General analysis method for isomeric myriapod locomotion coordination patterns

Abstract

This invention belongs to the field of biological movement behavior analysis, specifically involving a general analysis method for the movement coordination patterns of heterogeneous multipodial organisms. The method includes the following steps: graphical transformation of multivariate time series of multipodial movement; extraction of action features based on a microstructure visual dictionary; low-dimensional embedding of high-dimensional action features and clustering of movement behavior. The general analysis method for the movement coordination patterns of heterogeneous multipodial organisms proposed in this invention can not only convert the movement time series data of multipodial organisms with different numbers of limbs and different configurations into a unified representation, but also completely preserve the dynamic coordination relationship between limbs. It can achieve unified clustering and quantitative comparison of movement patterns across species, terrains, and injury states, breaking through the technical constraints of strong coupling between traditional methods and biological body configurations, and effectively providing core underlying technical support for the upgrade of multipodial robots from dedicated control to a generalized control paradigm.

Description

Technical Field

[0001] This invention belongs to the field of biological movement behavior analysis, specifically involving a general analysis method for the movement coordination patterns of heterogeneous multipodial organisms. Background Technology

[0002] Quantitative analysis of animal locomotion is a core foundation for revealing the mechanisms of biological locomotion control and promoting the development of biomimetic engineering technology. Research on the locomotion coordination patterns of myriapods is a crucial link between biological kinematics and biomimetic robot control. With the development of machine vision and deep learning technologies, high-precision pose estimation tools such as DeepLabCut and SLEAP have enabled accurate tracking of key points on animal limbs, providing data support for the quantitative analysis of locomotion. Meanwhile, unsupervised learning algorithms have become the core means of achieving automatic clustering and classification of locomotion.

[0003] Myriapods, with their flexible limb coordination, can achieve stable movement and damage compensation in complex environments. Their movement coordination patterns are important biological prototypes for the design of myriapod biomimetic robot control systems. However, current research on myriapod movement modalities still has significant gaps. Most existing work is limited to analysis under conditions of single species, specific limb configurations, and intact limbs. A systematic research method has not yet been developed for the mining and quantitative comparison of common movement patterns in heterogeneous myriapods, and it is difficult to accurately capture changes in movement coordination characteristics under different degrees of injury, such as leg amputation, making it difficult to explore their inherent movement fault-tolerant compensation mechanisms. This limitation directly restricts the transfer of biological movement patterns to the field of biomimetic control, causing myriapod robot development to be trapped in a dedicated control paradigm of "one machine, one model," making it difficult to achieve universal adaptation of a single model to multiple configurations.

[0004] Unsupervised behavioral pattern analysis is a core technique in the quantitative study of myriapod movement. Current mainstream methods include B-SOiD, Keypoint-MoSeq, VAME, and MotionMapper. While each algorithm has its own advantages in terms of functionality—B-SOiD is simple to operate and has high clustering consistency, Keypoint-MoSeq is robust and can capture complex, integrated movement sequences, and VAME can preserve the temporal characteristics of behavioral sequences through nonlinear dimensionality reduction—from a fundamental methodological perspective, these algorithms all rely on temporal data of limb keypoint movements as their core input. Their feature extraction, modal clustering, and behavior decoding processes are highly coupled with the biological body structure. The feature engineering and clustering logic of mainstream unsupervised algorithms rely on fixed limb dimensions and configuration designs, failing to construct dimensional adaptation and alignment mechanisms for heterogeneous myriapods. This makes them unable to handle scenarios with dimensional changes caused by differences in leg number and limb structure, becoming a core technical barrier in heterogeneous myriapod movement analysis.

[0005] In the field of high-dimensional motion data compression and pattern clustering, t-SNE and UMAP, as classic manifold learning methods, serve as core benchmarks for evaluating the effectiveness of low-dimensional embedding. SUDE, on the other hand, has improved upon manifold learning methods for mixed-distribution data, becoming a direct benchmark in this area. While these methods can effectively reduce dimensionality and cluster high-dimensional data, none of them address the needs of heterogeneous myriapod motion analysis by designing a unified motion representation paradigm independent of limb morphology. They lack a unified analytical logic across the entire chain from "time-series data - image transformation - feature extraction - manifold embedding," making them difficult to directly apply to the mining of heterogeneous myriapod motion coordination patterns and unable to provide end-to-end technical support for generalized motion feature extraction. Summary of the Invention

[0007] This invention addresses the limitation of existing biological locomotion behavior analyses, which are mostly confined to single species and specific configurations, failing to provide a unified analysis of locomotion across different species and limb morphologies, and thus hindering the revelation of the common underlying patterns of locomotion in heteropods. Therefore, this invention proposes a universal analytical method for the coordination patterns of locomotion in heteropods, with the following specific steps: Step 1: Graphical transformation of multivariate time series of myriapod locomotion; Step 2: Action feature extraction based on microstructure visual dictionary; Step 3: Low-dimensional embedding of high-dimensional action features and clustering of motion behavior; Furthermore, the specific steps of step one include: (1) High-speed cameras were used to capture the movement of myriapods. Two-dimensional coordinate information of the animal's foot was accurately extracted using key point detection technology to construct a multivariate time series describing the dynamic evolution of the animal's limb position. The sequence was segmented by a fixed-length sliding window, with the window moving by half the length of the time window each time. This divided a long multivariate time series into multiple overlapping short time series windows. The coordinate data was then normalized to eliminate interference from individual scale differences and shooting distance, resulting in normalized motion time series data. (2) Calculate the cosine similarity of the multidimensional state vectors at different times within the same time window, and construct and normalize the binary recursive matrix using the step function to transform the multivariate time series into a recursive matrix: .in, , A multidimensional state vector representing time i and time j within the same time window; (3) Logarithmic transformation is used to enhance and normalize the recursive matrix, avoiding logarithmic singularity problems and amplifying subtle distance differences to strengthen texture detail features. The logarithmic transformation formula is: .in, The log-enhanced matrix is ​​transformed to the grayscale range of [0, 255] through linear mapping to generate a recursive coordination pattern grayscale image, thereby realizing a graphical representation of the movement behavior of myriapods.

[0008] Furthermore, the specific steps of step two include: (1) Using the grayscale recursive image as input, a multi-scale Gaussian pyramid scale space is constructed to detect its microstructure. A Gaussian difference scale space is constructed by pixel-by-pixel difference between adjacent layers of the pyramid, and the key points are initially screened by comparing the extreme values ​​of the three-dimensional neighborhood in the scale space. Then, low-contrast unstable points are removed by using the grayscale contrast response threshold, and edge-sensitive points are removed by using the edge response criterion, so as to achieve accurate positioning and purification of effective key points.

[0009] (2) Centered on the effective keypoints, based on the gradient magnitude and direction information of the neighboring pixels of the corresponding scale Gaussian image, the gradient direction is quantized into multiple equally spaced intervals to construct a gradient direction histogram. The maximum value is selected as the main direction of the keypoint to ensure the rotation invariance of the feature. Finally, a local feature descriptor describing the recursive microstructure is constructed along the main direction.

[0010] (3) For each sample in the feature descriptor subset, calculate the sample's... The local importance of a sample is quantified by the total number of its nearest neighbor samples. Its local importance is expressed by "KNN containment". The total number of samples is used as the metric, that is, it satisfies: ;in, Indicates sample of A set of nearest neighbors.

[0011] Based on the above idea, the m nodes with the highest RNN values ​​are selected as microstructure landmarks, and the neighborhood parameters are initialized. minimum value and maximum value The optimal value is determined by reverse iterative search using binary search. The value is set such that the number of sampled landmarks meets the requirement of m. This ensures both the uniform distribution of landmarks in each cluster and the preservation of the cluster structure integrity and global distribution consistency of the original data to the greatest extent, laying the foundation for the accuracy and robustness of subsequent low-dimensional embedding.

[0012] After selecting m landmark points, each microstructure landmark is treated as a visual category. Based on the Euclidean distance between the non-landmark feature descriptor and its N nearest neighbor landmark feature descriptors, the non-landmark feature weights are probability-assigned to the corresponding landmark points. The weight assignment formula for landmarks is as follows: In the formula, For non-landmark feature descriptors, Let N be the p-th landmark feature descriptor, and N be the set of the nearest non-landmark points. Based on the calculated landmark weights, these values ​​are accumulated and normalized to form the action feature vector corresponding to each recursive graph. The calculation formula is as follows: In the formula, Describing the 1st dimension in the reduced space v The standard basis column vector, which is only the first standard basis column vector. v The first row has a value of 1, and all other rows have a value of 0, which is used to represent the dimensionality reduction result. v There are 1 dimension, where n is the number of all microstructure feature descriptors in a single recursive graph.

[0013] Furthermore, the specific steps of step three include: (1) For high-dimensional action features from different data sources, this invention employs multiple sampling strategies. For homogeneous biological action features, a global sampling strategy is adopted; for heterogeneous biological action features, a group sampling strategy is adopted, which divides the action feature vector into several subsets according to the data source, thereby improving the structure preservation effect of mixed distribution data.

[0014] (2) Filter out by calculating the RNN value of each high-dimensional action feature vector. The high-dimensional action feature landmark vectors are obtained. For each non-landmark vector, the vector with the closest Euclidean distance is selected. Given several landmark vectors, with the core objective of minimizing the local linear reconstruction error of non-landmark vectors in high-dimensional space, an optimization objective function for local linear embedding is constructed: ,in The linear reconstruction weight vector is a non-landmark vector. For the non-landmark vector to be embedded, for The set of nearest neighbor landmarks. Construct the Gram matrix using this objective function: By inverting the Gram matrix and combining it with the constraint that the sum of the weights is 1, the linear reconstruction weights can be obtained: Where 1 represents a column vector with all values ​​equal to 1. Finally, the initial low-dimensional embedding coordinates of the non-landmark vectors are calculated by weighted summation of the weights and the low-dimensional coordinates of the landmark points. The calculation formula is as follows: .

[0015] (3) To avoid non-landmark high-dimensional action feature vectors falling into inter-cluster gaps during low-dimensional embedding, each non-landmark vector is calculated. Its nearest landmark vector Euclidean distance in high-dimensional space And extract the vector of the nearest landmark. The pairwise distances between each nearest neighbor landmark vector in both the high-dimensional and low-dimensional spaces. and To minimize the mapping error between high- and low-dimensional pairwise distances, a least-squares optimization function is constructed: In the formula Let be the total number of pairwise distances, and 'scale' be the optimal distance scaling factor for the non-landmark vectors. The scaling factor is obtained by differentiation, and a high- and low-dimensional distance mapping is established. The constraint distance that the non-landmark vectors must satisfy in the low-dimensional space is then calculated. : .

[0016] (4) Construct a constrained optimization function with the objective of minimizing the deviation between the initial and final coordinates: , making ,in, '' represents the final low-dimensional coordinates of the non-landmark vector to be solved. The nearest landmark point in the low-dimensional space for this non-landmark vector. Low-dimensional coordinates. Introducing Lagrange multipliers. Constructing the Lagrangian function transforms the constrained optimization into an unconstrained optimization: ,right By taking the partial derivative and setting it to zero, and combining this with the constraints, we obtain the analytical solution for the final low-dimensional coordinates of the non-landmark vectors: This solution can strictly constrain non-landmark vectors to the neighborhood of the nearest landmark vector, preventing them from falling into the gaps between clusters and forming dirty clusters or outliers.

[0017] (5) Integrate the final low-dimensional coordinates of all landmark vectors and non-landmark vectors to generate a global low-dimensional embedding result that corresponds one-to-one with the original high-dimensional action feature vector. In the embedding process, the local neighborhood relationship and global topological structure of the high-dimensional data are fully preserved. Finally, a unified low-dimensional manifold space independent of the limb morphology of multi-legged animals is constructed to form a cluster of movement behaviors under different species and different limb morphologies.

[0018] The technical effects and advantages of this invention are as follows: 1) This invention breaks through the technical constraint of strong coupling between existing myriapod movement analysis methods and biological body configuration. The proposed universal analysis method for the movement coordination patterns of heterogeneous myriapods achieves unified analysis of heterogeneous myriapod movement data without configurational differences, and has significant technical advantages and application value: An innovative method for graphical representation of multivariate time series data. It transforms time series data with different numbers of legs and different limb configurations into a unified grayscale recursive map, completely removing morphologically relevant but irrelevant features, and realizing a unified projection of heterogeneous polypodary movement data into a coordination relationship space; 2) Low-dimensional embedding and clustering methods for heterogeneous multi-legged motion data. A group sampling strategy was designed to address the distribution characteristics of heterogeneous multi-legged motion data. Combined with a constrained local linear embedding algorithm, the topological structure of high- and low-dimensional spaces was effectively preserved, and a unified low-dimensional manifold space independent of morphology was constructed. This solved the key problem that traditional methods cannot directly quantify and compare heterogeneous multi-legged motion data. 3) A general analysis paradigm for biological motion coordination patterns was constructed, which can accurately uncover the universal motion control laws and fault-tolerance mechanisms of heterogeneous myriapods, providing core underlying technical support for the upgrade of myriapod robots from a dedicated control paradigm of "one machine, one model" to a generalized control paradigm of "single model adapting to multiple configurations".

[0019] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description and the drawings. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the general analysis method for the movement coordination pattern of heterogeneous multipodial organisms of the present invention. Figure 2 This is a low-dimensional embedding visualization of the motion coordination patterns of quadrupedal dogs walking on a flat surface and ants walking on different rough terrains, provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] In a specific embodiment of this invention, the general analysis method for heterogeneous multilegged animal movement coordination patterns uses six-legged ants and four-legged dogs as the core experimental research objects. The six-legged ants cover linear movement scenarios across six different terrain roughness levels, from low to high, while the four-legged dogs move linearly on a plane. This embodiment relies on a high-speed camera and key point detection technology. Through four core steps—collection and preprocessing of multilegged animal movement data, graphical representation of multivariate time series, construction of a visual bag-of-words model based on adaptive feature extraction, and low-dimensional embedding and clustering for mixed-distribution data—a unified analysis of heterogeneous multilegged animal movement coordination patterns is achieved, thereby uncovering the universal movement control laws and fault-tolerance mechanisms behind multilegged animals. The overall principle and flow of the method are as follows: Figure 1 As shown.

[0025] The first step is the graphical transformation of multivariate time series data on myriapod locomotion. The specific steps are as follows: 1) Multiple repeated experiments were conducted on ants moving in straight lines under terrains of varying roughness and quadruped dogs moving in straight lines in planar environments to collect motion data. High-speed cameras were used to record the entire movement process from a vertical overhead view, ensuring that the animals' limb movements were clearly visible in the footage. Using keypoint detection technology, two-dimensional coordinates of the feet relative to the body center were extracted to construct a raw multivariate time series representing dynamic changes in limb position. A fixed-length sliding window was used to segment the time series, with window parameters ensuring that each data window fully covered at least one typical movement pattern. Simultaneously, all foot coordinate data were normalized to eliminate systematic errors introduced by individual animal scale differences and shooting distance, ultimately obtaining standardized motion time series data, providing a unified and high-quality data foundation for subsequent graphical representation.

[0026] 2) In each time window, construct and normalize a binary recursive matrix using a step function: .in, , This represents the multidimensional state vectors at times i and j within the same time window. To address the issue of weak phase-coupled textures being easily masked in recursive graphs, a logarithmic transformation is applied to the constructed binary recursive matrix to enhance it. This avoids logarithmic singularities while amplifying subtle distance differences within the matrix, thus strengthening texture detail features. The logarithmic transformation formula is: ,in, The enhanced recursive matrix is ​​transformed to the standard grayscale range through linear mapping to generate a standard grayscale recursive map of fixed size, thus completing the effective transformation from time-series data to spatial texture features. Finally, a unified motion graphical representation dataset of heterogeneous multi-legged organisms, namely quadrupedal dogs and hexapods, is constructed.

[0027] The second step is action feature extraction based on a microstructured visual dictionary. The specific steps are as follows: 1) Using grayscale recursive images as the research object, a multi-scale Gaussian blur pyramid is constructed for each grayscale recursive image, with fixed scale interval coefficients and scale layers. The motion coordination features and microstructures at different scales are captured by calculating the standard deviation difference between adjacent scales. A fixed contrast threshold is used to filter low-contrast candidate keypoints. The scale and principal direction angle of the candidate keypoints are obtained by combining the central difference method and rotation invariance, generating feature descriptors describing the microstructure. Simultaneously, the pixel coordinates, scale, and principal direction angle of adjacent keypoints are compared to eliminate keypoints with completely duplicate geometric features, ensuring the uniqueness and representativeness of the extracted features. Finally, the microstructure feature descriptors of all grayscale recursive images are collected to construct a global feature descriptor set. 2) Introducing the reverse nearest neighbor concept to quantify the local importance of samples for adaptive sampling of microstructural landmarks. For each sample in the feature description subset, the local importance of that sample is statistically analyzed. The local importance of a sample is quantified by the total number of its nearest neighbor samples. Its local importance is expressed by "KNN containment". The total number of samples is used as the metric, that is, it satisfies: ;in, Indicates sample of A set of nearest neighbors; 3) Input the preset number of microstructure landmarks and initialize the maximum and minimum value range of the neighborhood parameters. Determine the optimal neighborhood parameter value through binary search and reverse iterative search. Then, sort the feature vectors in descending order of the reverse nearest neighbor value. Recursively select the samples with the highest ranking as microstructure landmarks and remove the nearest neighbor samples corresponding to the microstructure landmarks to avoid redundancy. Finally, obtain uniformly distributed microstructure landmarks as the core reference primitives of the visual bag of words. 4) Based on the sampled microstructure landmarks, a visual bag-of-words model is constructed. Using the feature descriptors of the microstructure landmarks as a benchmark, the Euclidean distance between non-landmark feature descriptors and all landmark feature descriptors is calculated. The N landmark points with the smallest distances are selected as matching objects, and weights are assigned based on the reciprocal of the distance. In the formula, For non-landmark feature descriptors, Let N be the feature descriptor for the landmark point, and let N be the set of the nearest landmark points to the non-landmark point.

[0028] 5) Construct a high-dimensional action feature vector for the recursive graph based on the normalized weights. The vector calculation formula is as follows: In the formula, Describing the 1st dimension in the reduced space v The standard basis column vector, which is only the first standard basis column vector. v The first row has a value of 1, and all other rows have a value of 0, which is used to represent the dimensionality reduction result. vThere are several dimensions. n represents the total number of feature elements in a single recursive graph. Finally, the feature vectors of all recursive graphs are sorted from highest to lowest according to the total feature weight of the landmarks. This feature vector can completely represent the global distribution pattern of leg coordination relationships in the recursive graph.

[0029] The third step involves low-dimensional embedding of high-dimensional action features and clustering of motion behaviors. The specific steps are as follows: 1) In this embodiment, the high-dimensional action feature vector set is divided into 7 subsets based on species type and terrain parameters of the movement scene. These subsets correspond to the feature data of quadrupedal dog planar linear motion and ant linear motion under 6 different roughness terrains. For the feature vectors within each subset, the reverse nearest neighbor sampling method is used to obtain the corresponding landmark vector set for each group. The landmark vectors of the 7 subsets are uniformly mapped to the same low-dimensional space, and then low-dimensional embedding processing is performed on all remaining non-landmark vectors based on the constrained local linear embedding algorithm. By introducing the nearest neighbor distance constraint to construct the optimization objective function, the problem of non-landmark vectors falling into the inter-cluster gap is effectively avoided, ensuring the clustering discriminativeness of features in the low-dimensional space.

[0030] 2) Based on the specified nearest neighbor pair distances of the landmark vector set in high-dimensional and low-dimensional spaces, the optimal distance scaling factor corresponding to each landmark point is solved by the least squares method to establish the distance mapping relationship between high-dimensional and low-dimensional spaces. For each non-landmark feature vector, its nearest several landmark points are selected, and the reconstruction weights are solved by local linear embedding. Then, the two-dimensional low-dimensional coordinates of the non-landmark vectors are solved by the Lagrange multiplier method. Finally, the two-dimensional low-dimensional coordinates of the landmark vectors and non-landmark vectors are integrated to generate a global low-dimensional embedding result that corresponds one-to-one with the original high-dimensional visual bag-of-words feature vectors, thus constructing a unified low-dimensional manifold space that is independent of shape.

[0031] The low-dimensional embedding visualization results of the motion coordination pattern representation obtained in this experiment are as follows: Figure 2As shown in the diagram, labels 0 to 1 correspond to the low-dimensional representations of the six-legged ant's linear motion in scenarios with increasing roughness, while the "dog" label corresponds to the low-dimensional representations of the four-legged dog's planar linear motion. Visual clustering results show that the vast majority of the four-legged dog's low-dimensional motion representations cluster independently, while a small number overlap with the ant's motion representations in low-roughness terrain. The ant's low-dimensional motion representations change with the gradient of terrain roughness from low to high, and can be roughly divided into three spatially differentiated clusters. These results indicate that the four-legged dog's planar linear motion and the ant's motion coordination mode in low-roughness terrain share a certain degree of similarity. Furthermore, the ant's motion coordination behavior mode exhibits three statistically significant differentiating behavioral modes as the surface roughness increases, and the transitions between modes show a significant correspondence with the gradient change in surface roughness. The above experimental results fully verify the technical effectiveness and advancement of the general analysis method for heterogeneous myriapod movement coordination patterns proposed in this invention. This method breaks through the technical constraints of strong coupling between existing myriapod movement analysis methods and biological body configuration, and realizes unified analysis of heterogeneous myriapod movement data without configurational differences. It can effectively explore the universal movement control laws and fault-tolerant compensation mechanisms of heterogeneous myriapods, and provide core underlying technical support for upgrading myriapod robots from a dedicated control paradigm of "one machine, one model" to a generalized control paradigm of "single model adapting to multiple configurations".

[0032] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A general analytical method for the movement coordination patterns of heterogeneous multipodial organisms, characterized in that, Includes the following steps: The first step is the graphical transformation of multivariate time series of myriapod locomotion; High-speed cameras were used to capture the movement of myriapods in a target scene. Relative coordinate information of the myriapod's feet was extracted through keypoint detection to construct a multivariate time series representing limb positions. Moving average filtering and linear interpolation were used to process the original time series, and the coordinate data was normalized to obtain standardized multivariate time series data. A sliding window was used to extract the standardized time series, moving half the time window length at each step. Within the same window, the cosine similarity of the multidimensional state vectors of the feet at different times was calculated to construct a binary recursive matrix. And a two-dimensional grayscale recursive map representing the leg coordination relationship is generated through linear mapping; The second step is action feature extraction based on a microstructure visual dictionary; Keypoint detection technology is applied to the grayscale recursive image. A Gaussian pyramid is constructed to filter unique and representative candidate keypoints and generate local feature vectors. The local density is represented by the inverse nearest neighbor value, and keypoints are recursively selected and removed. Nearest neighbor samples are used until a specified number of m microstructure landmarks are obtained; using the feature descriptors of the microstructure landmarks as a benchmark, the Euclidean distance between the non-landmark feature descriptors and the landmark descriptors is calculated, and the non-landmark feature weights are distributed to the top neighbors by normalizing the inverse of the distance. Based on the landmarks, an action feature vector based on a microstructure visual dictionary is finally constructed. The third step is the low-dimensional embedding of high-dimensional action features and the clustering of motion behaviors. Based on the data source, a global landmark point set is obtained by performing secondary landmark sampling on the visual bag-of-words histogram feature vectors. An objective function is constructed by introducing nearest neighbor distance constraints based on constrained local linear embedding, and non-landmark points are embedded. The optimal distance scaling factor of landmark points in high-dimensional and low-dimensional spaces is solved using the least squares method to establish a distance mapping relationship between high-dimensional and low-dimensional spaces, and the nearest neighbors of non-landmark points are selected. The system uses local linear embedding to solve for the reconstruction weights of each landmark point, and the Lagrange multiplier method to solve for the low-dimensional coordinates of non-landmark points. By integrating the low-dimensional coordinates of landmarks and non-landmarks, the system obtains the global low-dimensional embedding result, thereby realizing the low-dimensional embedding of high-dimensional motion features of myriapod movement coordination patterns and the clustering analysis of movement behavior.

2. The general analytical method for the movement coordination patterns of heterogeneous multipodial organisms as described in claim 1, characterized in that, The first step involves constructing a graphical representation dataset of multivariate time series of myriapod locomotion. This is achieved by using a sliding window to extract a standardized time series of length m, moving half the time window length at a time. Within the same window, the multidimensional state vectors of the foot at times i and j are calculated. , The cosine similarity is used to construct a binary recursive matrix through a step function. Then normalize, and then use logarithmic transformation on Enhancement and normalization processes were performed to obtain Finally A two-dimensional grayscale recursive map representing leg coordination is generated by linearly mapping the grayscale values ​​to the [0, 255] range; the normalization formula for the binary recursive matrix is ​​as follows: The formula for logarithmic transformation is: ,in, .

3. The general analytical method for the movement coordination patterns of heterogeneous multipodial organisms as described in claim 1, characterized in that, In the second step, when acquiring microstructure landmarks, input the preset number of target landmarks m and initialize the neighborhood parameters. minimum value and maximum value The optimal value is determined by reverse iterative search using binary search. The value ensures that the number of sampled landmarks meets the requirement of m; when assigning feature weights to non-landmark points, the weighting formula for the p-th landmark point is: In the formula, For non-landmark feature descriptors, Here, N is the set of landmark feature descriptors, where N is the set of landmarks closest to the non-landmark point; the formula for calculating the feature vector of the recursive graph is: In the formula, Describes the first dimension in the reduced space v The standard basis column vector, which is only the first standard basis column vector. v The first row has a value of 1, and all other rows have a value of 0, which is used to represent the dimensionality reduction result. v There are 1 dimension; n is the total number of features in a single recursive graph.

4. The general analytical method for the movement coordination patterns of heterogeneous multipodial organisms as described in claim 1, characterized in that, The data sources described in the third step combine global sampling and group sampling strategies. Global sampling can be used for the analysis of homogeneous polypodary movement data, while group sampling is used for the analysis of heterogeneous polypodary movement data to improve the structure preservation effect of mixed distribution data.

Images