A single-stage model-based action detection and evaluation method

Abstract

This invention relates to the field of sports training assistance technology, and in particular to a motion detection and evaluation method based on a single-stage model, comprising the following steps: acquiring sports kicking training videos in diverse scenarios, converting the video streams into image frames to form a raw dataset; manually annotating the image frames in the raw dataset and dividing them into training and validation sets; processing the training and validation sets; inputting the final training set into the model for training to obtain a trained motion detection model; inputting real-time acquired kicking motion image frames into the trained motion detection model, calculating a comprehensive motion standardization score, and providing real-time feedback of the detection results and quantitative scores. This invention uses a dual quantitative index of "hit accuracy + motion standardization score" to transform the hit accuracy and execution standardization of kicking motions into objective values, unifying the evaluation standard and avoiding interference from subjective factors such as experience and fatigue.

Description

Technical Field

[0001] This invention relates to the field of sports training assistance technology, and in particular to a motion detection and evaluation method based on a single-stage model. Background Technology

[0002] In sports like Taekwondo and Karate, where target kicking is a core training component, accurate movement detection and scientific evaluation are crucial for improving training effectiveness. Currently, target kicking training evaluation methods suffer from the following shortcomings: First, the coach's subjective judgment is limited. Traditional training relies entirely on the coach's visual observation, and evaluation results are influenced by subjective factors such as experience and fatigue, lacking unified and objective quantitative standards. Second, human supervision is inefficient; a single coach cannot accurately observe multiple trainees simultaneously, and the human eye struggles to capture and recall the high-speed details of the kicking motion. Third, existing visual technologies lack environmental adaptability. While some research has attempted to use computer vision technology for motion analysis, most directly employ standard target detection models. These solutions experience a sharp decline in performance when dealing with real-world training scenarios, especially under varying lighting conditions (such as fluctuating indoor lighting or alternating outdoor weather), leading to a significant increase in both false negatives and false positives for key targets such as feet and targets. Therefore, automated technical solutions that overcome light interference and achieve objective quantitative evaluation can promote the standardization and intelligentization of sports kicking training, improve the accuracy of detection and the fairness of competition; at the same time, they can provide coaches and managers with scientific and systematic training evaluation basis, and through model and algorithm analysis, assist in the formulation of more reasonable training guidance strategies to cope with the complex challenges of sports kicking evaluation. Summary of the Invention

[0003] The purpose of this invention is to provide a motion detection and evaluation method based on a single-stage model to improve the problems existing in the prior art.

[0004] This invention is implemented as follows: On the one hand, this invention provides an action detection and evaluation method based on a single-stage model, comprising the following steps:

[0005] S1: Use industrial-grade high-definition cameras to collect sports kicking training videos in diverse scenarios, and convert the video streams into image frames to form the original dataset;

[0006] S2: Manually annotate the image frames of the original dataset. The annotation content includes target bounding boxes, kicking end limb bounding boxes, and action type labels. According to different collection scenarios, all image frames corresponding to one or more complete scenarios are divided into a test set. All image frames corresponding to the remaining scenarios are merged and randomly divided into a training set and a validation set in a ratio of 8:2.

[0007] S3: Process the training set and validation set to obtain the processed training set and validation set;

[0008] S4: Using a YOLO series object detection model that includes a CSP backbone network and a PAN-FPN neck network, the processed training set is input into the model for training. The loss value is calculated using the CIoU loss function. The model parameters are iteratively updated using the backpropagation algorithm and gradient descent optimizer. The training convergence is monitored using the processed validation set. The model performance is evaluated using the test set to obtain the trained action detection model.

[0009] S5: Input the real-time acquired kicking action image frames into the trained action detection model, and output the target category label and target bounding box information; calculate the kicking accuracy based on the target bounding box information, calculate the action standardization comprehensive score by combining the action category confidence output by the model, and feed back the detection results and quantitative scores in real time.

[0010] More preferably, step S3, which processes the training set and validation set, specifically includes: a first-stage illumination correction and a second-stage data augmentation.

[0011] More preferably, the first-stage illumination correction specifically includes: adjusting the overall brightness of the image using an adaptive gamma transform, with the mapping relationship as follows:

[0012] ;

[0013] in, These are the brightness pixel values ​​before / after illumination correction. The average brightness of the current image or local area; For adaptive gamma exponent; For the upper and lower bounds of the adaptive gamma exponent, ;

[0014] Then, histogram equalization is used to optimize the contrast of the grayscale distribution. The formula is:

[0015] ;

[0016] in, The number of gray levels; grayscale The probability, This represents the cumulative probability of the original pixel value. The grayscale mapping value after equalization. Indicates rounding down.

[0017] More preferably, the two-stage data augmentation specifically includes at least one of the following transformations:

[0018] ① Random Affine Transformation: A random combination of scaling, rotation, and translation is applied to an image, and its coordinate transformation satisfies: ;

[0019] in, The vector of two-dimensional coordinate points before transformation. is the transformed coordinate point vector; s is the scaling factor, R is the two-dimensional rotation matrix, and t is the translation vector;

[0020] ② Color Gamut Perturbation: Randomly perturb the hue, saturation, and lightness channels in the Hue-Saturation-Value (HSV) color space: ;in, , , ; for Hue, saturation, and value in a color space. The hue, saturation, and brightness after perturbation. This corresponds to the random disturbance quantity; Indicates the interval Uniform distribution on, i.e. exist The values ​​are randomly selected evenly from each other. , Similarly, It controls the maximum perturbation amplitude of the hue channel. It controls the maximum perturbation amplitude of the saturation channel. It controls the maximum perturbation amplitude of the brightness channel;

[0021] ③ Horizontal random flipping and random scaling;

[0022] ④ Mosaic and Blending Enhancement: Blending enhancement generates new samples by linearly interpolating two images and their labels.

[0023] ;

[0024] in, This is a new sample generated after the hybrid enhancement. For two images involved in the blending enhancement; The mixing coefficient, Its prior distribution; for Distribution parameters.

[0025] More preferably, in step S4, the YOLO series object detection model is the YOLO11 model, and the gradient descent optimizer is the SGD or Adam optimizer; The formula for calculating the loss function is: ;

[0026] in, yes loss, It is the basic loss. It is a test frame With truth box The intersection and union ratio, Used to measure the consistency of aspect ratio. Based on Balance coefficient, The distance between the centers of the two frames. The minimum diagonal length of the bounding box; For prediction boxes Center and truth box The Euclidean distance from the center; The width and height of the prediction box and the width and height of the truth box; This is a term for consistency in aspect ratio.

[0027] More preferably, in step S5, a single kicking action includes The formula for the accuracy of a kick upon initial contact is: ,in Indicates the area of ​​the region. It is "kicking the end limb frame". With "target frame" The intersection, It is "kicking the end limb frame". With "target frame" The union of, It represents the accuracy of the kick, which is also the intersection-over-interference ratio of the kick end limb frame and the target frame, and its value ranges from [0,1].

[0028] More preferably, a single kicking motion includes Second contact, and At that time, the accuracy of the kick is determined by any of the following methods:

[0029] The accuracy of each contact is calculated in chronological order, and the maximum value is taken as the final accuracy score.

[0030] By time weight Perform a weighted average to obtain As the final accuracy score, among which, This indicates the sequence number of times the distal limb of the kick made contact with the target. Indicates the first The time weight of each contact. and , It is the total number of times the distal limb of the kick makes contact with the target during a single kicking motion. It is the first The intersection ratio of the kicking end limb frame and the target frame at the first contact.

[0031] More preferably, the formula for calculating the comprehensive score of action standardization in step S5 is: ;

[0032] in, It is the confidence score of the model for the identified action type, which can be regarded as a quantitative assessment of the standard of action execution; It is the accuracy index calculated above; and It is a weighting coefficient that can be adjusted by the coach according to the training objectives, and satisfies... .

[0033] On the other hand, the present invention also provides a motion detection and evaluation system based on a single-stage model, applied to any of the motion detection and evaluation methods described above, comprising:

[0034] The data acquisition module is used to acquire sports kicking training videos in diverse scenarios and convert them into raw image frames;

[0035] The annotation module is used to manually annotate the original image frames and output a dataset containing annotation information.

[0036] The dataset partitioning module is used to divide one or more complete unseen scene image frames into a test set, and the remaining scene image frames are randomly divided into a training set and a validation set in an 8:2 ratio;

[0037] The data processing module is used to perform adaptive gamma transformation and contrast-limited histogram equalization only on the training set and validation set, then perform data augmentation, and output the processed training set and validation set.

[0038] The model training module is used to deploy YOLO series object detection models and complete training, parameter optimization, and performance evaluation.

[0039] The quantitative evaluation module is used to receive real-time image frames and output detection results through the trained model, and calculate a comprehensive score of kick accuracy and action standardization.

[0040] The real-time feedback module is used to overlay and display the test results with the quantitative score and provide feedback to the user.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] 1. This invention uses a dual quantitative index of "hit accuracy + action standardization score" to convert the hit accuracy and execution standard of the kicking action into objective values, unify the evaluation standard, and avoid interference from subjective factors such as experience and fatigue.

[0043] 2. By combining "adaptive gamma transformation + histogram equalization" for illumination correction, the interference from variable lighting scenarios such as indoor lighting with varying brightness and outdoor lighting with alternating sunshine and rain is effectively reduced, significantly lowering the false negative and false positive rates for key targets.

[0044] 3. By employing diverse scene data collection, scene-based dataset partitioning strategies (the test set consists of entirely new scenes not seen by the model), and dynamic data augmentation techniques, combined with the CSP backbone network and PAN-FPN neck network of the YOLO11 model, accurate detection of targets at different scales is achieved.

[0045] 4. The CIoU loss function is adopted, which adds the distance between the center points of the bounding boxes and the aspect ratio constraints on the basis of the traditional IoU. This provides technical support for the quantification of scoring accuracy. Combined with validation set monitoring and test set evaluation, the model performance is guaranteed to be stable.

[0046] 5. By overlaying and displaying the detected target bounding boxes, action category labels, and comprehensive scores in real time, athletes are provided with an intuitive reference for their action shortcomings, and coaches are provided with scientific data to develop personalized training strategies and track training effects, thus promoting the standardization and intelligentization of training. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, 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. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0049] This embodiment provides a single-stage motion detection and evaluation method. Its core idea lies in overcoming the robustness challenge of existing technologies in complex real-world environments through systematic data engineering and sophisticated model design, and replacing traditional subjective evaluation with objective, quantifiable indicators. The specific implementation process of this method will be described in detail below, such as... Figure 1 As shown:

[0050] S1: Employs industrial-grade high-definition cameras to capture video of the sports kicking target training process at a resolution of no less than 1920×1080 pixels and a frame rate of no less than 30fps. The high frame rate clearly captures the high-speed limb movements at the moment of kicking, effectively reducing motion blur. To ensure data consistency, the camera position remains fixed throughout the acquisition process. To improve the model's environmental generalization ability, the acquisition work is conducted in diverse scenarios, including indoor training halls with uniform lighting, outdoor venues with both strong light and shadow, and complex backgrounds in multi-purpose stadiums. After acquisition, video processing tools are used... The video stream is converted into image frames by using equal-interval frame extraction or keyframe extraction based on motion changes, thus forming the original dataset.

[0051] S2: To provide supervision information to the model, professional annotation tools are used. The image frames are manually annotated. The annotations include a bounding box representing the target's location and a bounding box representing the area of ​​actual contact with the target by the kicking limb. To classify the action, the kicking limb is assigned different action type labels based on the action type, such as front_kick_foot and side_kick_foot.

[0052] In terms of dataset partitioning, this embodiment employs a strict scenario-based partitioning strategy to realistically evaluate the model's generalization ability. Specifically, one or more complete scenario datasets that the model has never encountered during training (e.g., all images collected in a complex stadium) are allocated as the test set. Data from all other scenarios are merged and then randomly divided into training and validation sets in an 8:2 ratio. This partitioning method ensures that the testing environment is entirely new to the model, and the evaluation results more accurately reflect the model's robustness in real-world deployments.

[0053] S3: The first stage of image preprocessing only processes the images in the training and validation sets offline to reduce the interference of illumination changes on model performance. The test set retains the original image frames and is only used for model performance evaluation. In this stage, an adaptive gamma transform is first used to adjust the overall brightness of the images, and its mapping relationship is as follows:

[0054] ;

[0055] in, These are the brightness pixel values ​​before / after illumination correction. The average brightness of the current image or local area; For adaptive gamma exponent; For the upper and lower bounds of the adaptive gamma exponent, ;

[0056] Then, histogram equalization is used to optimize the contrast of the grayscale distribution, ultimately resulting in an enhanced image frame with "uniform brightness and clear details," minimizing the interference of illumination changes. The formula is:

[0057] ;

[0058] in, The number of gray levels; grayscale The probability, This represents the cumulative probability of the original pixel value. The grayscale mapping value after equalization. This indicates rounding. After the above illumination correction process is completed, the generated enhanced image frames are merged with the corresponding original image frames to form the training dataset (including the training set and the validation set). This "hybrid" strategy enables the model to learn both the ideal features after illumination normalization and the features under the original variable illumination, thereby obtaining a more essential feature representation that is insensitive to illumination changes themselves.

[0059] The second stage performs data augmentation on the training dataset constructed in the first stage, resulting in processed training and validation sets. This stage is dynamically executed during the data loading process of model training to further expand the diversity of the data and suppress overfitting. This embodiment employs, but is not limited to, the following transformations:

[0060] ① Random Affine Transformation: A random combination of scaling, rotation, and translation is applied to an image, and its coordinate transformation satisfies: ;

[0061] in, The vector of two-dimensional coordinate points before transformation. is the transformed coordinate point vector; s is the scaling factor, R is the two-dimensional rotation matrix, and t is the translation vector;

[0062] ② Color Gamut Perturbation: Randomly perturb the hue, saturation, and lightness channels in the Hue-Saturation-Value (HSV) color space: ;

[0063] in, ; for Hue, saturation, and value in a color space. The hue, saturation, and brightness after perturbation. This corresponds to the random disturbance quantity; Indicates the interval Uniform distribution on, i.e. exist The values ​​are randomly selected evenly from each other. , Similarly, It controls the maximum perturbation amplitude of the hue channel. It controls the maximum perturbation amplitude of the saturation channel. It controls the maximum perturbation amplitude of the brightness channel.

[0064] ③ Horizontal random flipping and random scaling;

[0065] ④ Mosaic and Blending Enhancement: Blending enhancement generates new samples by linearly interpolating two images and their labels.

[0066]

[0067] in, This is a new sample generated after the hybrid enhancement. For two images involved in the blending enhancement; The mixing coefficient, Its prior distribution; for Distribution parameters.

[0068] S4: This embodiment uses the advanced YOLO11 model architecture, which has real-time detection capabilities, from the YOLO series. It should be clarified that the samples input to the model during the training / validation phase are the training / validation set samples processed in step S3. The first stage of illumination correction is processed offline, while the second stage of data augmentation is performed dynamically during data loading. The test set retains the original image frames and is used only for evaluation. The YOLO11 model typically includes a backbone network using a CSP structure and a... The neck network of the structure, in which the CSP (Content Scale Spinner) backbone network is used to extract multi-dimensional basic features from the input image and output multi-scale features. The neck network of the structure is used to fuse multi-scale features output by the backbone network to achieve feature adaptation for targets of different scales.

[0069] The processed training set was input into the YOLO11 model for training. During the training process, to improve the accuracy of bounding box localization, the following methods were employed: The (Complete Intersection over Union) loss function calculates the loss value. This loss function is used in traditional... In addition to the loss, the consistency of the distance between the center points of the bounding box and the aspect ratio is further considered, which is defined as: ;

[0070] in, yes loss, It is the basic loss. It is a test frame With truth box The intersection and union ratio, Used to measure the consistency of aspect ratio. Based on Balance coefficient, The distance between the centers of the two frames. The minimum diagonal length of the bounding box; For prediction boxes Center and truth box The Euclidean distance from the center; The width and height of the prediction box and the width and height of the truth box; This is a term related to aspect ratio consistency.

[0071] Based on the above The loss value is calculated by backpropagation algorithm to determine the gradient of network parameters. Gradient descent optimizers (such as SGD and Adam) are used to iteratively update the parameters of each layer of the model, such as the backbone network and the neck network, in order to minimize the loss value. After each round of training, the validation set data is input into the model for inference to evaluate the detection accuracy (such as mAP) of the model on the validation set. If the validation set accuracy does not improve for several consecutive rounds, the model training is considered to have converged and training is terminated.

[0072] It should be clear that an independent test set is used to input the trained model, calculate the model's detection accuracy, speed and other metrics on the test set, and complete the final evaluation of the model's performance.

[0073] S5: Analyze the real-time captured image frames of the kicking action using the trained model and output quantitative evaluation results. This step transforms the model's detection output into an objective score that provides practical guidance for athletes. The specific process is as follows:

[0074] S501: Hit accuracy evaluation; For each frame of the real-time video, the model outputs the detected target category and bounding box, such as "kicking the end limb bounding box". With "target frame" When both are detected simultaneously, the system calculates their crossover-union ratio (CUIR). This is used to quantify the accuracy of a kick; the indicator is called... Its definition is: ,in Indicates the area of ​​the region. It is "kicking the end limb frame". With "target frame" The intersection, It is "kicking the end limb frame". With "target frame" The union of, This indicates the accuracy of the kick, which is also the intersection-over-interference ratio of the kicking end limb frame and the target frame. Its value ranges from [0,1]. The larger the value, the more accurate the hit.

[0075] For multiple contacts in a single action, handle them using either of the following two methods:

[0076] The accuracy of each contact is calculated in chronological order, and the maximum value is taken as the final accuracy score.

[0077] By time weight Perform a weighted average to obtain As the final accuracy score, among which, This indicates the sequence number of times the distal limb of the kick made contact with the target. Indicates the first The time weight of each contact. and , It is the total number of times the distal limb of the kick makes contact with the target during a single kicking motion. It is the first The intersection ratio of the kicking end limb frame and the target frame at the first contact.

[0078] S502: Action Standardization Score; This invention constructs a comprehensive scoring model to generate the final action standardization score. The score is obtained by weighted summation of the confidence level and accuracy of the model in recognizing the action:

[0079] ;

[0080] in, It is the confidence score of the model for the identified action type, which can be regarded as a quantitative assessment of the standard of action execution; It is the accuracy index calculated above; and It is a weighting coefficient that can be adjusted by the coach according to the training objectives, and satisfies... For example, in the initial training phase, the intensity can be increased. The weighting is used to emphasize hit rate; in advanced training phases, it can be increased. The weighting is used to emphasize the standardization of the action.

[0081] S503: Finally, combine all detection results, including the target's bounding box, action category label, and calculated overall score. It is overlaid on the video screen in real time, providing athletes and coaches with immediate, intuitive, and quantifiable feedback.

[0082] Through the complete technical process described above, this invention provides a sports target kicking action detection and evaluation system that can work stably and accurately even in complex environments, offering strong technical support for scientific training.

[0083] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A motion detection and evaluation method based on a single-stage model, characterized in that, Includes the following steps: S1: Use industrial-grade high-definition cameras to collect sports kicking training videos in diverse scenarios, and convert the video streams into image frames to form the original dataset; S2: Manually annotate the image frames of the original dataset. The annotation content includes target bounding boxes, kicking end limb bounding boxes, and action type labels. According to different collection scenarios, all image frames corresponding to one or more complete scenarios are divided into a test set. All image frames corresponding to the remaining scenarios are merged and randomly divided into a training set and a validation set in a ratio of 8:

2. S3: Process the training set and validation set to obtain the processed training set and validation set; S4: Using a YOLO series object detection model that includes a CSP backbone network and a PAN-FPN neck network, the processed training set is input into the model for training. The loss value is calculated using the CIoU loss function. The model parameters are iteratively updated using the backpropagation algorithm and gradient descent optimizer. The training convergence is monitored using the processed validation set. The model performance is evaluated using the test set to obtain the trained action detection model. S5: Input the real-time acquired target kicking action image frames into the trained action detection model, and output the target category label and target bounding box information; The kicking accuracy is calculated based on the target bounding box information. The action standardization comprehensive score is calculated by combining the action category confidence score output by the model. The detection results and quantitative scores are fed back in real time. In step S5, a single kicking motion includes The formula for the accuracy of a kick upon initial contact is: ,in Indicates the area of ​​the region. It is "kicking the end limb frame". With "target frame" The intersection, It is "kicking the end limb frame". With "target frame" The union of, It represents the accuracy of the kick, and is also the intersection-over-interference ratio of the kick end limb frame and the target frame, with a value range of [0,1]. One kicking motion includes Second contact, and At that time, the accuracy of the kick is determined by any of the following methods: The accuracy of each contact is calculated in chronological order, and the maximum value is taken as the final accuracy score. By time weight Perform a weighted average to obtain As the final accuracy score, among which, This indicates the sequence number of times the distal limb of the kick made contact with the target. Indicates the first The time weight of each contact. and , It is the total number of times the distal limb of the kick makes contact with the target during a single kicking motion. It is the first The intersection ratio of the kicking end limb frame and the target frame at the first contact; The formula for calculating the comprehensive score of action standardization in step S5 is as follows: ; in, It is the confidence score of the model for the identified action type, which can be regarded as a quantitative assessment of the standard of action execution; It is the accuracy index calculated above; and It is a weighting coefficient that can be adjusted by the coach according to the training objectives, and satisfies... .

2. The action detection and evaluation method based on a single-stage model according to claim 1, characterized in that, Step S3, specifically the processing of the training set and validation set, includes: a first-stage illumination correction and a second-stage data augmentation.

3. The action detection and evaluation method based on a single-stage model according to claim 2, characterized in that, The first stage of illumination correction specifically includes: adjusting the overall brightness of the image using adaptive gamma transform, with the mapping relationship as follows: ； in, These are the brightness pixel values ​​before / after illumination correction. The average brightness of the current image or local area; For adaptive gamma exponent; For the upper and lower bounds of the adaptive gamma exponent, ; Then, histogram equalization is used to optimize the contrast of the grayscale distribution. The formula is: ； in, The number of gray levels; grayscale The probability, This represents the cumulative probability of the original pixel value. The grayscale mapping value after equalization. Indicates rounding down.

4. The action detection and evaluation method based on a single-stage model according to claim 3, characterized in that, Two-stage data augmentation specifically includes at least one of the following transformations: ① Random Affine Transformation: A random combination of scaling, rotation, and translation is applied to an image, and its coordinate transformation satisfies: ; in, The vector of two-dimensional coordinate points before transformation. is the transformed coordinate point vector; s is the scaling factor, R is the two-dimensional rotation matrix, and t is the translation vector; ② Color Gamut Perturbation: Randomly perturb the hue, saturation, and lightness channels in the Hue-Saturation-Value (HSV) color space: ;in, , , ; for Hue, saturation, and value in a color space. The hue, saturation, and brightness after perturbation. This corresponds to the random disturbance quantity; Indicates the interval Uniform distribution on, i.e. exist The values ​​are randomly selected evenly from each other. , Similarly, It controls the maximum perturbation amplitude of the hue channel. It controls the maximum perturbation amplitude of the saturation channel. It controls the maximum perturbation amplitude of the brightness channel; ③ Horizontal random flipping and random scaling; ④ Mosaic and Blending Enhancement: Blending enhancement generates new samples by linearly interpolating two images and their labels. ； in, This is a new sample generated after the hybrid enhancement. For two images involved in the blending enhancement; The mixing coefficient, Its prior distribution; for Distribution parameters.

5. The action detection and evaluation method based on a single-stage model according to claim 1, characterized in that, In step S4, the YOLO series object detection model is the YOLO11 model, and the gradient descent optimizer is either SGD or Adam optimizer; The formula for calculating the loss function is: ; in, yes loss, It is the basic loss. It is a test frame With truth box The intersection and union ratio, Used to measure the consistency of aspect ratio. For based on Balance coefficient, The distance between the centers of the two frames. The minimum diagonal length of the bounding box; For prediction boxes Center and truth box The Euclidean distance from the center; The width and height of the prediction box and the width and height of the truth box; This is a term for aspect ratio consistency.

6. A motion detection and evaluation system based on a single-stage model, applied to the motion detection and evaluation method according to any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire sports kicking training videos in diverse scenarios and convert them into raw image frames; The annotation module is used to manually annotate the original image frames and output a dataset containing annotation information. The dataset partitioning module is used to divide one or more complete unseen scene image frames into a test set, and the remaining scene image frames are randomly divided into a training set and a validation set in an 8:2 ratio; The data processing module is used to perform adaptive gamma transformation and contrast-limited histogram equalization only on the training set and validation set, then perform data augmentation, and output the processed training set and validation set. The model training module is used to deploy YOLO series object detection models and complete training, parameter optimization, and performance evaluation. The quantitative evaluation module is used to receive real-time image frames and output detection results through the trained model, and calculate a comprehensive score of kick accuracy and action standardization. The real-time feedback module is used to overlay and display the test results with the quantitative score and provide feedback to the user.

Images