Ice hockey movement collection sensing and strategy prediction method applied to smart viewing

Abstract

This invention relates to a method for collecting, sensing, and predicting curling motions in smart spectating. The method includes the following steps: fusing images of the same time frame from a single curling throw video obtained from cameras at multiple locations to obtain the current frame observation coordinates of each curling target; performing trajectory matching for each curling target based on the video frame order to obtain the motion trajectories of all curling stones during the current throw, which serves as the curling motion collection and sensing result; acquiring the current actual curling state; and combining an expert strategy method and an improved curling strategy prediction model to obtain the most likely action corresponding to the current actual curling state, which serves as the strategy prediction result. Compared with existing technologies, this invention has advantages such as improving the accuracy of predictions for the last few ends of a curling match and improving the accuracy of curling image recognition.

Description

Technical Field

[0001] This invention relates to the technical field of curling prediction and analysis, and in particular to a method for curling data acquisition, perception, and strategy prediction applied to smart spectating. Background Technology

[0002] Among current curling competition prediction solutions, CN202111107286.X (A Reinforcement Learning-Based Method for Curling Scene Analysis and Decision Support) discloses a reinforcement learning-based method for curling scene analysis and decision support. This method constructs a digital model of the curling competition scenario and situation, and establishes a corresponding decision support system. The system mainly includes three modules: a curling competition situation awareness module for real-time acquisition of information such as the position, speed, and stationary state of the curling stones; a curling rink digitization extraction module for accurately determining the position and type of the curling stones at key moments through position mapping between the rink and captured data; and a curling competition decision analysis module that, based on the above information, uses reinforcement learning algorithms to perform simulation calculations and outputs suggested striking positions for the next move to assist tactical decision-making.

[0003] CN202010435770.4 (A Method for Estimating the Motion State of a Curling Pole) discloses an image processing-based method for estimating the motion state of a curling pole, focusing on improving the accuracy of perceiving the state of the curling pole and the angle of the handle. It utilizes a curling pole target detection network, an angle detection network, and a target tracking network to detect and track the curling pole, obtaining its center coordinates. Finally, through coordinate transformation, the center coordinates and angles in the image coordinate system are converted to the curling pole coordinates and angles in the arena coordinate system.

[0004] CN202110774457.8 (A Curling Strategy Generation Method Based on Monte Carlo Reinforcement Learning) discloses a curling strategy generation method based on Monte Carlo reinforcement learning, addressing the problem of strategy decision-making when it is difficult to obtain effective datasets. It utilizes a policy value network model and the value network to obtain initial actions, combined with an improved Monte Carlo tree search algorithm. The policy value network is further updated using data from the self-game phase.

[0005] CN202410066718.4 (An Automated Learning and Decision Search Method for Curling) discloses a latent space-based curling decision search method. It maps the actual game state to a low-dimensional latent state representation, uses Monte Carlo tree search to search for strategies in the latent space, and finally calculates the optimal strategy based on node value and visit frequency.

[0006] However, there are two main problems with the existing technology:

[0007] (1) In the curling positioning and tracking problem, existing vision-based solutions mainly have the following drawbacks:

[0008] a. Existing technologies all use relatively traditional target detection algorithms, which have limited accuracy and generalization. When applied to new curling venues, a large amount of image data needs to be collected and labeled again due to changes in camera position and angle.

[0009] b. Existing methods typically use object detection networks such as YOLO, which can only locate the bounding box of the target. However, due to the relatively large size of a curling stone (approximately 30 cm in diameter), the position of the bounding box varies significantly in images viewed from different angles. For example, when viewing the curling stone at an angle, the center of the bounding box is not exactly at the center of the stone. Therefore, directly calculating the center coordinates of the stone from the detected bounding box area will result in a large error. These techniques can only achieve high accuracy when the camera is shooting the curling stone vertically from above, which places high demands on the number of cameras and their installation positions.

[0010] Existing technologies for generating curling strategies have the following drawbacks:

[0011] a. Insufficient long-term strategy modeling. Existing strategy methods only consider the outcome of the current end during training and inference, neglecting the overall outcome of the match. Therefore, the predicted strategies given in the final few stones of the last two ends often differ significantly from the strategies actually adopted by the athletes. For example, in the final stone of the last end of a match, if the player is trailing by 3 points, they will typically adopt an aggressive strategy to gain a chance of >= 3 points to win the match. However, existing strategies lack a global perspective, often generating strategies that aim to win the current end by 1 or 2 points but ultimately lose the match.

[0012] b. Existing strategies rely solely on data-driven algorithms, which may be limited to local optima. The reasoning process does not support the integration of commentators' and coaches' judgments on the current situation with their professional knowledge, and cannot effectively support the actual needs of commentators in selecting and predicting tactical types in smart viewing. Summary of the Invention

[0013] The purpose of this invention is to provide a curling motion acquisition, perception, and strategy prediction method for smart spectator applications, which aims to improve the accuracy of curling game prediction in the last few ends of a game and to improve the accuracy of curling image recognition.

[0014] The objective of this invention can be achieved through the following technical solutions:

[0015] A method for data acquisition, perception, and strategy prediction in curling for smart spectating, comprising the following steps:

[0016] S1. By fusing images of the same time frame from a single curling video obtained from cameras at multiple locations, the current frame curling observation coordinates of each curling target are obtained. Based on the video frame order of the curling video, trajectory matching is performed on each curling target to obtain the motion trajectory of all curling stones during the current curling process, which serves as the curling motion acquisition and perception result.

[0017] S2. Obtain the current actual curling state, combine the expert strategy method and the improved curling strategy prediction model to obtain the maximum possible action corresponding to the current actual curling state, and use it as the strategy prediction result. Render the strategy prediction result.

[0018] The training process of the improved curling strategy prediction model is as follows:

[0019] Construct an initial policy network and an initial value network, obtain training data, and perform supervised training on the policy network and value network based on the training data to obtain the first policy network and the first value network.

[0020] An improved curling policy prediction model is obtained by retraining the first policy network and the first value network based on the improved self-game reinforcement learning and Monte Carlo tree search method.

[0021] Furthermore, the training data includes training label data {(TS)} t Ta t Tz t The training labeled data includes labeled data from multiple matches, and each match labeled data includes labeled data from multiple ends. Each end labeled data includes all states of a curling end and the final score Tz of that end. t A curling game consists of all states, including the practice state for each throw. The practice state TS for the t-th throw is defined within all states of a curling game. t Transition to the training state TS for the (t+1)th curling throw t+1 The corresponding training action is Ta. t The policy network and value network are trained under supervision based on the training label data to obtain the first policy network and the first value network.

[0022] Furthermore, the specific steps for fusing images of the same time frame from a single curling throw video obtained from cameras at multiple locations to obtain the current frame curling observation coordinates for each curling target are as follows:

[0023] Multiple curling targets are identified from the same time frame images of a single curling throw video obtained from cameras at multiple locations. The same time frame segmentation map of each curling target from multiple perspectives is obtained. Based on the same time frame segmentation map of each curling target from each perspective, the actual coordinates of the curling stone on the track in the current frame of that perspective are determined. The actual coordinates of the same curling target in the current frame of each perspective are fused to obtain the current frame curling observation coordinates of each curling target.

[0024] Furthermore, the specific steps for performing trajectory matching on each curling target based on the video frame order to obtain the motion trajectory of all curling stones are as follows:

[0025] For the same video frame t, assuming *a* curling targets are identified, obtain the current frame observation coordinates of *a* curling targets and their state vectors in video frame t. Predict the state vector of video frame t+1 using the state vector of video frame t, the current frame observation coordinates of video frame t, and the Kalman filter state transition model. Obtain the predicted coordinates of the *a* curling targets in video frame t+1. Calculate the Euclidean distance between the current frame observation coordinates and the predicted coordinates of the curling targets in video frame t+1. Use the Hungarian algorithm based on the Euclidean distance for optimal matching. For successfully matched targets, update the current frame observation coordinates of the curling targets in video frame t+1 to the corresponding motion trajectory sequence, and obtain the state vector of the curling targets in video frame t+1. For unmatched targets, use the current frame observation coordinates of the curling targets in video frame t+1 as newly appearing targets. Initialize the state vector of the newly appearing targets in video frame t+1, update the video frame to t=t+1, and repeat the above steps to obtain the motion trajectories of all curling stones during the current throw.

[0026] Furthermore, the first policy network and the first value network are retrained based on the improved self-game reinforcement learning and Monte Carlo tree search method to obtain the improved curling policy prediction model. The specific steps are as follows:

[0027] A1. Set the initial curling state as the current curling state;

[0028] A2. Input the current state of the curling stone throwing into the first policy network and the first value network respectively to obtain the probability distribution of the action corresponding to the current state of the curling stone throwing. The probability distribution of a score in a game v t The probability distribution of actions Sample a set of discrete actions The set of basic actions for the current state of curling. ={ }, Let m represent the 1st, 2nd, ..., mth discrete action in the base action set, where m represents the total number of discrete actions. Then, we obtain the discrete actions. The corresponding probabilities P t ;

[0029] A3. Set the current curling state and basic actions. Probability P t The probability distribution of a score in a game v t As the root node of the Monte Carlo tree, after sequentially performing selection, expansion, simulation, and backtracking, the cumulative reward and number of visits of the nodes in the Monte Carlo tree are updated.

[0030] A4. Use the normalized number of visits to the child nodes of the root node as the probability distribution of the first action;

[0031] Sampling the probability distribution of the first action, the first action corresponding to the current state of the curling is obtained. The first action is executed based on the current state of the curling, and the obtained state is used as the new state of the current curling. Returning to A2 until the end of the current game, the end of the current game means that the current number of throws in the current game has reached the threshold.

[0032] A5. Integrate the current state of each curling stone, the first action corresponding to each current state of each curling stone, the cumulative score from the first end to the current end, the number of remaining ends in the current game up to the current end, and the win or loss of the first end in the current game as a global Q network training set.

[0033] A6. Use the first action corresponding to each current curling state as the label of the first policy network, and the result of the current game at the end of the current game as the label of the first value network for training, to obtain the second policy network and the second value network.

[0034] A7. Construct a global Q-network, train the global Q-network based on the global Q-network training set, and obtain the trained Q-network;

[0035] A8. Set the initial curling state as the current curling state;

[0036] A9. Input the current curling state into the second policy network and the second value network respectively to obtain the probability distribution of the action corresponding to the current curling state. And the probability distribution of a game's score v t ', probability distribution of actions Sample a set of discrete actions , , ..., The set of basic actions for the current state of curling. , , ..., }, , , ..., Represents the 1st, 2nd... of the base action set. A discrete action, This represents the total number of discrete actions, and yields the discrete actions. , , ..., The corresponding probabilities P t ';

[0037] A10. Set the current curling state and basic actions. Probability P t 'And the probability distribution of a game's score v t As the root node of the Monte Carlo tree, after sequentially performing selection, expansion, simulation, and backtracking, the cumulative reward and number of visits of the nodes in the Monte Carlo tree are updated;

[0038] A11. Use the normalized number of visits to the child nodes of the root node as the second probability distribution. Sample a set of actions corresponding to the current curling state from the second probability distribution without replacement. Indicates the 1st, 2nd... k Each sampling action and the probability corresponding to each sampling action ,in i =1, 2... k , k Indicates the total number of sampling actions;

[0039] A12. Input the current curling state, k sampled actions, the cumulative score from the first end to the current end corresponding to the current curling state, and the remaining end in the current match corresponding to the current curling state into the trained Q network to obtain the probability of the second player winning or losing in the current match for each sampled action in the first end. Based on the probability of winning or losing and the probability corresponding to each sampling action Calculate the probability of fusing local and global information Based on the current state of curling, perform a probabilistic calculation that integrates local and global information. The largest sampling action is used as the new current curling state, and the state is returned to A9 until the end of the current game.

[0040] A13. The probability of fusing local and global information for each current curling state. Integrate the maximum action, the cumulative score from the first inning to the current inning corresponding to each current curling state, the remaining innings in the current game corresponding to each current curling state, and the win-loss situation of the second hand in the first inning in the current game as the global Q-network fine-tuning dataset, and fine-tune the trained global Q-network based on the fine-tuning dataset to obtain the fine-tuned global Q-network.

[0041] Further, to obtain the current actual curling state and combine the expert strategy method and the improved curling strategy prediction model, the specific steps to obtain the maximum possible action corresponding to the current actual curling state are as follows:

[0042] B1. Input the current actual curling state into the second policy network and the second value network respectively. The second policy network and the second value network output the probability distribution of the action corresponding to the current actual curling state and the probability distribution of the score for one inning corresponding to the current actual curling state respectively;

[0043] B2. Sample a set of actual actions and the probabilities corresponding to the actual actions from the probability distribution of the action corresponding to the current actual curling state, and select f (f < e) expert strategies from the predefined e expert strategies , denote the 1st, ……, the f th expert strategies. One expert strategy represents one action, and at the same time, obtain the normalized credibility of each expert strategy;

[0044] B3. The expert strategies and a set of actual actions form the actual discrete action set, and the probabilities corresponding to the actual actions and the normalized credibility of each expert strategy form the probabilities corresponding to the actual discrete action set;

[0045] B4. Take the current actual curling state, the actual discrete action set, the probabilities corresponding to the actual discrete action set, and the probability distribution of the score for one inning corresponding to the current actual curling state as the root node of the Monte Carlo tree. After performing selection, expansion, simulation, and backtracking in sequence, update the cumulative rewards and visit counts of the nodes of the Monte Carlo tree;

[0046] B5. Take the normalized visit counts of the child nodes of the root node as the first probability distribution ’, and sample a set of sampling actions corresponding to the current actual curling state without replacement from the first probability distribution denote the 1st, 2nd, …… k th sampling actual actions, and the probabilities corresponding to each sampling actual action , where i = 1, 2, …… k , k denote the total number of sampling actual actions;

[0047] B6. Input the current actual curling state, k sampled actual actions, the cumulative score from the first end to the current end corresponding to the current actual curling state, and the number of remaining ends in the current game corresponding to the current actual curling state into the fine-tuning global Q network. The fine-tuning global Q network outputs the probability of the second player winning or losing in the current game for each sampled actual action in the first end. Calculate the actual probability that integrates local and global information, and take the sampled actual action with the highest actual probability as the most likely action corresponding to the current actual curling state.

[0048] Furthermore, the probability of fusing local and global information is:

[0049]

[0050] in, This represents the probability of fusing local and global information. Indicates the first i Each sampling action, This is a dynamic coefficient.

[0051] Furthermore, the dynamic coefficient is:

[0052]

[0053] in, Here are hyperparameters, where c is half the total number of ends in a curling match, and n... t This represents the number of remaining ends in the current curling game, up to the current end, corresponding to the current curling state.

[0054] Furthermore, the state space corresponding to the state consists of the sequential positions of all curlers and the house position, whether the current curler is being thrown by the team that goes second in the first end, which number of the current curler in the end, and the current end number.

[0055] Furthermore, the action space corresponding to the action is the target position of the curling stone and whether to rotate clockwise or counterclockwise to reach the target position.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] The improved self-game reinforcement learning algorithm designed in this invention enhances long-term strategy modeling by introducing a global Q-network. Based on dynamic coefficients, it effectively balances the single-game wins and losses with the overall cumulative score at different stages of the game. By integrating local and global information, the generated strategy focuses more on single-game scores in the early stages (with many games remaining) and more on the overall cumulative score in the later stages (with fewer games remaining). This dynamic adjustment mechanism, integrating local and global information, gives the invention both the flexibility and specificity of single-game decision-making at the micro level and the stability and goal orientation of the overall strategy at the macro level. This ensures the consistency and continuity of the strategy at each stage, ultimately achieving an optimal balance between single-game benefits and overall cumulative score. This significantly improves the adaptability and accuracy of the curling strategy prediction method in complex game scenarios.

[0058] Furthermore, this invention presents a curling target detection and segmentation scheme based on a large visual model. This scheme benefits from the powerful generalization ability of the large visual model, achieving relatively accurate curling target segmentation without requiring a large amount of labeled data for the curling scene. This invention employs a curling coordinate localization method based on differentiable rendering optimization. By utilizing the consistency constraints between the standard 3D model of the curling target and the actual scene, the true coordinates of the curling stone can be accurately solved. Compared with other schemes that use coarse target detection and localization methods, this scheme not only obtains more accurate curling stone position information but also does not require the camera to be installed directly above the rink, thus having a wider range of application scenarios. Attached Figure Description

[0059] Figure 1 This is a structural diagram of the system corresponding to the present invention;

[0060] Figure 2 This is a diagram of the global Q network architecture. Detailed Implementation

[0061] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0062] This invention proposes a method for collecting, sensing, and predicting strategies in curling for smart spectating, the method comprising the following steps:

[0063] S1. By fusing images of the same time frame from a single curling video obtained from cameras at multiple locations, the current frame curling observation coordinates of each curling target are obtained. Based on the video frame order of the curling video, trajectory matching is performed on each curling target to obtain the motion trajectory of all curling stones during the current curling process, which serves as the curling motion acquisition and perception result.

[0064] S2. Obtain the current actual curling state, combine the expert strategy method and the improved curling strategy prediction model to obtain the maximum possible action corresponding to the current actual curling state, and use it as the strategy prediction result. Render the strategy prediction result.

[0065] The training process of the improved curling strategy prediction model is as follows:

[0066] Construct an initial policy network and an initial value network, obtain training data, and perform supervised training on the policy network and value network based on the training data to obtain the first policy network and the first value network.

[0067] An improved curling policy prediction model is obtained by retraining the first policy network and the first value network based on the improved self-game reinforcement learning and Monte Carlo tree search method.

[0068] The basic content and core inventive points of this invention are as follows:

[0069] (1) In the acquisition and perception section, this invention uses multiple cameras set up around and above the curling venue to capture images of the same curling lane, while simultaneously acquiring images of the curling venue. Images from various perspectives are processed using a Segment Anything Model (SAM) visual model to segment and locate the curling targets in each frame. Then, for each segmented curling stone, a coordinate rotation angle optimization algorithm based on differentiable rendering is used to obtain the absolute coordinates and angles of the stone on the venue. Finally, based on the continuity of curling motion, the position of each stone is tracked using a Kalman filter algorithm, thus obtaining the trajectory of each stone throw. At the same time, according to the throwing rules of curling, it is also possible to determine which team each stone came from and which throw it was.

[0070] (2) Based on the KR-DL-UCT (Kernel regression-Deep Learning-Upper ConfidenceBound Apply to Tree Monte Carlo Tree Search) algorithm, this invention introduces a global Q network, proposes a mechanism for the policy model to integrate global wins and losses with single-game scores, optimizes the sampling process of the Monte Carlo Tree Search algorithm, and guides the policy model to optimize the globality of the policy in long-term games.

[0071] (3) Data-driven strategy models may fail in extreme situations (being far behind), while expert strategies are stable in certain contexts. This invention proposes an expert-learning strategy hybrid control system, which effectively integrates expert strategies with deep reinforcement learning-based strategies to improve the robustness and generalization of the strategies.

[0072] This invention consists of three parts: a data acquisition and sensing module, a simulation and strategy prediction module, and a VR animation rendering module. The structure is as follows: Figure 1 As shown.

[0073] 1. Data Acquisition and Sensing Module:

[0074] a. Multi-camera system: This invention involves setting up five or more cameras around the curling rink. Each camera needs to be distributed across a different viewing angle, spread out over an area of ​​at least 270 degrees around the rink. The specific camera positions can be adjusted according to the playing area, ensuring a clear view of the target area. Video signals from multiple cameras are transmitted via fiber optic cable to a switch using Network Device Interface (NDI) transmission equipment and protocols. A server connected to the switch via a network cable can simultaneously read each video signal. Each camera needs to have its intrinsic parameters pre-calibrated for its chosen resolution and focal length settings, using a checkerboard calibration method. After the camera positions are determined, the extrinsic parameters of the cameras are calibrated using the marking lines and points on the curling rink.

[0075] b. Curling Object Detection: For images acquired from each viewpoint, a large-scale visual model is used for segmentation to obtain segmentation images of each curling stone on the track. This solution uses the SAM model, which has been thoroughly trained on a large-scale segmentation dataset and can automatically detect and segment various objects without additional training data, exhibiting strong generalization capabilities. The pre-trained SAM segmentation model is a general-purpose model that requires input prompts to segment specific objects. To apply it to the specific scenario of curling, this solution fine-tunes the model: First, images containing curling stones are collected from the internet or at the event venue. Then, these images are labeled with the assistance of the pre-trained SAM model. This involves inputting the images into the SAM model while manually providing prompts indicating the location of the curling stones (points or bounding boxes), causing the SAM model to output segmentation images of all curling stones. Approximately 1000 labeled curling stone images are needed. These images are used to fine-tune the pre-trained SAM model, enabling it to complete curling stone segmentation without any prompts.

[0076] c. Curling Coordinate Prediction: Due to the relatively standard shape and size of the curling stone, and the fact that it only slides on the rink surface, the actual coordinates of the stone on the rink can be calculated through optimization, using these constraints and the stone segmentation image obtained in step b. First, a standard 3D mesh model of a curling stone needs to be prepared, consistent with the actual curling stone in a match (3D model colors normalized to binary grayscale). For each stone segmentation image, the average value of the segmented pixels is calculated. Then, using the intrinsic and extrinsic parameters of the corresponding camera, this coordinate is back-projected back to the ground coordinates of the playing area, serving as the initial coordinates for optimization. In virtual 3D space, the 3D curling stone model is initialized to the corresponding position. Then, using the intrinsic and extrinsic parameters of the corresponding camera, an approximately differentiable renderer (OpenDR) is used for projection to obtain the projected image. When the stone coordinates and rotation angle are accurate, this 3D projected segmentation image should overlap as much as possible with the segmentation image obtained in step b. Therefore, the planar coordinates and rotation angle of the stone can be optimized using the following formula:

[0077]

[0078] In this model, OpenDR is the renderer, Obj is a fixed 3D curling model, x, y, and θ are the horizontal and vertical coordinates and rotation angle of the curling stone in world coordinates (track plane coordinates), respectively, and M is the predicted curling stone segmentation image in b. The absolute value error between the rendered image and the predicted segmentation image is used as the loss function between the two. Since the entire calculation process is differentiable, the gradient of the loss function can be backpropagated to the three variables x, y, and θ. Therefore, gradient optimization methods such as Newton's method can be used to optimize the coordinates and rotation variables, thus obtaining the actual coordinates and rotation of the curling stone. The reciprocal of the optimized nearest distance is used as the confidence level of the coordinates.

[0079] d. Multi-camera data fusion: Since multiple cameras are simultaneously capturing the competition venue, excluding obstructed views, a single curling stone may be captured by multiple cameras. Therefore, it is necessary to fuse the coordinates of the curling stone obtained from each viewpoint. First, the Hungarian matching algorithm is used to match the same curling stone from different views. Then, the coordinates from each viewpoint are weighted and averaged using confidence levels to obtain the final coordinates.

[0080] e. Curling Tracking: Following the video frame order, a Kalman filter tracking algorithm is used to track and concatenate all the curling stones to obtain the motion trajectory of each stone. Specifically: A state vector is constructed using the x, y coordinates, rotation angle θ, and velocity of each curling stone. When a new curling stone appears, the state vector is initialized, generating a new curling stone trajectory. At the next moment (new frame), the current state is predicted based on the previous moment's state and the state transition model. The Euclidean distance between the observed coordinates and the predicted trajectory coordinates obtained through 'd' in the next frame is calculated. The Hungarian algorithm is used to perform optimal matching between each observation and trajectory. Successfully matched observations are updated to the corresponding trajectory sequence, while those that fail to match are considered newly appearing curling stones. In this way, each curling stone will have a continuous motion trajectory. According to the curling rules and the order of the teams, the color and sequence number of each newly appearing curling stone trajectory can be identified.

[0081] In the matching process of frame t+1, the predicted state vector of frame t+1 is obtained by using the observed state vectors of the historical curling trajectory (frames t, t-1, ...) and predicting them using the Kalman filter state prediction formula. The state vector of frame t+1 is obtained by calculating data such as speed using the current trajectory (frames t+1, t...) when updating the trajectory of the corresponding curling stone after the matching of frame t+1 is completed. The sequence of curling stone motion trajectories is the sequence of observed state vectors.

[0082] For the same video frame t, let the current frame curling observation coordinates of a identified curling targets and the state vectors of a curling targets in video frame t be respectively. By predicting the state vector of video frame t+1 using the state vector of video frame t, the current frame curling observation coordinates of video frame t, and the Kalman filter state transition model, the predicted coordinates of a curling targets in video frame t+1 are obtained. The Euclidean distance between the current frame curling observation coordinates and the predicted coordinates of the curling targets in video frame t+1 is calculated. The Hungarian algorithm is used to perform optimal matching based on the Euclidean distance. The current frame curling observation coordinates of the successfully matched video frame t+1 are updated to the motion trajectory sequence of the corresponding curling target, and the state vector of the curling target in video frame t+1 is calculated based on the trajectory sequence. The current frame curling observation coordinates of the unmatched video frame t+1 are used as the state vector of the newly appearing curling target. The state vector of the newly appearing curling target in video frame t+1 is initialized, the video frame is updated to t+2, and the above steps are repeated to obtain the motion trajectory of all curling targets. The current frame curling observation coordinates are obtained from the video frames using a target recognition algorithm.

[0083] In the above steps, after identifying the current frame observation coordinates of *a* curling targets and their respective state vectors in video frame *t*, for the first curling target, its state vector in video frame *t* is used as the optimal estimate for the previous moment. Based on this optimal estimate, the prior estimate for frame *t+1* (i.e., the predicted value of the state vector in frame *t+1*) is calculated, along with the prior error covariance and Kalman gain. Using the current frame observation coordinates of the first curling target as the measurement vector, the posterior state estimate and posterior error covariance for frame *t+1* are obtained. These are then used as inputs for the next iteration. The predicted value of the state vector in frame *t+1*, combined with the observation coordinates in video frame *t*, yields the predicted coordinates of the curling target in video frame *t+1*. This process is repeated for all curling targets to obtain the predicted coordinates of the *a* curling targets in frame *t+1*. Match the current frame observation coordinates of each curling target in the known video frame t+1 with the predicted coordinates of each of the a curling targets. For a curling target, the current frame observation coordinates that successfully match its predicted coordinates are used as the next coordinates in the motion trajectory sequence of that curling target.

[0084] 2. Simulation and Policy Prediction:

[0085] a. Curling physics simulation platform:

[0086] This invention is based on a digital curling physics simulation platform. By utilizing the aforementioned acquisition and sensing modules, multiple sets of curling motion trajectories without ice sweeping are collected on-site. Based on these collected trajectories, parameters such as the friction coefficient in the physics simulation platform are configured using a system identification method to make the physics simulation environment more closely resemble the real competition environment.

[0087] b. Overview of the strategy module framework:

[0088] The strategy prediction module in this invention is based on the KR-DL-UCT algorithm. Its network architecture includes a strategy-value network. First, the sequential positions of all stones in the current round and the house position are converted into a 32*32 pixel image and input into the network. Latent features are obtained through multiple convolutional layers. These latent features are then input into the strategy tap and the value tap. The strategy tap returns the probability distribution (action) of whether the stone stops at any of the aforementioned 32*32 pixels and arrives at that location via clockwise or counter-clockwise rotation, i.e., the probability distribution of 32*32*2 candidate actions. The value network outputs the probability distribution of the score in the current round, with the score ranging from [-8, 8] (value), totaling 17 values ​​(each player throws 8 stones per round).

[0089] To address the limitations of discrete states and action spaces in high-dimensional continuous action scenarios, a Monte Carlo Tree Search (MCTS) method is further employed. First, the state S of the adaptive policy network at that moment is constructed based on the perceived current game information. t， The input is fed into the policy-value network to obtain the probability distribution of actions. Score probability distribution v t Based on the action probability distribution output by the policy network Sample a set of discrete actions As the set of base actions in the current state and the probability P corresponding to these actions t The KR-DL-UCT algorithm uses S t , P t v t The root node of the Monte Carlo tree is constructed using the input. Combining the physical simulation platform and policy-value network from 5.2a, the four-step process (selection-expansion-simulation-backtracking) is called multiple times in continuous states and action space. Finally, based on the probability distribution of the normalized visit counts of the root node's child nodes, the current state S is sampled. t Action strategy .

[0090] Finally, based on the mapping function of the curling kinematics characteristics developed in the physical simulation platform, the above action strategy is applied. Map this to the corresponding curling throw speed and spin direction. Set the current state S... t The predicted speed and direction of the curling throw are input into the simulation platform, which then simulates the trajectory of the thrown curling stone and its collision with the stones on the ice.

[0091] The training of the aforementioned strategy-value network is divided into two stages: supervised learning and self-game reinforcement learning. In the supervised learning stage, real-world data {(S)} is used. t a t z t )} is a tag (z t (The final score of the current game), the strategy-value network parameters are optimized through the cross-entropy loss function, so that the corresponding output {(S)} is equal to the final score of the current game. t , v t The labels tend to be consistent.

[0092] In the self-game reinforcement learning phase, action strategies obtained from Monte Carlo tree search during the self-game process are collected. With the game outcome z t For the labels, the same loss function optimization strategy as in the supervised learning phase is used – the value network parameters.

[0093] c. Improved self-game reinforcement learning algorithm

[0094] Based on the massive results of the first stage of self-game training (the original self-game training process of KR-DL-UCT), we construct {(S t a t n t d t Data set , r)} where n t d represents the number of games remaining up to the current game. t This represents the cumulative score across all games (positive points for a win and negative points for a loss in each game for the player who is second in the first game), and r = {1, 0, -1} indicates whether the player who is second in the first game has won the entire match (1 for a win, 0 for a draw, and -1 for a loss).

[0095] A global Q-network was designed, which is based on (S t a t n t d t Given the input r, output the probability distribution of r in the three categories {1, 0, -1}. The network architecture of the global Q-network is as follows: Figure 2 As shown, it trains a global Q-network using the dataset constructed above, with cross-entropy as the loss function.

[0096] In the second stage of self-game training, the final sampling process of Monte Carlo tree search is modified as follows: the normalized number of visits to the child nodes of the root node is used as the probability of the corresponding action of the child node, which is called the local policy probability distribution. .according to Sample a set of data based on the current state S without replacement. t action and the corresponding probability , i = 1, 2, ..., k Furthermore, n is calculated. t d t , will (S t , n t d t Inputting the global Q-network yields the corresponding global score probability distribution, from which the probability of that side winning the entire match is derived (r = 1 for the second player in the first game, and r = -1 otherwise), denoted as... The index that integrates local and global information is calculated based on the formula.

[0097]

[0098]

[0099] Pick Return the largest corresponding action. Among them, is a hyperparameter, and c is generally taken as half of the total number of games. Through the dynamic coefficient , the generated strategy pays more attention to the score of a single game (β is small) in the early stage of the game (more remaining games), and focuses more on the cumulative score of the whole game (β is large) in the later stage (less remaining games)

[0100] From the second stage of self-play training, based on the above improved method, perform algorithm self-play again, and update the dataset {(S t , a t , n t , d t , r)} accordingly, and train and fine-tune the global Q-network based on the updated dataset, still using cross-entropy as the loss function. The a in the updated dataset t is the largest corresponding action.

[0101] d. Expert-Learning Strategy Hybrid Control System

[0102] Fuse the rule-based expert strategy (such as defensive / offensive tactics) with the above data-driven strategy to form a hybrid action probability distribution, so as to integrate the professional knowledge of coaches, athletes, commentators, etc. during the game, improve the strategy diversity and interpretability, and better support intelligent viewing.

[0103] Specifically, we pre-construct an expert strategy library, define e expert strategies {π exp1 ,…,π expe}, each strategy corresponds to a specific tactic (such as "protect the center pot", "clear the opponent's pot"), and each expert strategy is implemented using a rule-based algorithm, which returns the action of the corresponding strategy according to the current state . (The above strategies do not need to be trained.)

[0104] In the inference stage, the system accepts f (f < e) expert strategies selected from the pre-defined e expert strategies and the corresponding credibility input manually (the value range is [0,1], and the higher the credibility, the more the expert tends to this strategy), and normalize the credibility of these strategies so that their sum is. Further, we expand the input of Monte Carlo tree search from the discrete action set sampled from the output distribution of the policy-value network to , and at the same time P t is expanded to P tThe normalized confidence of the expert policy actions is calculated, and the root node of the Monte Carlo tree is constructed from the expanded inputs described above. The number of expert policies *f* and the number of data-driven algorithmic policies *m* can be adjusted to accommodate different system preferences. Subsequent steps remain unchanged.

[0105] In the above process, the self-game training process of the model, that is, retraining the first policy network and the first value network based on the improved self-game reinforcement learning and Monte Carlo tree search method to obtain the improved curling policy prediction model, specifically involves the following steps:

[0106] A1. Set the initial curling state as the current curling state;

[0107] A2. Input the current state of the curling stone throwing into the first policy network and the first value network respectively to obtain the probability distribution of the action corresponding to the current state of the curling stone throwing. The probability distribution of a score in a game v t The probability distribution of actions Sample a set of discrete actions The set of basic actions for the current state of curling. ={ }, Let m represent the 1st, 2nd, ..., mth discrete action in the base action set, where m represents the total number of discrete actions. Then, we obtain the discrete actions. The corresponding probabilities P t ;

[0108] A3. Set the current curling state and basic actions. Probability P t The probability distribution of a score in a game v t As the root node of the Monte Carlo tree, after sequentially performing selection, expansion, simulation, and backtracking, the cumulative reward and number of visits of the nodes in the Monte Carlo tree are updated.

[0109] A4. Use the normalized number of visits to the child nodes of the root node as the probability distribution of the first action;

[0110] Sampling the probability distribution of the first action, the first action corresponding to the current state of the curling is obtained. The first action is executed based on the current state of the curling, and the obtained state is used as the new state of the current curling. Returning to A2 until the end of the current game, the end of the current game means that the current number of throws in the current game has reached the threshold.

[0111] A5. Integrate the current state of each curling stone, the first action corresponding to each current state of each curling stone, the cumulative score from the first end to the current end, the number of remaining ends in the current game up to the current end, and the win or loss of the first end in the current game as a global Q network training set.

[0112] A6. Use the first action corresponding to each current curling state as the label of the first policy network, and the result of the current game at the end of the current game as the label of the first value network for training, to obtain the second policy network and the second value network.

[0113] A7. Construct a global Q-network, train the global Q-network based on the global Q-network training set, and obtain the trained Q-network;

[0114] A8. Set the initial curling state as the current curling state;

[0115] A9. Input the current curling state into the second policy network and the second value network respectively to obtain the probability distribution of the action corresponding to the current curling state. And the probability distribution of a game's score v t ', probability distribution of actions Sample a set of discrete actions , , ..., The set of basic actions for the current state of curling. , , ..., }, , , ..., Represents the 1st, 2nd... of the base action set. A discrete action, This represents the total number of discrete actions, and yields the discrete actions. , , ..., The corresponding probabilities P t ';

[0116] A10. Set the current curling state and basic actions. Probability P t 'And the probability distribution of a game's score v t As the root node of the Monte Carlo tree, after sequentially performing selection, expansion, simulation, and backtracking, the cumulative reward and number of visits of the nodes in the Monte Carlo tree are updated;

[0117] A11. Use the normalized number of visits to the child nodes of the root node as the second probability distribution. Sample a set of actions corresponding to the current curling state from the second probability distribution without replacement. Indicates the 1st, 2nd... k Each sampling action and the probability corresponding to each sampling action ,in i =1, 2...k , k represents the total number of sampling actions;

[0118] A12. Input the current state of throwing a curling stone, k sampling actions, the cumulative score from the first game to the current game corresponding to the current state of throwing a curling stone, and the remaining number of games in the current game in the current curling game corresponding to the current state of throwing a curling stone into the trained Q-network, and obtain the probability of winning or losing for the second player in the first game of the current curling game corresponding to each sampling action , based on the probability of winning or losing and the probability corresponding to each sampling action calculate the probability of integrating local and global information , on the basis of the current state of throwing a curling stone, execute the probability of integrating local and global information the sampling action with the maximum value, and use the obtained state as the new current state of throwing a curling stone, and return to A9 until the end of the current game;

[0119] A13. Integrate the states of each current curling stone, the probabilities of integrating local and global information corresponding to the states of each current curling stone the action with the maximum value, the cumulative score from the first game to the current game corresponding to the states of each current curling stone, the remaining number of games in the current curling game corresponding to the states of each current curling stone, and the winning or losing situation of the second player in the first game of the current curling game as the fine-tuning dataset for the global Q-network, and fine-tune the trained global Q-network based on the fine-tuning dataset to obtain the fine-tuned global Q-network.

[0120] When actually making predictions, the specific steps to obtain the most likely action corresponding to the current actual curling state by combining the expert strategy method and the improved curling strategy prediction model are as follows:

[0121] B1. Input the current actual curling state into the second policy network and the second value network respectively. The second policy network and the second value network respectively output the probability distribution of the actions corresponding to the current actual curling state and the probability distribution of the score of one game corresponding to the current actual curling state;

[0122] B2. Sample a group of actual actions and the probabilities corresponding to the actual actions from the probability distribution of the actions corresponding to the current actual curling state, and select f (f < e) expert strategies from the predefined e expert strategies , represent the 1st, ……, the f th expert strategies. One expert strategy represents one action, and at the same time obtain the normalized credibility of each expert strategy;

[0123] B3. An expert strategy and a set of actual actions constitute the actual discrete action set. The probability corresponding to the actual action and the normalized confidence of each expert strategy constitute the probability corresponding to the actual discrete action set.

[0124] B4. Using the current actual curling state, the actual discrete action set, the probability corresponding to the actual discrete action set, and the score probability distribution of the current actual curling state as the root node of the Monte Carlo tree, after selecting, expanding, simulating, and backtracking in sequence, update the cumulative reward and access count of the nodes in the Monte Carlo tree.

[0125] B5. Use the normalized number of visits to the child nodes of the root node as the first probability distribution. 'Sampling actions corresponding to the current actual curling state without replacement from the first probability distribution' Indicates the 1st, 2nd... k Each sampled actual action, and the probability corresponding to each sampled actual action. ,in i =1, 2... k , k Indicates the total number of actual actions sampled;

[0126] B6. Input the current actual curling state, k sampled actual actions, the cumulative score from the first end to the current end corresponding to the current actual curling state, and the remaining end in the current game corresponding to the current actual curling state into the fine-tuning global Q network. The fine-tuning global Q network outputs the probability of the second player winning or losing in the current game for each sampled actual action in the first end. Calculate the actual probability of fusing local and global information, and take the sampled actual action with the highest actual probability as the most likely action corresponding to the current actual curling state. The formula for the actual probability of fusing local and global information is the same as the probability of fusing local and global information in A12. The formula is the same.

[0127] 3. Animation rendering

[0128] The VR event footage is reconstructed using information obtained from the acquisition and perception module, and the predicted strategy effect is rendered based on information obtained from the strategy simulation based on the strategy module of the physical simulation platform.

[0129] This invention adopts a curling target detection and segmentation scheme based on a large visual model. This scheme benefits from the powerful generalization ability of the large visual model and can achieve relatively accurate curling target segmentation without the need for a large amount of labeled data of curling scenes.

[0130] This invention employs a curling coordinate positioning method based on differentiable rendering optimization. By utilizing the consistency constraints between the standard 3D model of the curling target and the actual scene, the true coordinates of the curling stone can be accurately solved. Compared with other solutions that use coarse target detection and positioning methods, this method can not only obtain more accurate curling stone position information, but also does not require the camera to be installed directly above the field, thus having a wider range of application scenarios.

[0131] The improved self-game reinforcement learning algorithm designed in this invention improves the long-term policy modeling of the algorithm by introducing a global Q-network, and effectively balances the single-game wins and losses of the generated strategy with the total score of the whole game based on dynamic coefficients at different stages of the game.

[0132] The expert-learning strategy hybrid control system method adopted in this invention integrates expert domain knowledge through an expert strategy library, and further improves tactical diversity and strategy quality by combining data-driven methods and Monte Carlo tree search. It avoids deep reinforcement learning strategies from getting stuck in local optima and can effectively support the actual needs of commentators to select prediction tactical types in smart viewing.

[0133] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for data acquisition, perception, and strategy prediction in curling for smart spectating, characterized in that, The method includes the following steps: S1. By fusing images of the same time frame from a single curling video obtained from cameras at multiple locations, the current frame curling observation coordinates of each curling target are obtained. Based on the video frame order of the curling video, trajectory matching is performed on each curling target to obtain the motion trajectory of all curling stones during the current curling process, which serves as the curling motion acquisition and perception result. S2. Obtain the current actual curling state, combine the expert strategy method and the improved curling strategy prediction model to obtain the maximum possible action corresponding to the current actual curling state, and use it as the strategy prediction result. Render the strategy prediction result. The training process of the improved curling strategy prediction model is as follows: Construct an initial policy network and an initial value network, obtain training data, and perform supervised training on the policy network and value network based on the training data to obtain the first policy network and the first value network. An improved curling strategy prediction model is obtained by retraining the first policy network and the first value network based on the improved self-game reinforcement learning and Monte Carlo tree search method. An improved self-play reinforcement learning algorithm enhances long-term policy modeling by introducing a global Q-network. Based on dynamic coefficients, it effectively balances the single-game wins and losses and the total score of the generated policy at different stages of the game. By integrating local and global information and using dynamic coefficients, the generated policy focuses more on single-game scores in the early stages of the game and more on the total score of the game in the later stages.

2. The method for collecting, sensing, and predicting strategies for curling in smart spectating, as described in claim 1, is characterized in that... The training data includes training label data {(TS)} t Ta t Tz t The training labeled data includes labeled data from multiple matches, and each match labeled data includes labeled data from multiple ends. Each end labeled data includes all states of a curling end and the final score Tz of that end. t A curling game consists of all states, including the practice state for each throw. The practice state TS for the t-th throw is defined within all states of a curling game. t Transition to the training state TS for the (t+1)th curling throw t+1 The corresponding training action is Ta. t The policy network and value network are trained under supervision based on the training label data to obtain the first policy network and the first value network.

3. The method for collecting, sensing, and predicting strategies for curling in smart spectating, as described in claim 1, is characterized in that... The specific steps for obtaining the current frame observation coordinates of each curling target by fusing images of the same time frame from multiple cameras in a single curling video are as follows: Multiple curling targets are identified from the same time frame images of a single curling throw video obtained from cameras at multiple locations. The same time frame segmentation map of each curling target from multiple perspectives is obtained. Based on the same time frame segmentation map of each curling target from each perspective, the actual coordinates of the curling stone on the track in the current frame of that perspective are determined. The actual coordinates of the same curling target in the current frame of each perspective are fused to obtain the current frame curling observation coordinates of each curling target.

4. The method for collecting, sensing, and predicting strategies for curling in smart spectating, as described in claim 1, is characterized in that... The specific steps for matching the trajectory of each curling stone based on the video frame sequence to obtain the motion trajectory of all curling stones are as follows: For the same video frame t, assuming *a* curling targets are identified, obtain the current frame observation coordinates of *a* curling targets and their state vectors in video frame t. Predict the state vector of video frame t+1 using the state vector of video frame t, the current frame observation coordinates of video frame t, and the Kalman filter state transition model. Obtain the predicted coordinates of the *a* curling targets in video frame t+1. Calculate the Euclidean distance between the current frame observation coordinates and the predicted coordinates of the curling targets in video frame t+1. Use the Hungarian algorithm based on the Euclidean distance for optimal matching. For successfully matched targets, update the current frame observation coordinates of the curling targets in video frame t+1 to the corresponding motion trajectory sequence, and obtain the state vector of the curling targets in video frame t+1. For unmatched targets, use the current frame observation coordinates of the curling targets in video frame t+1 as newly appearing targets. Initialize the state vector of the newly appearing targets in video frame t+1, update the video frame to t=t+1, and repeat the above steps to obtain the motion trajectories of all curling stones during the current throw.

5. The method for collecting, sensing, and predicting strategies for curling in smart spectating, as described in claim 1, is characterized in that... The specific steps for retraining the first policy network and the first value network based on the improved self-game reinforcement learning and Monte Carlo tree search method to obtain the improved curling policy prediction model are as follows: A1. Set the initial curling state as the current curling state; A2. Input the current state of the curling stone throwing into the first policy network and the first value network respectively to obtain the probability distribution of the action corresponding to the current state of the curling stone throwing. The probability distribution of a score in a game v t The probability distribution of actions Sample a set of discrete actions The set of basic actions for the current state of curling. ={ }, Let m represent the 1st, 2nd, ..., mth discrete action in the base action set, where m represents the total number of discrete actions. Then, we obtain the discrete actions. The corresponding probabilities P t ; A3. Set the current curling state and basic actions. Probability P t The probability distribution of a score in a game v t As the root node of the Monte Carlo tree, after sequentially performing selection, expansion, simulation, and backtracking, the cumulative reward and number of visits of the nodes in the Monte Carlo tree are updated. A4. Use the normalized number of visits to the child nodes of the root node as the probability distribution of the first action; Sampling the probability distribution of the first action, the first action corresponding to the current state of the curling is obtained. The first action is executed based on the current state of the curling, and the obtained state is used as the new state of the current curling. Returning to A2 until the end of the current game, the end of the current game means that the current number of throws in the current game has reached the threshold. A5. Integrate the current state of each curling stone, the first action corresponding to each current state of each curling stone, the cumulative score from the first end to the current end, the number of remaining ends in the current game up to the current end, and the win or loss of the first end in the current game as a global Q network training set. A6. Use the first action corresponding to each current curling state as the label of the first policy network, and the result of the current game at the end of the current game as the label of the first value network for training, to obtain the second policy network and the second value network. A7. Construct a global Q-network, train the global Q-network based on the global Q-network training set, and obtain the trained Q-network; A8. Set the initial curling state as the current curling state; A9. Input the current curling state into the second policy network and the second value network respectively to obtain the probability distribution of the action corresponding to the current curling state. And the probability distribution of a game's score v t ', probability distribution of actions Sample a set of discrete actions , , ..., The set of basic actions for the current state of curling. , , ..., }, , , ..., Represents the 1st, 2nd... of the base action set. A discrete action, This represents the total number of discrete actions, and yields the discrete actions. , , ..., The corresponding probabilities P t '; A10. Set the current curling state and basic actions. Probability P t 'And the probability distribution of a game's score v t As the root node of the Monte Carlo tree, after sequentially performing selection, expansion, simulation, and backtracking, the cumulative reward and number of visits of the nodes in the Monte Carlo tree are updated; A11. Use the normalized number of visits to the child nodes of the root node as the second probability distribution. Sample a set of actions corresponding to the current curling state from the second probability distribution without replacement. , Indicates the 1st, 2nd... k Each sampling action and the probability corresponding to each sampling action ,in i =1, 2... k , k Indicates the total number of sampling actions; A12. Input the current curling state, k sampled actions, the cumulative score from the first end to the current end corresponding to the current curling state, and the remaining end in the current match corresponding to the current curling state into the trained Q network to obtain the probability of the second player winning or losing in the current match for each sampled action in the first end. Based on the probability of winning or losing and the probability corresponding to each sampling action Calculate the probability of fusing local and global information Based on the current state of curling, perform a probabilistic calculation that integrates local and global information. The largest sampling action is used as the new current curling state, and the state is returned to A9 until the end of the current game. A13. The probability of fusing local and global information for each current curling state. The largest action, the cumulative score from the first end to the current end for each current curling state, the number of remaining ends in the current game for each current curling state, and the win / loss situation of the first end in the current game are integrated into a global Q-network fine-tuning dataset. The trained global Q-network is then fine-tuned based on the fine-tuning dataset to obtain the fine-tuned global Q-network.

6. The curling motion data acquisition, perception, and strategy prediction method applied to smart spectating, as described in claim 5, is characterized in that... The specific steps to obtain the current actual curling state, and to derive the most likely action corresponding to the current actual curling state by combining the expert strategy method and the improved curling strategy prediction model, are as follows: B1. The current actual curling state is input into the second policy network and the second value network respectively. The second policy network and the second value network output the probability distribution of the action corresponding to the current actual curling state and the probability distribution of the score of a game corresponding to the current actual curling state respectively. B2. Sample a set of actual actions and the probabilities corresponding to the actual actions from the probability distribution of the actions corresponding to the current actual curling state, and select f expert strategies from the predefined e expert strategies , denote the 1st, ……, the f th expert strategies, f < e, 1 expert strategy represents 1 action, and at the same time obtain the normalized credibility of each expert strategy; B3. An expert strategy and a set of actual actions constitute the actual discrete action set. The probability corresponding to the actual action and the normalized confidence of each expert strategy constitute the probability corresponding to the actual discrete action set. B4. Using the current actual curling state, the actual discrete action set, the probability corresponding to the actual discrete action set, and the score probability distribution of the current actual curling state as the root node of the Monte Carlo tree, after selecting, expanding, simulating, and backtracking in sequence, update the cumulative reward and access count of the nodes in the Monte Carlo tree. B5. Use the normalized number of visits to the child nodes of the root node as the first probability distribution. 'Sampling actions corresponding to the current actual curling state without replacement from the first probability distribution' , Indicates the 1st, 2nd... k Each sampled actual action, and the probability corresponding to each sampled actual action. ,in i =1, 2... k , k Indicates the total number of actual actions sampled; B6. Input the current actual curling state, k sampled actual actions, the cumulative score from the first end to the current end corresponding to the current actual curling state, and the number of remaining ends in the current game corresponding to the current actual curling state into the fine-tuning global Q network. The fine-tuning global Q network outputs the probability of the second player winning or losing in the current game for each sampled actual action in the first end. Calculate the actual probability that integrates local and global information, and take the sampled actual action with the highest actual probability as the most likely action corresponding to the current actual curling state.

7. The curling motion data acquisition, perception, and strategy prediction method applied to smart spectating, as described in claim 5, is characterized in that... The probability of fusing local and global information is: Wherein, represents the probability of fusing local and global information, and represents the th... i Each sampling action, This is a dynamic coefficient.

8. The curling motion data acquisition, perception, and strategy prediction method applied to smart spectating, as described in claim 7, is characterized in that... The dynamic coefficient is: in, Here are hyperparameters, where c is half the total number of ends in a curling match, and n... t This represents the number of remaining ends in the current curling game, up to the current end, corresponding to the current curling state.

9. The method for collecting, sensing, and predicting strategies for curling in smart spectating, as described in claim 1, is characterized in that... The state space corresponding to the state is the sequential position of all curlers and the house position, whether the current curler is the first end's second team's curler, which curler is the current curler's number in the end, and the current end number.

10. The method for collecting, sensing, and predicting strategies for curling in smart spectating, as described in claim 1, is characterized in that... The action space corresponding to the movement is the target position of the curling stone and whether to rotate clockwise or counterclockwise to reach the target position.

Images