aplicadas, Matemáticas generales, Física, Física, otros

RSS Feed de revista

Design and Implementation of Digital Dance Teaching Platform Based on Kinect

Yiheng Li

Publicado en línea: 17 mar 2025

Recibido: 02 nov 2024

Aceptado: 20 feb 2025

DOI: https://doi.org/10.2478/amns-2025-0253

Palabras claveKinect, Dance teaching platform, Human movement, Hidden Markov Model, Hausdorff distance, Longest common subsequence algorithm

© 2025 Yiheng Li, published by Sciendo

This work is licensed under the Creative Commons Attribution 4.0 International License.

Palabras clave
Kinect, Dance teaching platform, Human movement, Hidden Markov Model, Hausdorff distance, Longest common subsequence algorithm