Ice hockey training method and apparatus based on ice hockey quantification indicators

Abstract

The application provides an ice hockey training method and device based on an ice hockey quantitative index, and belongs to the technical field of artificial intelligence, and specifically comprises the following steps: acquiring a plurality of historical ice hockey game video data, learning and processing training sample data sets by using a lightweight YOLO v8n model to obtain a target detection model, collecting real-time training video of a to-be-tested player by using a mobile terminal camera, extracting frame images from the real-time training video according to timestamps to obtain real-time training images, inputting the real-time training images into the lightweight YOLO v8 model, outputting detection results, associating the detection results to the same timestamp, inputting the detection results into an evaluation and analysis model, obtaining ice hockey quantitative indexes, inputting the ice hockey quantitative indexes into a multi-modal feedback model, comparing the ice hockey quantitative indexes with preset standard values by using the multi-modal feedback model, and generating sports guidance information. Through the processing scheme, the ice hockey and the human hockey stick collaborative scene of a small target and high speed during sports can be detected and tracked in real time.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to ice hockey training methods and devices based on quantitative indicators of ice hockey. Background Technology

[0002] Ice hockey is a high-intensity, high-speed, and highly competitive team sport. The precision and efficiency of passing and control movements directly affect game performance, but traditional training and evaluation methods rely heavily on the ice rink and on-site coaching. However, ice hockey rinks are expensive to build and maintain, making them inaccessible to the average enthusiast. Furthermore, coaches' subjective evaluations often lack quantitative data support, making it difficult to quantify the details and quality of movements.

[0003] Existing ball sports training servers calculate the ball's trajectory, landing point, and speed after capturing training videos and performing target detection, then use this data for virtual training. However, directly transferring this technology to ice hockey training scenarios presents the following unresolved technical difficulties: 1. Due to the small size and high speed of ice hockey, existing target detection methods have a high false negative rate. A standard ice hockey puck is only 7.62 centimeters in diameter, often smaller than 10×10 pixels in a long-distance camera shot, making it a typical small target. Existing target detection algorithms have insufficient recall for small targets. Furthermore, ice hockey pucks can reach speeds of up to 160 km / h, and current technologies lack optimized solutions for detecting small targets moving at high speeds, leading to serious missed detections and frame skipping.

[0004] 2. Hockey stick control movements cannot be evaluated using general human posture estimation. In ice hockey, the point of contact between the puck and the end of the stick can be 1.5 to 1.8 meters away from the human joint. General attitude estimation models do not include this key point of the stick end, and therefore cannot directly determine ice hockey-specific technical characteristics such as stick face angle and puck direction.

[0005] 3. Lack of real-time interactive training feedback, making it difficult to operate independently of coach guidance. Existing technologies primarily provide post-training data analysis and VR virtual training, but do not address real-time voice correction, visual guidance, or virtual opponent interaction for athletes during actual physical training. Ice hockey training, in particular, requires immediate feedback, which cannot be met by purely offline analysis. Summary of the Invention

[0006] Therefore, in order to overcome the shortcomings of the prior art, the present invention provides an ice hockey training method based on quantitative ice hockey indicators that can be applied to ice hockey training, comprising: Multiple historical ice hockey game video datasets were acquired to obtain a training sample dataset. Each historical ice hockey game video dataset included an annotated ice hockey puck, key points of the player's body posture, and the end of the hockey stick. A lightweight YOLO v8n model was used to perform frame image learning processing on the training sample dataset to obtain a target detection model for recognizing the ice hockey puck, player posture, and hockey stick end. The lightweight YOLO v8n model was developed using a pruning algorithm on YOLO v8n. The v8n model is obtained by pruning redundant channels; real-time training videos of the players under test are captured by a mobile terminal camera, and frame images are extracted from the real-time training videos according to timestamps to obtain real-time training images; the real-time training images are input into the target detection model, and the detection results are output and associated with the same timestamp; the detection results are input into the evaluation and analysis model according to the time order of the timestamps to obtain hockey quantitative indicators. The evaluation and analysis model calculates hockey quantitative indicators based on the geometric relationship between the puck trajectory and the human posture to evaluate the technical execution quality of the players under test; the hockey quantitative indicators are input into a multimodal feedback model, which compares the hockey quantitative indicators with preset standard values ​​to generate sports guidance information.

[0007] Preferably, the method further includes: extracting the defensive player's motion posture data based on the sequence of key points of the defensive player's human posture and the stick end label in the training sample dataset; inputting the motion posture data into an intent reasoning model constructed by a trained neural network, and outputting a multi-dimensional intent probability distribution vector, each dimension of the multi-dimensional intent probability distribution vector corresponding to a preset hockey defensive action; generating the virtual defender's ice surface movement trajectory under hockey kinematic constraints according to the dominant intent in the multi-dimensional intent probability distribution vector and the motion posture data; and calculating the optimal blocking posture of the virtual defender at each moment according to the ice surface movement trajectory and a geometric blocking algorithm, wherein the geometric blocking algorithm is used to maximize the coverage area of ​​the virtual defender over the shooting angle.

[0008] Preferably, the target detection model is trained through the following steps: The YOLO v8n model is sparsely trained by adding an L1 regularization term to the loss function and applying an absolute value penalty to the convolutional kernel weights to make some convolutional kernel weights approach zero, thus obtaining a trained detection model; the detection accuracy of the trained detection model is determined to be no less than a preset accuracy threshold; when the detection accuracy is no less than the preset accuracy threshold, the trained detection model is pruned to obtain a pruned detection model; the pruned detection model is optimized using TensorRT to convert its accuracy from FP32 to INT8 quantization accuracy.

[0009] Preferably, the real-time training video of the player to be tested is acquired by a mobile terminal camera, and the real-time training image is obtained by extracting frame images from the real-time training video according to the timestamp. This includes: acquiring the real-time training video of the player to be tested by a mobile terminal camera; generating a rectangular calibration area covering the movement area according to the viewing angle and shooting range of the mobile terminal camera; sequentially acquiring the image coordinates of the four corners of the movement area; obtaining a spatial calibration mapping matrix based on the image coordinates using the direct linear transformation method, and performing condition number verification on the matrix; after the verification is passed, confirming the spatial calibration accuracy by verifying that the deviation of the virtual grid intersection point is less than a preset threshold; performing perspective transformation preprocessing on the real-time training video that has completed spatial calibration, and extracting frame images according to the timestamp to obtain the real-time training image.

[0010] Preferably, the step of inputting the real-time training image into the target detection model, outputting the detection result, and associating the detection result with the same timestamp includes: inputting the real-time training image into a lightweight YOLOv8 model, and the detection head outputting sphere detection boxes and human pose key points in the image, wherein the sphere detection box consists of bounding box coordinates and target confidence, and the human pose key points consist of pixel coordinates and visibility; merging redundant sphere detection boxes using NMS, sorting them according to the target confidence, retaining the sphere detection box with the highest confidence, and deleting other sphere detection boxes that overlap with the sphere detection box by more than a threshold; converting the pixel coordinates of the retained sphere detection box center coordinates and the human pose key point pixel coordinates into real-world plane coordinates using a spatial calibration mapping matrix; applying a Kalman filter algorithm to the sphere detection box center coordinates to obtain smooth sphere center world coordinates, and applying a moving average filter to the human pose key point coordinates to obtain smooth human pose key point coordinates; and associating the smooth sphere center world coordinates and human pose key point coordinates with the same timestamp.

[0011] Preferably, the step of inputting the detection results into the evaluation and analysis model according to the time sequence of the timestamps to obtain hockey quantitative indicators, wherein the evaluation and analysis model calculates hockey quantitative indicators based on the geometric relationship between the puck trajectory and the human posture, and is used to evaluate the technical execution quality of the player under test, including: aligning the puck detection frame sequence, the human posture key point sequence, and the stick end sequence according to the timetamp sequence; inputting the aligned detection results into the evaluation and analysis model; the evaluation and analysis model constructing the puck trajectory, the human posture sequence, and the stick end trajectory; and, based on the puck trajectory, the preset human skeletal key points, and the spatial coordinates of the stick end, calculating the hockey quantitative indicators of the puck relative to the player under test, the stick face, and the human posture, and evaluating the technical execution quality of the athlete, wherein the hockey quantitative indicators include at least the following parameters: the puck contact frequency calculated based on the puck velocity direction change rate, the deviation value of the puck trajectory from the preset ideal path, the torso direction determined by the human posture, and the angle between the torso direction and the direction of attack of the virtual defender.

[0012] Preferably, it further includes adaptive ball recognition: acquiring a preprocessed image captured by a mobile terminal camera, extracting the shape, size, and color of the sphere in the image; inputting the shape, size, and color of the sphere into a pre-trained classification model, and outputting a ball category identifier; and dynamically adjusting the anchor frame size, confidence threshold, and physical size conversion coefficient based on the identified ball category identifier and the actual diameter of the sphere.

[0013] Preferably, the step of inputting the hockey quantification index into a multimodal feedback model, wherein the multimodal feedback model compares the hockey quantification index with a preset standard value to generate exercise guidance information, includes: the multimodal feedback model acquiring the hockey quantification index and the preset standard value, wherein the preset standard value corresponds to the optimal hitting parameter range for hockey; calculating the deviation value between the hockey quantification index and the preset standard value; and generating guidance information based on the deviation value, wherein the guidance information includes at least one of the following: action correction prompts, force adjustment suggestions, and angle optimization suggestions.

[0014] Preferably, the method further includes dynamic adjustment of training parameters: statistically analyzing performance indicators within a fixed time window, the performance indicators including at least one of success rate, average reaction time, number of consecutive successes or number of consecutive failures; calculating the success rate of the current window based on the performance indicators, and dynamically adjusting training parameters based on the success rate, the training parameters including at least one of virtual defender speed, reaction time tolerance, and task duration.

[0015] An ice hockey training device based on ice hockey quantitative indicators includes: a sample acquisition module for acquiring multiple historical ice hockey game video data to obtain a training sample dataset, wherein each historical ice hockey game video data includes an annotated ice hockey puck and key points of the player's body posture and the end of the hockey stick corresponding to each player; and a model improvement module for learning and processing the training sample dataset using a lightweight YOLO v8n model to obtain a target detection model, wherein the lightweight YOLO v8n model is obtained by using a pruning algorithm to improve the YOLO... The v8n model is obtained by pruning redundant channels; the image acquisition module is used to acquire real-time training videos of the players under test through the mobile terminal camera, and extract frame images from the real-time training videos according to the timestamps to obtain real-time training images; the image processing module is used to input the real-time training images into the target detection model, output the detection results, and associate the detection results with the same timestamp; the evaluation and analysis module is used to input the detection results into the evaluation and analysis model according to the time order of the timestamps to obtain hockey quantitative indicators. The evaluation and analysis model calculates hockey quantitative indicators based on the geometric relationship between the puck trajectory and the human posture to evaluate the technical execution quality of the players under test; the feedback module inputs the hockey quantitative indicators into a multimodal feedback model. The multimodal feedback model compares the hockey quantitative indicators with preset standard values ​​to generate sports guidance information.

[0016] Compared with the prior art, the advantages of the present invention are as follows: It can detect and track small targets and high-speed ice hockey and human stick coordination scenarios in real time, and automatically calculate ice hockey-specific indicators such as touch frequency, trajectory deviation rate, torso direction and the angle between the torso direction and the direction of the virtual defender's attack (exposure angle).

[0017] Users only need a mobile device with a camera and a small area of ​​ordinary ground to obtain professional-level quantitative feedback and competitive training, without the need for ice, professional cameras, calibration boards, or coaches.

[0018] The multimodal feedback mechanism generates corrective exercise guidance information immediately after an incorrect movement occurs, preventing the reinforcement of bad habits. Compared to traditional coaching based on observation and verbal instruction, the feedback speed is significantly improved. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1This is a flowchart illustrating the ice hockey training method based on quantitative indicators in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an ice hockey training device based on quantitative indicators of ice hockey in an embodiment of the present invention. Detailed Implementation

[0021] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0022] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] It should be noted that the following description covers various aspects of embodiments within the scope of protection of this invention. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0024] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0025] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0026] To address the limitations of existing ball sports training servers, this approach involves capturing training videos, performing target detection, calculating the ball's trajectory, landing point, and speed, and then using this data for virtual training. However, directly applying this technology to ice hockey training scenarios presents the following unresolved technical challenges: 1. Due to the small size and high speed of ice hockey, existing target detection methods have a high false negative rate. A standard ice hockey puck is only 7.62 centimeters in diameter, often smaller than 10×10 pixels in a long-distance camera shot, making it a typical small target. Existing target detection algorithms have insufficient recall for small targets. Furthermore, ice hockey pucks can reach speeds of up to 160 km / h, and current technologies lack optimized solutions for detecting small targets moving at high speeds, leading to serious missed detections and frame skipping.

[0027] 2. Hockey stick control movements cannot be evaluated using general human posture estimation. In ice hockey, the point of contact between the puck and the end of the stick can be 1.5 to 1.8 meters away from the human joint. General attitude estimation models do not include this key point of the stick end, and therefore cannot directly determine ice hockey-specific technical characteristics such as stick face angle and puck direction.

[0028] 3. Lack of real-time interactive training feedback, making it difficult to operate independently of coach guidance. Existing technologies primarily provide post-training data analysis and VR virtual training, but do not address real-time voice correction, visual guidance, or virtual opponent interaction for athletes during actual physical training. Ice hockey training, in particular, requires immediate feedback, which cannot be met by purely offline analysis.

[0029] In view of this, this invention proposes an ice hockey training method based on quantitative indicators. This method can detect and track small-target, high-speed ice hockey and player-stick coordination scenarios in real time, automatically calculating ice hockey-specific indicators such as puck contact frequency, trajectory deviation rate, torso direction, and the angle between the torso direction and the virtual defender's attack direction (exposure angle). Users only need a mobile terminal with a camera and a small area of ​​ordinary ground; no ice surface, professional camera, calibration board, or coach are required to obtain professional-level quantitative feedback and competitive training. The multimodal feedback mechanism can generate movement guidance information for correction as soon as an incorrect movement occurs, avoiding the reinforcement of incorrect habits. Compared with traditional coach-observed verbal guidance, the feedback speed is greatly improved.

[0030] This hockey training method based on quantitative indicators can be applied to servers or terminals. Terminals can be, but are not limited to, various personal computers, laptops, smartphones, tablets, and portable smart devices. Servers can be implemented using independent servers or server clusters composed of multiple servers.

[0031] like Figure 1As shown, this application provides an ice hockey training method based on quantitative indicators of ice hockey. Taking the application of this method to a server as an example, the method includes the following steps: Step 101: Obtain multiple historical ice hockey game video data to obtain a training sample dataset. Each historical ice hockey game video data includes an annotated ice hockey puck and key points of the player's body posture and stick end corresponding to each player.

[0032] The server retrieves video data from multiple historical ice hockey games as training samples. This data comes from publicly available game recordings, professional training logs, and simulations of games played with other pucks. These videos cover real-life scenarios with varying venue conditions, puck types, lighting conditions, intensity of competition, and player skill levels.

[0033] The server extracts raw images from historical ice hockey game video data by frame extraction and cleaning. These raw images are then annotated, including: bounding box annotations for each puck in each frame, identifying the puck type and selecting the corresponding puck; pose keypoint annotations for each player's body; and position annotations for the stick tip. The resulting raw images contain the annotated information. After denoising, enhancement, and size normalization, these annotated raw images are used for subsequent target detection model training.

[0034] To achieve high-precision recognition of hockey puck and human postures, this invention constructs a dedicated hockey training dataset, which includes the following subsets:

[0035] Step 102: The target detection model is obtained by learning the training sample dataset using a lightweight YOLO v8n model. The lightweight YOLO v8n model is obtained by pruning redundant channels of the YOLO v8n model using a pruning algorithm.

[0036] Using YOLOv8n as the base model, a fourth output scale 160×160 feature map is added. By adding an upsampling branch after the backbone network, the deep feature map is fused with the shallow feature map to generate a high-resolution feature map, resulting in the adjusted YOLO v8n model.

[0037] The server performs sparse training on the adjusted YOLO v8n model, adding an L1 regularization term to the loss function to make the weights of some convolutional kernels approach zero. The strength coefficient of L1 regularization can be set according to actual needs; in this embodiment, it is set to 0.001. A global threshold pruning strategy is adopted to remove convolutional kernels whose contribution is less than 1% of the global maximum. The pruning ratio is approximately 30%, reducing the number of parameters from 3.2M to 2.2M, significantly reducing the model size. The pruned model is then fine-tuned to recover the accuracy loss caused by pruning. Testing shows that the mean accuracy (mAP) of the fine-tuned model decreases by no more than 1.5%, maintaining high detection accuracy while ensuring lightweight design.

[0038] The server converts the pruned and fine-tuned model from FP32 precision to INT8 quantization precision, resulting in a lightweight YOLOv8n model. During the conversion, a calibration set is used to statistically analyze the distribution range of activation values ​​in each layer, determining the quantization parameters (including scaling factors and zeros) for each layer. The quantized model size is compressed from approximately 13MB to approximately 4.5MB, a compression ratio exceeding 65%. Inference tests on mobile devices reduced the processing speed from 45ms / frame to 28ms / frame, achieving the frame rate requirement of over 30fps for real-time detection. In one embodiment, the FP32 model is converted to INT8 quantization format using the TensorRT optimization tool, and a layer-by-layer calibration method is employed to reduce accuracy loss during quantization.

[0039] Step 103: Collect real-time training videos of the players to be tested using the mobile terminal camera, and extract frame images from the real-time training videos based on timestamps to obtain real-time training images.

[0040] The mobile terminal's camera captures real-time training videos of the players under test. After verifying the real-time training videos, the server scans and locates the target time point frame by frame according to a pre-configured sampling strategy (e.g., fixed-frequency sampling of one frame every N frames, dynamic sampling based on motion detection algorithms to automatically identify key movement changes, or sampling at specific time points specified by external instructions). This is done based on the timestamps within the real-time training videos or the timeline recorded externally. Each located frame is then decoded and decompressed using a video decoder, converting the original video stream into a digital image matrix in RGB or BGR format.

[0041] Step 104: Input the real-time training image into the target detection model, output the detection result, and associate the detection result with the same timestamp.

[0042] The server trains a lightweight YOLO v8 model in real time using image inputs. After the model completes inference, the detection head outputs two detection results: (1) The first detection result contains candidate boxes, and each candidate box contains the coordinates of the sphere bounding box and the target confidence.

[0043] (2) The second detection result includes the coordinates of 17 key points of a series of human instances and the coordinates of the end of the cue stick and its visibility score. Each key point consists of pixel coordinates and visibility, which is predicted by the model to indicate whether it is occluded or outside the image range.

[0044] The server associates the first and second detection results output by the detection head with the same image to ensure that they belong to the same trainer, and binds the association result to the corresponding timestamp for subsequent quantitative evaluation.

[0045] Step 105: Input the detection results into the evaluation and analysis model according to the time sequence of the timestamps to obtain hockey quantitative indicators. The evaluation and analysis model calculates hockey quantitative indicators based on the geometric relationship between the hockey trajectory and the human posture, which are used to evaluate the technical execution quality of the player under test.

[0046] The server aligns the sphere detection box sequence, the human pose keypoint sequence, and the club tip sequence in timestamp order to form a spatiotemporally consistent fusion detection result. The server then inputs the aligned detection result into a pre-trained evaluation and analysis model. Based on the kinematic biomechanical geometric relationship between the sphere trajectory and the human pose, the evaluation and analysis model performs the following operations: The evaluation and analysis model constructs the ball's trajectory, the temporal sequence of the player's human posture, and the spatial trajectory of the end of the cue stick based on the input fusion detection results.

[0047] The evaluation and analysis model is based on the constructed ball trajectory, the spatial coordinates of preset key points of the human skeleton (including but not limited to the shoulder, elbow, wrist, hip, and knee joints), and the real-time spatial coordinates of the hockey stick end. Through geometric calculation and dynamic modeling, it calculates a set of quantitative indicators of hockey with biomechanical significance. These quantitative indicators include, but are not limited to: instantaneous velocity and acceleration of the puck, the orientation deviation angle of the stick face at the moment of impact, the synchronization parameter between the player's wrist force and the torso rotational angular velocity, and the launch angle of the puck when it leaves the stick.

[0048] The evaluation and analysis model uses these quantitative indicators as outputs to quantitatively assess the technical execution quality of the player in this hitting action. The evaluation results can objectively reflect the standardization of the player's movements, the efficiency of power transfer, and the accuracy of the hit, providing a scientific basis for subsequent training guidance and technical improvement.

[0049] Step 106: Input the hockey quantitative index into the multimodal feedback model. The multimodal feedback model compares the hockey quantitative index with the preset standard value to generate sports guidance information.

[0050] The server inputs hockey quantification metrics into a pre-trained multimodal feedback model. The core function of this model is to compare the quantification metrics with a pre-defined database of standard values ​​in real time. Based on the comparison results, it simultaneously generates and delivers motion guidance information that the user can immediately perceive through two independent output channels: voice and vision. In a realistic physical training environment where the user is focused on hockey control movements, the multimodal feedback model provides low-latency, high-definition, and non-intrusive feedback, enabling the user's erroneous movements to be identified and corrected instantly, while maximizing the continuity and immersion of the training.

[0051] The above steps can detect and track small targets and high-speed ice hockey and human stick coordination scenarios in real time, and realize the automatic calculation of ice hockey-specific indicators such as touch frequency, trajectory deviation rate, torso direction and the angle between the torso direction and the direction of the virtual defender's attack (exposure angle).

[0052] Users only need a mobile device with a camera and a small area of ​​ordinary ground to obtain professional-level quantitative feedback and competitive training, without the need for ice, professional cameras, calibration boards, or coaches.

[0053] The multimodal feedback mechanism generates corrective exercise guidance information immediately after an incorrect movement occurs, preventing the reinforcement of bad habits. Compared to traditional coaching based on observation and verbal instruction, the feedback speed is significantly improved.

[0054] In one embodiment, the method further includes generating a virtual defender and planning offensive tactics accordingly, comprising the following steps: Step 1: Based on the sequence of key points of the defensive player's human posture and the label at the end of the cue stick in the training sample dataset, extract the motion posture data of the defensive player.

[0055] The server uses algorithms such as OpenPose and HRNet to detect and track the coordinates of skeletal key points (e.g., head, shoulder, elbow, wrist, knee, ankle) of defensive players from the training sample dataset. To improve accuracy, multi-view or 3D pose estimation methods can be employed. The server concatenates the coordinates of the defensive player's body center point (e.g., midpoint of the hip) or key points (e.g., both feet) in consecutive frames to calculate the player's movement trajectory.

[0056] The server extracts the following key features from a single frame's pose: the angles of joints such as arms and torso, the direction and slope of arms and legs, the displacement of key human body points between adjacent frames, and the distance and orientation of the human body to key areas such as the ball and goal.

[0057] Step 2: Input the motion posture data into the intention reasoning model constructed by the trained neural network, and output a multi-dimensional intention probability distribution vector. Each dimension of the multi-dimensional intention probability distribution vector corresponds to a preset hockey defensive action.

[0058] The server builds a trained intention reasoning neural network model, which is constructed using models such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or Transformers, to capture dynamic patterns of poses over time.

[0059] The model takes motion posture data as input and outputs a multidimensional intent probability distribution vector. Each dimension of the multidimensional intent probability distribution vector corresponds to a preset hockey defensive primitive action category, which includes at least: spacing control, stick grabbing, body blocking, blocking shots, chasing back to block passing lines, positioning in front of the goal, and forcing the wall. The training process employs at least one algorithm selected from behavior cloning, generative adversarial imitation learning, or hierarchical reinforcement learning.

[0060] Step 3: Based on the dominant intent in the multidimensional intent probability distribution vector and the motion posture data, generate the virtual defender's ice surface movement trajectory under ice hockey kinematic constraints.

[0061] Based on the dominant intent category in the multidimensional intent probability distribution vector and the hockey training data at the current moment, a virtual defender's ice surface movement trajectory is generated over a preset future time period. The generation of the ice surface movement trajectory is constrained by a hockey kinematics constraint model, which limits the virtual defender's movement ability on the ice surface to ensure that the generated trajectory conforms to the physical movement laws of real hockey players.

[0062] The kinematic constraint model of ice hockey is a mathematical model used to describe the motion of a virtual defender on ice. Its core lies in characterizing the defender's motion evolution through state variables (position, velocity, body orientation angle) and their differential equations, and imposing multiple constraints such as acceleration, turning angular velocity, speed limit, and posture switching. The most critical feature is that the extremely low coefficient of friction of the ice surface (about 0.03 to 0.06) leads to large gliding inertia and the inability to start and stop instantaneously. In addition, the defender is often in a backward gliding state where the body orientation is inconsistent with the gliding direction. This makes the model fundamentally different from land motion models, and can ensure that the generated virtual defender trajectory conforms to the physical motion laws of real ice hockey players.

[0063] Step four: Based on the ice surface movement trajectory and the geometric blocking algorithm, calculate the optimal blocking posture of the virtual defender at each moment. The geometric blocking algorithm is used to maximize the coverage area of ​​the virtual defender over the shooting angle.

[0064] Based on the ice surface trajectory, the server uses a geometric blocking algorithm to calculate the optimal blocking posture of the virtual defender at each moment. The geometric blocking algorithm includes: calculating the shooting angle formed by the line connecting the attacking player with the ball and the two goalposts; determining the desired position of the virtual defender on the ice surface trajectory, maximizing the coverage area of ​​the virtual defender's torso and stick on the projection plane perpendicular to the shooting direction of the shooting angle; and calculating the virtual defender's body orientation angle, center of gravity height, stick extension length, and stick edge angle based on the desired position.

[0065] Step 5: Based on the movement trajectory and blocking posture characteristics, synthesize a virtual defender.

[0066] Based on the extracted movement trajectories and blocking posture features of the defending players, the server synthesizes one or more virtual defenders with specific behaviors. The server maps the generated posture and position data onto the real-time captured footage to generate virtual defenders and drive them to perform corresponding actions.

[0067] Step 6: Based on the location, movement pattern, and blocking area of ​​the virtual defender, generate an offensive tactic targeting the virtual defender.

[0068] The server uses virtual defenders as the adversarial environment to generate or recommend offensive tactics. It quantifies the virtual defenders' positions, movement patterns, and blocking areas into dynamic threat fields, inertial prediction vectors, and coverage probability maps, respectively. Based on this, Monte Carlo tree search is used to conduct multi-step offensive and defensive game theory, selecting the optimal running routes and passing timings. Simultaneously, deep reinforcement learning is used to generate breakthrough action instructions targeting weaknesses in the defensive center of gravity. The final output tactics include running hot zone corrections and passing / shooting time windows, and dynamic replanning is achieved through rapid simulated adversarial scenarios, ensuring that the strategy always targets the real-time exposed gaps in the defense.

[0069] The aforementioned method utilizes key point sequences of human posture combined with stick tip tags to extract the defensive player's movement trajectory and blocking posture features, accurately depicting the player's realistic blocking behavior. Based on these features, a virtual defender with realistic movement characteristics and blocking range is synthesized, and appropriate offensive tactics are generated according to the virtual defender's position, movement pattern, and blocking area. On the one hand, this overcomes the limitation of simply relying on player coordinate modeling in identifying effective blocking areas, making the virtual defender's behavior more closely resemble the real game; on the other hand, it can dynamically generate differentiated offensive tactics for different defensive individuals, realizing intelligent offensive and defensive simulation and effectively supporting player tactical training and tactical plan verification.

[0070] In one embodiment, the target detection model is trained through the following steps: Step 1: Perform sparsity training on the YOLO v8n model by adding an L1 regularization term to the loss function. By applying an absolute value penalty to the convolution kernel weights, some convolution kernel weights are made to approach zero, thus obtaining a trained detection model.

[0071] The adjusted YOLO v8n model was trained with sparsity by adding an L1 regularization term to the loss function, causing the weights of some convolutional kernels to approach zero. Global threshold pruning was then applied: convolutional kernels with a contribution less than 1% of the global maximum were removed. The pruning ratio was approximately 30%, reducing the number of parameters from 3.2M to 2.2M. Fine-tuning was performed after pruning to restore accuracy, ultimately resulting in a mAP decrease of no more than 1.5%.

[0072] Step 2: Determine whether the detection accuracy of the trained detection model is greater than or equal to a preset accuracy threshold.

[0073] The server determines whether the detection accuracy of the trained detection model is greater than or equal to a preset accuracy threshold. The preset accuracy threshold is adaptively adjusted according to the current hardware computing power level. On edge GPUs, the threshold is relaxed to mAP≥0.90, and on server-grade GPUs, the threshold is increased to mAP≥0.94.

[0074] Step 3: If the detection accuracy is greater than or equal to a preset accuracy threshold, then the trained detection model is pruned to obtain a pruned detection model.

[0075] If the detection accuracy is greater than or equal to a preset accuracy threshold, the server performs structured pruning on the trained detection model to obtain a pruned detection model. The structured pruning uses individual convolutional kernels in the convolutional layers as pruning units, calculates the importance score of each convolutional kernel, and removes the lowest 10% to 30% of convolutional kernels. The pruning ratio is dynamically adjusted according to the current hardware computing power: an upper limit of 30% for edge GPUs and a lower limit of 10% for server-level GPUs. After pruning, the model undergoes short-cycle fine-tuning for no more than 50 epochs to compensate for accuracy loss. If the mAP of the fine-tuned model on the validation set decreases by more than 1% compared to before pruning, the model weights are reverted to their pre-pruning state, and the pruning ratio is readjusted until the accuracy constraint is met.

[0076] Step four: Optimize the pruned detection model using TensorRT to convert its accuracy from FP32 to INT8 quantization accuracy.

[0077] The server uses TensorRT to convert the FP32 model to INT8 quantization format and employs a layer-by-layer calibration method to reduce accuracy loss. The calibration set is used to statistically analyze the activation range of each layer to determine the quantization parameters. After quantization, the model size is compressed from approximately 13MB to approximately 4.5MB, and the inference speed on an iPhone 8 is reduced from 45ms / frame to 28ms / frame, achieving the ≥30fps requirement.

[0078] The above method, through sparse training, accuracy verification, model pruning and INT8 quantization optimization, significantly simplifies the model structure, reduces inference computing power consumption and latency while ensuring the model's detection accuracy. This achieves lightweight and high real-time performance of the detection model, making it more suitable for real-time detection scenarios on terminal devices and more practical for deployment.

[0079] In one embodiment, real-time training videos of the player to be tested are captured using a mobile terminal camera, and frame images are extracted from the real-time training videos based on timestamps to obtain real-time training images. This includes the following steps: Step 1: Collect real-time training videos of the players to be tested using the mobile terminal's camera.

[0080] Real-time training videos of the players being tested are captured using a mobile terminal camera mounted on a tripod, desktop, or wall-mounted stand. The camera should be 80-120 cm above the ground, with a downward angle of 15°-30°, ensuring the training area occupies approximately 60% of the central area of ​​the frame. The training area should ideally be a flat, hard surface such as wood flooring, rubber flooring, or short-pile carpet, avoiding highly reflective or dark, light-absorbing surfaces. Sufficient safety buffer space should be provided around the training area. Ambient light intensity should be no less than 200 lux, with no significant backlighting or dynamic light and shadow changes.

[0081] Step 2: Based on the viewing angle and shooting range of the mobile terminal camera, generate a rectangular calibration area that covers the motion area.

[0082] Images captured by mobile terminal cameras are two-dimensional perspective projections, exhibiting effects such as near objects appearing larger than distant ones and angular distortion. To convert the coordinates of the detected sphere center pixels in the image into planar coordinates in the real world, a homography mapping relationship must be established between the image plane and the actual training ground plane.

[0083] Based on the mobile terminal's camera field of view and the recommended working distance of 1.5 meters to 2.0 meters, the server generates a rectangular calibration area by default. This rectangular calibration area is 1.25 meters long and 1 meter wide. The size of this rectangular calibration area covers the user's arm extension and the range of motion of the golf club. Users can adjust the rectangular calibration area in the settings according to the actual available space.

[0084] Step 3: Collect the image coordinates of the four corners of the motion area in sequence.

[0085] The server draws four target points on the image, corresponding to the four vertices of the calibration area. The user holds the training ball and moves it sequentially to the actual position of each vertex, aligning the ball image with the target point on the screen and clicking "OK". Each time the user confirms, the server records the center pixel coordinates of the ball detection box in the current frame, using these coordinates as the image coordinates of that vertex. The server also knows the world coordinates of that vertex. To ensure accuracy, each vertex can be sampled 2-3 times, and the server takes the average to eliminate hand tremor errors.

[0086] Step four: Based on the image coordinates, obtain the spatial calibration mapping matrix using the direct linear transformation method, and perform condition number verification on the matrix.

[0087] Based on the four sets of corresponding point pairs collected, the server uses the direct linear transformation method to solve for the 3×3 perspective transformation matrix. This matrix describes the linear mapping relationship from image pixel coordinates to world plane coordinates and can be decomposed into rotation, scaling, shearing, and translation components. After the solution is completed, the server performs a condition number check on the matrix. If the condition number is too high, the user is prompted to recalibrate or adjust the region size.

[0088] Step 5: After the verification is passed, the spatial calibration accuracy is confirmed by verifying that the deviation of the virtual grid intersection is less than the preset threshold.

[0089] After calibration, the server overlays a virtual grid onto the image and guides the user to place the sphere near the grid intersections. If the deviation between the blue halo at the sphere's center and the grid intersection is less than 2 cm, the calibration is considered successful; otherwise, the source of the error is indicated and recalibration is suggested. During training, the server continuously monitors the device's gyroscope and accelerometer data. If a change in device angle exceeding ±2° or a displacement exceeding 1 cm is detected, training is automatically paused, and a notification of device position change is displayed, requesting recalibration. This mechanism ensures the effectiveness of the mapping relationship throughout the entire training session.

[0090] Step 6: Perform perspective transformation preprocessing on the real-time training video that has completed spatial calibration, and extract frame images according to timestamps to obtain real-time training images.

[0091] The server performs perspective transformation preprocessing on the real-time training video that has completed spatial calibration, and extracts frame images from the real-time training video according to the timestamp to obtain the real-time training image.

[0092] The above method requires no depth sensor; it utilizes only a monocular RGB image and a ground plane assumption to establish a mapping from 2D pixel coordinates to 2D world plane coordinates by solving the homography matrix. No professional calibration is needed; the user moves the ball sequentially to four target points on the screen and clicks, and the server automatically collects four sets of corresponding point pairs, solving the matrix using a direct linear transformation. During training, the device's gyroscope and accelerometer are continuously monitored; if the detected device displacement or angle change exceeds a threshold, a recalibration prompt is automatically triggered.

[0093] Ordinary users can use the above method to complete the calibration within 2 minutes on any flat surface such as the living room or garage using their own mobile phone, with a positional accuracy of ±2 cm, which meets the requirements of ice hockey training.

[0094] In one embodiment, the real-time training image is input into the target detection model, the detection result is output, and the detection result is associated with the same timestamp, including the following steps: Step 1: Input the real-time training image into the target detection model. The detection head outputs the sphere detection box and human pose key points in the image. The sphere detection box consists of bounding box coordinates and target confidence, and the human pose key points consist of pixel coordinates and visibility.

[0095] The server will train the image input to the object detection model in real time. The detection head of the object detection model will output the sphere detection box and human pose key points in the image. The sphere detection box consists of bounding box coordinates and object confidence, and the human pose key points consist of pixel coordinates and visibility.

[0096] The server performs an initial screening of the detector head output based on a confidence threshold, retaining candidate boxes with a confidence level greater than a preset lower limit. For pose keypoints, all keypoints are retained, and outliers are subsequently removed through temporal filtering.

[0097] Step 2: Merge redundant sphere detection boxes using NMS, sort them according to the target confidence level, retain the sphere detection box with the highest confidence level, and delete other sphere detection boxes that overlap with this sphere detection box by more than a threshold.

[0098] Since the model may generate multiple overlapping candidate boxes around the same sphere, the server merges redundant boxes using Non-Maximum Summarization (NMS). Preferably, the confidence threshold is set to 0.4; candidates below this value are discarded. The core logic of NMS is to sort by confidence from high to low, retain the box with the highest confidence, and then delete other boxes whose overlap area with that box exceeds a certain proportion. Because hockey moves rapidly and may have motion blur, the LoU threshold is set relatively low to retain candidate boxes that may be separated due to occlusion; preferably, the LoU threshold is 0.3. NMS is not performed on pose keypoints because keypoints of different parts of the human body do not overlap.

[0099] Step 3: Using a spatial calibration mapping matrix, the center coordinates of the retained sphere detection box and the pixel coordinates of the human pose key points are converted into real-world planar coordinates.

[0100] After the sphere detection bounding box is preserved by NMS, the server takes its center as the pixel coordinates of the sphere's center in the image. The server then uses an established spatial calibration mapping matrix to convert these pixel coordinates into real-world planar coordinates. The converted coordinates can be used to calculate the sphere's position, velocity, and direction of movement. For attitude keypoints, the same mapping matrix is ​​used to convert them to the world coordinate system, providing a unified plane for subsequent calculations.

[0101] Step four: Apply the Kalman filter algorithm to the center coordinates of the sphere detection box to obtain smooth world coordinates of the sphere center. Apply the moving average filter to the coordinates of the human body pose key points to obtain smooth coordinates of the human body pose key points.

[0102] The original detection results may exhibit jitter, jumps, or brief loss. To obtain a continuous and stable trajectory, the server employs Kalman filtering on the ball's center coordinates. Kalman filtering uses the position and velocity of the previous frame to predict the position of the current frame, then uses the model's detection results as observations for correction, ultimately outputting a smoother estimate. Even if a frame is lost, the filter can still make short-term predictions based on the motion model, ensuring uninterrupted feedback. When more than 5 frames are lost, the server determines the ball is out of control and resets the filter state. For attitude keypoints, moving average filtering is used to reduce joint position jitter caused by rapid movements.

[0103] If no ball is detected for 2-5 consecutive frames, the server does not immediately issue an alarm but continues to display the virtual device using Kalman scattering to predict the position and continues voice training guidance. If the prediction fails to recover after more than 5 frames, a message indicating that the ball is out of control is displayed. If other people or pets are present in the scene, the model may falsely detect multiple balls or multiple poses. The server maintains tracking by selecting the detection result with the highest confidence and closest to the last position, and highlights the selected target on the image.

[0104] Step 5: Associate the smoothed world coordinates of the sphere's center with the coordinates of the key points of the human pose to the same timestamp.

[0105] After obtaining the smoothed world coordinates of the ball's center and the coordinates of the human body's keypoints, the server associates them with the same timestamp for use by the quantitative evaluation module. The core of this association is ensuring that the ball and the human body belong to the same trainee. Since this server currently only supports single-user operation, complex matching is unnecessary; it only needs to confirm that the distance between the ball's center and any keypoint is within a certain range. Otherwise, it may be a false detection or the user not holding the ball. Furthermore, the server generates a data validity flag for each frame. If the ball's center coordinates exceed the calibration area boundary, the keypoint visibility is too low, or the filter diverges, the frame is marked as invalid and not included in the score.

[0106] The above method integrates, filters, transforms coordinates, and performs preliminary analysis on the original detection results in each frame of the image output by the target detection model, providing a clean and continuous data stream for subsequent quantitative evaluation and real-time feedback. It also integrates the sphere detection box and human pose key points output by the target detection model into a unified world coordinate system for temporal smoothing and correlation.

[0107] In one embodiment, the detection results are input into an evaluation and analysis model according to the time sequence of timestamps to obtain hockey quantification indicators. The evaluation and analysis model calculates these indicators based on the geometric relationship between the puck's trajectory and the player's posture, and is used to evaluate the technical execution quality of the player being tested. This includes the following steps: Step 1: Align the sphere detection frame sequence, the human posture key point sequence, and the club end sequence according to the timestamp order.

[0108] The server will associate the detection results with the same timestamp with a sequence of sphere detection boxes, a sequence of human pose key points, and a sequence of coordinates at the end of the cue stick, organized in chronological order of image acquisition time.

[0109] Step two: Input the aligned detection results into the evaluation and analysis model; The server inputs the aligned detection results frame by frame into the pre-trained evaluation and analysis model. The pre-trained evaluation and analysis model is a temporal multi-task deep network for hockey training. The server inputs spatiotemporally aligned standard hockey training video data into the model. The standard hockey training video is a pre-recorded sports training video of a professional hockey player, including the athlete's three-dimensional skeletal key points, joint angles, spatial coordinates of the hockey stick and hockey puck, and temporal features of the contact state of the ice skates during the movement.

[0110] Step 3: Evaluate and analyze the model to construct the ball's trajectory, the human posture sequence, and the trajectory of the club's end point; The evaluation and analysis model is based on the kinematic biomechanical geometric relationship between the sphere detection box sequence, the human posture key point sequence, and the club end coordinate sequence. The spatial motion trajectory of the sphere is generated by fitting the center coordinates of the sphere detection box from multiple consecutive frames; the temporal posture changes of the torso, upper limbs, and lower limbs are constructed based on the human posture key point sequence; and the club end trajectory is generated based on the club end coordinate sequence.

[0111] Step four: Based on the ball's trajectory, the preset key points of the human skeleton, and the spatial coordinates of the end of the hockey stick, the evaluation and analysis model calculates the hockey quantitative indicators of the ball relative to the player, the stick face, and the human posture to evaluate the quality of the athlete's technical execution.

[0112] The server calculates hockey quantification metrics based on the puck's current position, the player's posture, and the trajectory of the stick's end. These metrics include dribbling skills assessment, puck protection skills assessment, puck protection skills assessment, breakthrough techniques assessment, and toe-heel flicking techniques assessment.

[0113] The core of the dribbling technique assessment focuses on the rhythm and spatial stability of the cue stick's control over the ball, primarily calculating the contact frequency, trajectory deviation error, and ball control stability score. The contact frequency is calculated by detecting the rate of change in the ball's velocity direction; a valid contact occurs when the angle of change in the ball's velocity direction exceeds a certain value. The server counts the number of contact times per unit time and compares it to a preset rhythm target. A low frequency prompts for faster ball movement, while a high frequency prompts for controlled rhythm. In straight or curved dribbling training, the server generates an ideal path, such as a straight line from the left starting point to the right ending point. The vertical distance from the ball's center to this path is calculated in real time, and the average deviation throughout the training process is calculated to obtain the trajectory deviation error. The smaller the deviation, the higher the ball control accuracy. When the deviation exceeds 10 centimeters, a voice prompt indicates the ball has deviated from its path. A ball control stability score from 0 to 100 is given by combining the success rate of the contact frequency and the average trajectory deviation. This score also considers the smoothness of the movement; the more even the intervals, the higher the score.

[0114] The core of the ball protection technique assessment focuses on the body's ability to shield the ball, i.e., the ability to prevent the defender from stealing it. The server uses the midpoints of the shoulders and hips to determine the torso's orientation. In ball protection training, the ideal posture is to face the defender sideways, keeping the ball away from the defender. The server calculates the angle between the torso's direction and the direction of the incoming defense to obtain the body's facing angle. The exposure angle is the geometric relationship between the ball's center, the torso's facing angle, and the defensive point. The smaller the exposure angle, the better the ball is shielded by the body; when the exposure angle exceeds a certain value, the ball is exposed to the defender's side and easily stolen. The server provides real-time voice commands such as "turn left" and "turn right" based on the exposure angle. In mobile ball protection training, the defensive point simulates the movement of a real defender. The server continuously tracks the rate of change of the exposure angle and dynamically adjusts the ball protection commands. If the exposure angle increases rapidly, the user is prompted to accelerate their positioning adjustment.

[0115] The core evaluation of puck protection techniques focuses on the deceptiveness of the movement, i.e., whether the opponent can be misled. The server comprehensively judges the deceptive strength by analyzing the velocity direction before the change of direction, the magnitude of the acceleration change during the change of direction, and the shift in the body's center of gravity. A high-quality feint should first show a clear movement in one direction, then suddenly change direction, with a genuine shift in the center of gravity. The server calculates the peak velocity during the deception phase and the peak acceleration during the change of direction; the higher the ratio, the stronger the deception. The feint needs to be smooth and natural, without any abrupt pauses. The server analyzes the smoothness of joint angle changes, such as the hip and shoulder joints, throughout the movement sequence to calculate the smoothness of the movement. If there is a sudden turn, the smoothness score is reduced, and the smoothness score accounts for 30% of the total feint score. The server has built-in multiple standard feint templates recorded by professional hockey players, such as body feints and puck pulls, and provides a completion percentage through similarity matching of the movement sequences.

[0116] The core evaluation of breakthrough techniques focuses on the quality of tactical execution and speed changes. A key aspect of breakthroughs is the rapid reverse acceleration after a sudden stop. The server measures the time interval from the start of the sudden stop to the reverse acceleration and the magnitude of the reverse acceleration to determine the efficiency of the sudden stop and change of direction. Generally, a delay of less than 0.2 seconds and an acceleration greater than a certain threshold are considered excellent. The server breaks down the breakthrough action into three phases: sudden stop, change of direction, and acceleration, evaluating the cleanliness of each phase. For example, the sudden stop phase requires the ball speed to drop to near zero in a short time, the change of direction phase requires a significant change in direction without losing possession, and the acceleration phase requires a rapid start. If any phase fails to meet the standards, it will be clearly indicated in the report. In tangential breakthroughs, the server checks whether the user successfully cuts into the target direction without exceeding the calibration boundary. Touching the boundary is considered a failed breakthrough, and points for that item are deducted.

[0117] The toe-heel flick technique assessment focuses on the coordinated ball control ability between the foot and the cue stick, and is considered an advanced technique. A valid foot contact event is defined as the distance between the ankle and the center of the ball is less than a certain threshold, and the ball speed changes significantly in the next frame. The server distinguishes between the left and right feet, recording their usage frequency separately. In rhythm training, users are required to touch the ball with their feet according to a metronome. The server calculates the deviation between the actual contact time and the theoretical beat time, calculates the deviation value, and also records the evenness of foot alternation. The magnitude of the ball speed change before and after foot contact reflects the precision of control. Ideally, foot contact should cause the ball to move at a constant speed or smoothly change direction, rather than bouncing violently. The server calculates the standard deviation of the ball speed before and after contact; a smaller standard deviation indicates more precise control. This indicator is particularly important for advanced users.

[0118] The above method introduces five exclusive indicators, including exposure angle, deception intensity, emergency stop and change of direction efficiency, foot contact event detection, and clubface orientation deviation angle. It fills the gaps in the evaluation of existing technologies in dimensions such as body shielding of the puck, deceptiveness of feints, breakdown of action phases, hand-foot coordination of puck control, and accuracy of hitting the puck, and achieves a comprehensive, refined, and quantitative assessment of the execution quality of ice hockey techniques.

[0119] In one embodiment, adaptive ball recognition is further included, comprising the following steps: Step 1: Obtain the pre-processed image captured by the mobile terminal camera, and extract the shape, size, and color of the sphere in the image.

[0120] The server acquires preprocessed images captured by the mobile terminal's camera, locates candidate sphere regions in the images, and extracts the ellipse fitting shape parameters, physical size estimates based on camera intrinsic back projection, and black-white ratio and hue histogram in HSV space as multi-dimensional feature descriptions of the sphere.

[0121] Step 2: Input the shape, size, and color of the sphere into the pre-trained classification model and output the sphere category identifier.

[0122] The server concatenates the sphere's shape features (ellipse-fitted shape parameters), size features (back-projected physical dimensions), and color features (HSV space black-white ratio and hue histogram) extracted in step one into a multi-dimensional feature vector, which is then input into a pre-trained classification model. The classification model is built on a lightweight convolutional neural network (such as MobileNetV3-Small) and outputs the sphere's class identifier. The class identifier includes at least ice hockey, tennis, golf, and hockey.

[0123] The classification model has a built-in table of sphere feature parameters, which contains the following key fields.

[0124] Step 3: Perform adaptive operations based on the identified ball category identifier.

[0125] Because the YOLO v8 model itself is robust to color, it does not forcibly change the RGB channels of the input image during normal use, but color enhancement options can be enabled in extreme scenarios. The server performs histogram equalization or contrast stretching on the image before it is fed into the model, based on the typical color range of the selected ball type, to improve the distinction between the target and the background. This operation does not change the model weights; it only performs real-time filtering on the input frames, with an overhead of less than 2ms / frame.

[0126] The lightweight YOLO v8 model has a fixed input of 640×640 pixels, and the original camera capture resolution is usually 1280×720. When scaling down, if the diameter of the sphere is too small, the server uses a high-resolution ROI cropping strategy. It first crops a 960×960 pixel area centered on the sphere's previous position, and then scales it down to 640×640 pixels to increase the pixel ratio of small objects.

[0127] Step 4: Dynamically modify the anchor frame size of the lightweight YOLO v8 model by scaling the preset anchor frame proportionally according to the actual diameter corresponding to the ball category.

[0128] The YOLO v8 model predefines multiple anchor boxes at each scale of the output feature map to regress target bounding boxes. The standard anchor box size is statistically derived from the training sample dataset. For small targets such as ice hockey pucks, the default anchor boxes may be too large. After the model is loaded, the server dynamically modifies the anchor box parameters based on the type of ball. The anchor box size is scaled proportionally according to the diameter of the identified ball. For example, the standard ice hockey puck anchor box is 40×40 pixels, but when switching to a golf ball, the anchor box size is adjusted to 22×22 pixels. This modification only requires replacing the anchor box list in the model configuration file; recompiling the model is not necessary.

[0129] Step 5: Adjust the confidence threshold and use the benchmark confidence threshold corresponding to the ball category to filter the detection results.

[0130] Different types of balls exhibit varying stability in their appearance features within images. For instance, the dents on a golf ball's surface can create irregular highlights, potentially causing fluctuations in detection confidence. The server pre-sets baseline confidence thresholds for each ball type: 0.4 for standard ice hockey, 0.5 for tennis balls, and 0.3 for golf balls. The server selects the confidence threshold based on the identified ball type, and users can also manually adjust the confidence threshold.

[0131] Step 6: Set the physical size conversion factor, verify the consistency of the conversion of the pixel width of the detection frame after perspective transformation based on the actual diameter of the ball category, and trigger calibration or prompt when the deviation exceeds the preset range.

[0132] After the pixel width of the detection box is mapped to world coordinates through perspective transformation, the server needs to verify the consistency of the conversion based on the actual diameter of the sphere. If the converted diameter of the detection box deviates from the preset diameter by 20%, it is judged as a detection anomaly, triggering temporary calibration or prompting the user to check the lighting. This mechanism can effectively avoid false positives caused by misidentification of background objects.

[0133] The above method, through preprocessing adaptation and postprocessing parameter dynamic adjustment, enables a single YOLOv8 model to stably recognize multiple types of balls without retraining.

[0134] In one embodiment, the step of inputting the hockey quantification index into a multimodal feedback model, wherein the multimodal feedback model compares the hockey quantification index with preset standard values ​​to generate exercise guidance information, includes the following steps: Step 1: The multimodal feedback model acquires the hockey quantification index and preset standard value, where the preset standard value corresponds to the optimal range of hitting parameters for hockey motion.

[0135] The multimodal feedback model acquires quantitative indicators for ice hockey and pre-stored preset standard values. The preset standard values ​​correspond to the optimal execution range of various key technical parameters in ice hockey, and are derived from: the mean and standard deviation obtained from large-scale motion capture data of national and above-level athletes; ideal execution parameter thresholds calibrated by a senior coaching team based on the principles of sports biomechanics; and target reference values ​​obtained through iterative optimization of the ice hockey dynamics simulation model.

[0136] The preset standard values ​​should include at least the following baseline ranges: beginner ball contact frequency ≥ 80 times / minute, exposure angle in ball protection state: 15°~30°, deception intensity: ≥ 2.5 (quantitative threshold for the effectiveness of feints), emergency stop change of direction efficiency: ≤ 0.4 seconds (time interval from emergency stop to reverse start), clubface orientation deviation angle: ≤ 5° (angle between the clubface and the target direction at the moment of impact), foot contact frequency: ≥ 40 times per minute (high-intensity ball control training standard).

[0137] Step 2: Calculate the deviation between the ice hockey quantitative index and the preset standard value.

[0138] The multimodal feedback model compares the current frame's hockey quantization index vector with a preset standard value vector dimension by dimension, calculating the deviation value for each index. The deviation value includes: Absolute deviation: The difference between the measured value of an indicator and the median of the standard value; Relative deviation: Absolute deviation divided by the width of the standard value range; Exceeding limits flag: If the measured value exceeds the upper or lower bound of the standard value range, it is marked as abnormal.

[0139] Step 3: Based on the deviation value, generate guidance information, which includes at least one of the following: action correction prompts, force adjustment suggestions, and angle optimization suggestions.

[0140] The multimodal feedback model converts the deviation value into user-perceptible guidance information, which is then transmitted simultaneously through both voice and visual channels.

[0141] Voice feedback is not played in every frame; instead, it's event-driven. The server has built-in trigger rules that generate a voice command when a quantifiable metric meets a trigger condition for a given period of time. Common rules include: Deviation accumulation trigger: When a certain indicator is lower than the set deviation value for 3 consecutive times, the command to increase the ball speed is triggered.

[0142] Instantaneous cross-boundary trigger: When the exposure angle exceeds the safe value and the duration exceeds 0.3 seconds, the turn to protect the ball command is immediately triggered.

[0143] Success incentive trigger: When a user completes a difficult action, an encouragement command is generated.

[0144] Phase reminder: During rhythm training, directional words such as left, right, and forward are played before and after each beat to guide the movement.

[0145] The server employs an offline TTS engine, pre-compiling frequently used phrases into audio files for caching. For dynamically generated instructions, the TTS engine synthesizes them in real-time. To ensure real-time feedback, the total latency from indicator triggering to voice playback is controlled within 200 milliseconds. This is achieved through audio file preloading, playback queue priority management, and the use of low-latency audio APIs. Frequently used phrases can include action commands such as "start," "shift left," and "change direction"; real-time error correction such as "wrist too stiff," "lower center of gravity," and "speed up the pace"; and encouraging phrases such as "good shot," "perfect," and "one more time."

[0146] Visual feedback includes trajectory heatmaps, virtual defenders, bullseye and path guidance, and status indicators. All visual feedback is overlaid on real-time training images.

[0147] The ball trajectory heatmap is generated by the server recording the historical position of the ball's center in the last 30 frames and drawing semi-transparent trajectory points on the real-time training images. The color of each point gradually changes from blue to red, forming the heatmap effect. The heatmap visually displays the dense areas of ball control. If the red area is concentrated in a corner, it indicates that the user's ball control range is too narrow; if the trajectory is scattered and exceeds the calibration box, it indicates insufficient ball control stability.

[0148] In ball protection training and high-pressure challenge modes, the server generates a virtual defender. The virtual defender's position and cueing direction are dynamically determined by the server based on the exposure angle. When the exposure angle exceeds a warning value, the virtual defender will appear in the exposed direction on the real-time training image and perform a ball-grabbing animation. The virtual defender's speed and attack frequency are controlled by an adaptive difficulty mechanism. Trainees need to adjust their body orientation according to the virtual defender's movements, ensuring the virtual defender is behind them. If the virtual defender successfully touches the ball, the defense is considered inadequate.

[0149] In Rapid Breakthrough Training and Custom Training, the server draws guiding elements on the live training image. A circular bullseye is drawn on the live training image, and the player needs to dribble the ball into the bullseye. After hitting the bullseye, an animation shows the bullseye transforming into a hit, and a new bullseye appears at a random position. The server draws a white or green semi-transparent curve on the live training image, indicating the trajectory the ball should follow. The player needs to keep the ball close to the path line; if they deviate, the path line turns red. In phased training, a series of checkpoints are added to the live training image, and the player needs to dribble the ball to each checkpoint in sequence.

[0150] The server displays brief information about the current training mode in the corner of the real-time training image: remaining time, current action name, and number of times the target has been completed. It also displays the server's recognition status: green indicates the ball is being stably tracked, yellow indicates unstable tracking, and red indicates the ball has been lost. If the ball is lost, a message will appear on the screen asking the user to move it back to the recognition area.

[0151] Voice and visual feedback do not work independently but rather in tandem, forming redundant encoding to ensure that users receive information even if they miss one modality. The collaborative strategies include: Simultaneous presentation: When an important event occurs, voice announcement and visual highlighting happen simultaneously to enhance perception.

[0152] Complementary presentation: For spatial information, the virtual defender's position is displayed visually, while the voice directly instructs the player to turn left; for rhythmic information, a progress bar is displayed visually, while the beat points are announced via voice.

[0153] Priority arbitration: When multiple events occur simultaneously, the system determines which one to broadcast first based on a preset priority order. For example, security alarms have higher priority than rhythmic alerts. Users can also adjust the priority of each event in the settings.

[0154] During training, you can also choose to enable only the pure audio mode. In this mode, on-screen prompts are disabled, and all feedback is delivered via voice. Pure audio mode is suitable for scenarios where users want to focus on sensory feedback and reduce visual distractions. In pure audio mode, the server still performs video analysis, but only plays voice prompts at key moments, and all visual feedback elements are hidden.

[0155] The above method provides low-latency, high-definition, and non-intrusive feedback while the user focuses on hockey control movements, enabling errors to be identified and corrected immediately, while maintaining the continuity and immersion of training.

[0156] In one embodiment, dynamic adjustment of training parameters is also included, comprising the following steps: Step 1: Statistically analyze the performance metrics within a fixed time window. The performance metrics include at least one of the following: success rate, average reaction time, number of consecutive successes, or number of consecutive failures.

[0157] The server uses a fixed time window to collect user performance metrics within that window. For example, a window length of 10 seconds allows for timely responses to changes while avoiding drastic fluctuations caused by a single, accidental event. Performance metrics include, but are not limited to: success rate = number of successful ball protections / total number of defensive plays (or number of successful shots / total number of attempts), average reaction time (the time from when a defensive play occurs to when the user makes an effective move), and consecutive successes / failures.

[0158] Step 2: Calculate the success rate of the current window based on the performance indicators, and dynamically adjust the training parameters according to the success rate. The training parameters include at least one of the following: virtual defender speed, reaction time tolerance, and task duration.

[0159] The server calculates the success rate of the current window based on performance metrics. Based on this success rate, the server adopts the following strategies: If the success rate is greater than 75%, it indicates the current mode's challenge is too easy, and the server appropriately increases the difficulty, for example, by increasing the speed of the virtual defender by one step or shortening the reaction time tolerance; if the success rate is less than 65%, it indicates the current mode's challenge is too difficult, and the server reduces the difficulty, for example, by decreasing the speed of the virtual defender by one step or extending the task time; if the success rate is between 65% and 75%, it indicates the current mode's challenge is appropriate, and the current difficulty is maintained without adjustment.

[0160] To ensure smoothness, the server waits at least 5 seconds after each adjustment before conducting the next evaluation to avoid frequent oscillations. Furthermore, the adjustment increments gradually decrease: the further away from the target success rate, the larger the step size per adjustment; the closer to the target, the smaller the step size. When the difficulty parameter approaches the preset maximum or minimum value, the server halves the adjustment step size to provide a buffer. The maximum or minimum value must never be exceeded to prevent situations where there is no solution or the difficulty is too easy. If a user's success rate is below 40% for three consecutive windows, the server determines that the user may be fatigued or there may be an abnormal environment, automatically pausing the challenge and prompting a rest or recalibration.

[0161] The server records the final difficulty parameters of the last 10 high-pressure challenges and takes the median as the initial speed for the next challenge. Thus, the server will start based on the highest difficulty recorded last time; if no record is available, it will start from the default medium difficulty. Furthermore, the server also records the user's performance differences across different skills, using different initial difficulty coefficients in the corresponding modes. For example, if a user has strong ball protection skills but weak dribbling, the initial speed for the ball protection challenge can be set to 1.3 m / s, and for the dribbling challenge, it can be set to 0.9 m / s.

[0162] The above method introduces an adaptive difficulty mechanism, which dynamically adjusts the speed, attack frequency, or task threshold of the virtual opponent based on the user's real-time performance, keeping the user in a state of flow.

[0163] In one embodiment, such as Figure 2 As shown, an ice hockey training device based on ice hockey quantitative indicators includes a sample acquisition module 201, a model improvement module 202, an image acquisition module 203, an image processing module 204, an evaluation and analysis module 205, and a feedback module 206.

[0164] The sample acquisition module 201 is used to acquire multiple historical ice hockey game video data to obtain a training sample dataset. Each historical ice hockey game video data includes an labeled ice hockey puck and key points of the player's body posture and the end of the hockey stick corresponding to each player. The model improvement module 202 is used to learn and process the training sample dataset using a lightweight YOLO v8n model to obtain an object detection model. The lightweight YOLO v8n model is obtained by pruning redundant channels of the YOLO v8n model using a pruning algorithm. The image acquisition module 203 is used to acquire real-time training videos of the player to be tested through the mobile terminal camera, and extract frame images from the real-time training videos according to the timestamps to obtain real-time training images. Image processing module 204 is used to input the real-time training image into the target detection model, output the detection result, and associate the detection result with the same timestamp; The evaluation and analysis module 205 is used to input the detection results into the evaluation and analysis model according to the time sequence of the timestamp to obtain the hockey quantitative index. The evaluation and analysis model is based on the geometric relationship between the hockey trajectory and the human posture to calculate the hockey quantitative index in order to evaluate the technical execution quality of the player under test. Feedback module 206 inputs the ice hockey quantitative index into a multimodal feedback model, which compares the ice hockey quantitative index with a preset standard value to generate exercise guidance information.

[0165] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An ice hockey training method based on quantitative indicators of ice hockey, characterized in that, include: Multiple historical ice hockey game video data are acquired to obtain a training sample dataset. Each historical ice hockey game video data includes an annotated ice hockey puck and key points of the player's body posture and stick end corresponding to each player. A lightweight YOLO v8n model is used to perform frame image learning processing on the training sample dataset to obtain a target detection model for recognizing the hockey puck, player posture, and stick tip. The lightweight YOLO v8n model is obtained by pruning redundant channels of the YOLO v8n model using a pruning algorithm. Real-time training videos of the players to be tested are captured by the mobile terminal camera, and real-time training images are obtained by extracting frame images from the real-time training videos according to the timestamps. The real-time training image is input into the target detection model, the detection result is output, and the detection result is associated with the same timestamp. The detection results are input into the evaluation and analysis model according to the time sequence of the timestamp to obtain hockey quantitative indicators. The evaluation and analysis model is based on the geometric relationship between the hockey trajectory and the human posture to calculate hockey quantitative indicators in order to evaluate the technical execution quality of the player under test. The quantitative indicators of ice hockey are input into a multimodal feedback model, which compares the quantitative indicators of ice hockey with preset standard values ​​to generate sports guidance information.

2. The ice hockey training method based on quantitative indicators of ice hockey according to claim 1, characterized in that, Also includes: Based on the sequence of key points of the defensive player's human posture and the label at the end of the cue stick in the training sample dataset, the motion posture data of the defensive player is extracted. The motion posture data is input into the intention reasoning model constructed by the trained neural network, and a multi-dimensional intention probability distribution vector is output. Each dimension of the multi-dimensional intention probability distribution vector corresponds to a preset hockey defensive action. Based on the dominant intent in the multidimensional intent probability distribution vector and the motion posture data, a virtual defender's ice surface movement trajectory is generated under ice hockey kinematic constraints. Based on the ice surface movement trajectory and the geometric blocking algorithm, the optimal blocking posture of the virtual defender at each moment is calculated. The geometric blocking algorithm is used to maximize the coverage area of ​​the virtual defender over the shooting angle.

3. The ice hockey training method based on quantitative indicators of ice hockey according to claim 1, characterized in that, The target detection model is trained through the following steps: The YOLO v8n model is trained in a sparsity manner by adding an L1 regularization term to the loss function and applying an absolute value penalty to the convolution kernel weights to make some convolution kernel weights approach zero, thus obtaining a trained detection model. Determine whether the detection accuracy of the trained detection model is not less than a preset accuracy threshold; When the detection accuracy is not less than a preset accuracy threshold, the trained detection model is pruned to obtain a pruned detection model. The pruned detection model was optimized using TensorRT, converting its accuracy from FP32 to INT8 quantization accuracy.

4. The ice hockey training method based on quantitative indicators of ice hockey according to claim 1, characterized in that, Real-time training videos of the players to be tested are captured using a mobile terminal camera. Frame images are extracted from the real-time training videos based on timestamps to obtain real-time training images, including: Real-time training videos of the players being tested are captured using the camera on a mobile device. Based on the viewing angle and shooting range of the mobile terminal camera, a rectangular calibration area covering the motion area is generated; The image coordinates of the four corners of the motion area are acquired sequentially; Based on the image coordinates, a spatial calibration mapping matrix is ​​obtained using the direct linear transformation method, and the matrix is ​​then subjected to condition number verification. After the verification is passed, the spatial calibration accuracy is confirmed by verifying that the deviation of the virtual grid intersections is less than the preset threshold. Perspective transformation preprocessing is performed on the real-time training video that has completed spatial calibration, and frame images are extracted according to timestamps to obtain real-time training images.

5. The ice hockey training method based on quantitative indicators of ice hockey according to claim 1, characterized in that, The step of inputting the real-time training image into the target detection model, outputting the detection result, and associating the detection result with the same timestamp includes: The real-time training image is input into the lightweight YOLOv8 model, and the detection head outputs the sphere detection box and human pose key points in the image. The sphere detection box consists of bounding box coordinates and target confidence, and the human pose key points consist of pixel coordinates and visibility. Redundant sphere detection boxes are merged using NMS, sorted according to the target confidence level, and the sphere detection box with the highest confidence level is retained while other sphere detection boxes that overlap with this sphere detection box by more than a threshold are deleted. The center coordinates of the retained sphere detection box and the pixel coordinates of the human pose key points are converted into real-world planar coordinates using a spatial calibration mapping matrix. The Kalman filter algorithm is used to obtain smooth world coordinates of the sphere center for the detection box, and the moving average filter is used to obtain smooth coordinates of the human body pose key points. Link the smoothed world coordinates of the sphere's center and the coordinates of key points in the human pose to the same timestamp.

6. The ice hockey training method based on quantitative indicators of ice hockey according to claim 1, characterized in that, The detection results are input into the evaluation and analysis model according to the time sequence of the timestamps to obtain hockey quantitative indicators. The evaluation and analysis model calculates hockey quantitative indicators based on the geometric relationship between the puck trajectory and the human posture, which are used to evaluate the technical execution quality of the player under test, including: Align the sphere detection frame sequence, human posture key point sequence, and club end sequence according to the timestamp order; The aligned detection results are then input into the evaluation and analysis model. The evaluation and analysis model constructs the ball's motion trajectory, the human posture sequence, and the trajectory of the club's end effector. Based on the ball's trajectory, pre-defined key points of the human skeleton, and the spatial coordinates of the stick tip, the evaluation and analysis model calculates quantitative indicators of the puck relative to the player, the stick face, and the player's posture to assess the quality of the athlete's technical execution. The quantitative indicators for ice hockey include at least the following parameters: The ball contact frequency is calculated based on the rate of change of the ball's velocity direction; the deviation of the ball's trajectory from the preset ideal path is calculated; the torso direction is determined by the human body posture; and the angle between the torso direction and the direction of the virtual defender's attack is calculated.

7. The ice hockey training method based on quantitative indicators of ice hockey according to claim 1, characterized in that, It also includes adaptive ball recognition: Acquire preprocessed images captured by the mobile terminal camera, and extract the shape, size, and color of the sphere in the images; The shape, size, and color of the sphere are input into a pre-trained classification model, which outputs a sphere category identifier. An adaptive operation is performed to dynamically adjust the anchor frame size, confidence threshold, and physical size conversion coefficient based on the identified ball category identifier and the actual diameter of the ball.

8. The ice hockey training method based on quantitative indicators of ice hockey according to claim 1, characterized in that, The step involves inputting the ice hockey quantification index into a multimodal feedback model, which compares the ice hockey quantification index with preset standard values ​​to generate exercise guidance information, including: The multimodal feedback model acquires the hockey quantification index and preset standard value, the preset standard value corresponding to the optimal hitting parameter range of hockey motion; Calculate the deviation between the quantitative index of ice hockey and the preset standard value; Based on the deviation value, guidance information is generated, which includes at least one of the following: action correction prompts, force adjustment suggestions, and angle optimization suggestions.

9. The ice hockey training method based on quantitative indicators of ice hockey according to claim 1, characterized in that, It also includes dynamic adjustment of training parameters: The performance metrics within the current window are statistically analyzed using a fixed time window. The performance metrics include at least one of the following: success rate, average reaction time, number of consecutive successes, or number of consecutive failures. The success rate of the current window is calculated based on the performance indicators, and the training parameters are dynamically adjusted according to the success rate. The training parameters include at least one of the following: virtual defender speed, reaction time tolerance, and task duration.

10. An ice hockey training device based on quantitative indicators of ice hockey, characterized in that, The device includes: The sample acquisition module is used to acquire multiple historical ice hockey game video data to obtain a training sample dataset. Each historical ice hockey game video data includes an labeled ice hockey puck and key points of the player's body posture and the end of the hockey stick corresponding to each player. The model improvement module is used to learn and process the training sample dataset using a lightweight YOLO v8n model to obtain an object detection model. The lightweight YOLO v8n model is obtained by pruning redundant channels of the YOLO v8n model using a pruning algorithm. The image acquisition module is used to acquire real-time training videos of the players to be tested through the mobile terminal camera, and extract frame images from the real-time training videos according to the timestamps to obtain real-time training images. The image processing module is used to input the real-time training image into the target detection model, output the detection result, and associate the detection result with the same timestamp; The evaluation and analysis module is used to input the detection results into the evaluation and analysis model according to the time sequence of the timestamp to obtain hockey quantitative indicators. The evaluation and analysis model is based on the geometric relationship between the hockey trajectory and the human posture to calculate hockey quantitative indicators in order to evaluate the technical execution quality of the player under test. The feedback module inputs the ice hockey quantitative indicators into a multimodal feedback model, which compares the ice hockey quantitative indicators with preset standard values ​​to generate exercise guidance information.

Images