CN-121999323-A - Intelligent evaluation method and system for quality of pattern skating action based on process guidance

CN121999323ACN 121999323 ACN121999323 ACN 121999323ACN-121999323-A

Abstract

The invention discloses a process-oriented intelligent evaluation method and system for quality of pattern skating actions, wherein the method comprises the steps of obtaining data, preprocessing, integrating and constructing a pattern skating action process data set, extracting video features, obtaining feature data, outputting specific start-stop time of each sub-action through a constructed sub-action start-stop recognition module, dividing continuous action paragraphs in video according to the start-stop time to obtain a plurality of continuous sub-action video fragments, inputting the divided sub-action video fragments into a trained process quality evaluation module, outputting a scoring result of each sub-action, summarizing and integrating the scoring result of each sub-action based on official labeling and model reasoning, and obtaining a final evaluation result. The method not only improves the accuracy and interpretation of the action quality score, but also provides scientific technical support for actual training and competition, and promotes the application and development of intelligent evaluation technology of pattern skating.

Inventors

ZHAO BIN
LI LIANGWEI
LI CHAO
YIN SIYUAN
Ao Cancan

Assignees

南京师范大学

Dates

Publication Date: 20260508
Application Date: 20260211

Claims (10)

1. The intelligent evaluation method for the quality of the pattern skating action based on the process guidance is characterized by comprising the following steps: S1, acquiring video data of a pattern skating match, preprocessing the data, and integrating and constructing a pattern skating action process data set; s2, extracting video features aiming at the pattern skating action process data set to obtain feature data; S3, outputting specific start-stop time of each sub-action through the constructed sub-action start-stop recognition module according to the characteristic data, and dividing continuous action paragraphs in the video according to the start-stop time to obtain a plurality of continuous sub-action video clips; S4, inputting the divided sub-action video clips into a trained process quality evaluation module, and outputting a grading result of each sub-action; and S5, summarizing and integrating the scoring result of each sub-action based on official labeling and model reasoning to obtain a final evaluation result.
2. The intelligent evaluation method for quality of pattern skating action based on process guidance according to claim 1, wherein the step S2 is characterized in that a VST model is adopted for video feature extraction, and the method specifically comprises the steps of extracting feature sequences from non-overlapping video segments, each video segment consists of a plurality of continuous frames, adopting a two-layer multi-layer perceptron to reduce dimension of features extracted from a main network, and comprising the steps of The feature sequences obtained for video of the individual segments are expressed as: Wherein 。
3. The intelligent evaluation method for the quality of the pattern skating action based on the process guidance is characterized in that a sub-action start-stop recognition module in the step S3 comprises an extensible granularity sensing layer and a shared weight detection head, the extensible granularity sensing layer is used for constructing a multi-scale pyramid feature representation, the detection head comprises a classification head and three regression heads, the classification head is responsible for determining whether a time slice contains sub-actions or not and ensuring the detection accuracy and recall rate, and the three-bin boundary regression head is used for jointly estimating the initial boundary, the end boundary and the center offset of the sub-actions based on relative distribution learning.
4. The intelligent evaluation method for quality of pattern skating motion based on process guidance according to claim 3, wherein in the sub-motion start-stop recognition module of step S3, the extensible granularity sensing layer enhances the discrimination and multi-scale capturing capability of the time sequence features through a double-branch structure, and provides a foundation for multi-scale pyramid construction, and the core formula is as follows: ; Wherein, the The method comprises the steps of inputting original video time sequence characteristics; As an example-level enhancement factor, The global average feature is obtained through a full connection layer and an activation function and is used for strengthening the feature difference between the sub-actions and the non-sub-actions; For the full connection layer, carrying out information coding on the single time step characteristics; For the window-level weighting factor, Dynamically adjusting the contribution degrees of different window characteristics through one-dimensional convolution; 、 The window sizes are respectively And Is convolved with one dimension depth Granularity expansion factor) for capturing timing semantics at different time scales; Original characteristic information is reserved for residual connection, so that information loss in deep learning is avoided; to enhance the time sequence characteristics, the multi-scale pyramid characteristics are formed after the maximum pooling with the multi-round step length of 2.
5. The intelligent evaluation method for quality of pattern skating motion based on process guidance according to claim 4, wherein in the sub-motion start-stop recognition module in step S3, the joint estimation formula is: Set the time step of the feature pyramid under a certain scale as The number of neighborhood boxes for boundary prediction is Then: start boundary relative probability distribution and offset: ; ; end boundary relative probability distribution and offset: ; ; Wherein: boundary feature sequences of B neighborhoods on the left side and the right side of the time step t respectively; respectively starting and ending boundary constraint features of the central offset branch output; The relative probability distribution of the starting boundary and the ending boundary is obtained through Softmax normalization; The characteristic domain offset from the time step t to the start and end boundaries of the sub-actions is calculated by the mathematical expectation of probability distribution.
6. The intelligent evaluation method for quality of pattern skating based on process guidance according to claim 5, wherein in the sub-action start-stop recognition module of step S3, specific start-stop time of each sub-action is as follows: Based on the feature domain offset and the time resolution of the multi-scale pyramid, mapping the feature domain result into the actual time of the video, wherein the formula is as follows: ; ; Wherein, the A characteristic time step which is judged to contain sub-actions in the scale layer l; 、 the initial and final boundary feature domain offsets output in the scale layer l are respectively; is the temporal resolution factor of the scale layer i, , wherein, For the original video frame rate, Downsampling multiple for the layer; 、 the actual start and end times of the sub-actions in the video, respectively.
7. The intelligent evaluation method for quality of pattern skating based on process guidance according to claim 6, wherein the process quality evaluation module in step S4 comprises a time sequence fusion module, a scoring analysis module and a coupling integration module, the time sequence fusion module is used for performing time sequence fusion on sub-action video clips, the scoring analysis module is used for outputting coarse-granularity and fine-granularity action scores according to the characteristics after time sequence fusion, and the coupling integration module is used for fusing the coarse-granularity and fine-granularity action scores to obtain a final scoring result of each sub-action.
8. The intelligent evaluation method for the quality of the pattern skating motion based on the process guidance according to claim 7, wherein the scoring analysis module in the step S4 decomposes the features into two layers of coarse granularity and fine granularity, the coarse granularity features capture the overall performance and key technical point information of the motion, the fine granularity features focus on the detail and the execution quality of the motion, and the scoring prediction is carried out on the two types of features through a multi-layer perceptron and a fine granularity scoring module respectively to output the scoring of the motion with the coarse granularity and the fine granularity.
9. The intelligent evaluation method for the quality of the pattern skating action based on the process guidance according to claim 8, wherein the training of the process quality evaluation module in the step S4 divides the real score into two parts of coarse granularity and fine granularity for supervision, so that the progressive score learning from coarse to fine is realized, and the discrimination capability and the generalization performance of a score model are improved.
10. A process-oriented intelligent system for evaluating quality of a pattern skating action, which is characterized by being used for realizing the method of claim 1, and comprising the following steps: the preprocessing module is used for preprocessing the data of the video data of the pattern skating match, and integrating and constructing a pattern skating action process data set; The feature extraction module is used for extracting video features and acquiring feature data; The sub-action start-stop recognition module is used for outputting specific start-stop time of each sub-action; The division module is used for dividing continuous action paragraphs in the video according to the start-stop time to obtain a plurality of continuous sub-action video clips; The procedural quality evaluation module is used for outputting the grading result of each sub-action; And the evaluation module is used for summarizing and integrating the scoring results of each sub-action to obtain a final evaluation result.

Description

Intelligent evaluation method and system for quality of pattern skating action based on process guidance Technical Field The invention belongs to the field of data processing analysis, relates to a sports data analysis and intelligent evaluation technology, and particularly relates to a process-oriented intelligent evaluation method and system for quality of pattern skating actions. Background With the rapid development of digital camera technology and the wide popularization of social media platforms, the number and detail level of pattern skating videos are increased in a bursting manner. The event video is widely broadcast through each large broadcasting platform. In addition, more than one hundred video related to pattern skating is uploaded on the video platform during the season. The analysis of the figure skating sports video not only can assist automatic scoring, wonderful lens generation and video abstraction, but also can provide support for the statistical analysis of physical ability and action advantages of athletes, and further provides scientific basis for training. The figure skating is used as a competition item integrating artistic expression and technical difficulty, and the action quality of the figure skating directly influences the competition result and training effect of athletes. However, the traditional scoring mainly depends on subjective judgment of professional referees, has certain subjectivity and instability, and is difficult to realize objective, fair and real-time feedback of the score. The automatic evaluation of the action quality can not only improve the scientificity and transparency of the score, but also provide fine-granularity technical analysis for coaches, help athletes optimize actions and improve the athletic level. In recent years, action quality evaluation based on deep learning and computer vision technology is gradually rising, and the action quality evaluation becomes a research hotspot in the field of sports video analysis. Most related work has focused on motion recognition, video classification, or overall motion scoring, e.g., fis-V, etc., public data sets have driven research into automated scoring of patterned skating videos. The existing model mostly adopts a long-short time memory network LSTM, a convolutional neural network CNN and a self-attention mechanism to try to capture the time sequence characteristics and the spatial information of the motion. These techniques improve the accuracy and efficiency of automatic scoring to some extent. Despite significant progress, existing researches have the problems that the video time of the pattern skating match is long, key score actions are unevenly distributed, so that a common model is difficult to capture all important action fragments, tag-dependent professional scoring is that action scoring needs to depend on scores given by professional referees, tag data are high in acquisition cost and limited in quantity, model generalization capability is restricted, details of an action process are not fully utilized, and an existing public data set is generally lack of score marking at a fine granularity and a process level. Although some datasets provide action annotations, their class categories are limited and do not contain corresponding scores, severely limiting the depth of model training and breadth of analytical exploration. Disclosure of Invention The invention aims to: in order to overcome the defects in the prior art, the intelligent evaluation method and the intelligent evaluation system for the quality of the pattern skating action based on process guidance are provided, and the accurate capturing and scoring of the key nodes and the process of the action are realized by constructing a pattern skating data set containing rich technical action labels and combining a multi-mode data fusion and depth time sequence model. Not only improves the accuracy and interpretation of the action quality scoring, but also provides scientific technical support for practical training and competition, and promotes the application and development of intelligent evaluation technology of pattern skating. The intelligent evaluation method for the quality of the pattern skating action based on the process guidance comprises the following steps of: S1, acquiring video data of a pattern skating match, preprocessing the data, and integrating and constructing a pattern skating action process data set; s2, extracting video features aiming at the pattern skating action process data set to obtain feature data; S3, outputting specific start-stop time of each sub-action through the constructed sub-action start-stop recognition module according to the characteristic data, and dividing continuous action paragraphs in the video according to the start-stop time to obtain a plurality of continuous sub-action video clips; S4, inputting the divided sub-action video clips into a trained process quality evaluation module, and outputting