Automatic generation method of basketball game player point map

Abstract

The invention relates to the technical field of computer vision and data processing, and discloses an automatic generation method of a basketball game athlete point map, which comprises the following steps: instantiating a dynamic kinematics posture model for an athlete; analyzing and identifying tactical roles based on the states of all athletes; according to the tactical role, adaptively selecting a state transfer function to carry out state prediction to obtain a prior state; correcting the prior state by using the visual observation data of the current video frame to obtain a posterior state; and finally, extracting position coordinates from the posterior state to generate a point location map. According to the method, the prediction model is dynamically adjusted by introducing tactical role recognition, high-level semantic judgment and bottom-level state estimation are combined, the tracking accuracy and robustness of athletes in a high-dynamic tactical scene are remarkably improved, and meanwhile output three-dimensional attitude and semantic information provide data support for advanced tactical analysis.

Description

Technical Field

[0001] This invention relates to the field of computer vision and data processing technology, specifically a method for automatically generating a basketball game player position map. Background Technology

[0002] In modern professional sports, especially basketball, accurate acquisition and recording of the positional information of all players on the court at every moment is crucial for tactical analysis. This positional information is usually presented in the form of a point map, providing basic data support for coaches to review tactics, evaluate player performance, and formulate game strategies. Currently, the mainstream technology for automatically generating player point maps relies on computer vision and multi-target tracking algorithms.

[0003] These algorithms typically detect athletes in each frame of a video and correlate them across frames to form a motion trajectory. When making state predictions, i.e., estimating the athlete's possible position in the next moment, existing technologies generally use a single motion model based on general physical laws, such as a constant velocity model, a constant acceleration model, or a simple maneuver model.

[0004] However, basketball players' on-court behavior is not a simple physical movement, but is driven by complex and ever-changing tactical intentions. A player may be setting a screen one second and remaining almost still, while the next second he needs to make a quick cut to the basket, exhibiting a highly dynamic sprint. General motion models are difficult to effectively describe this kind of movement mode switching caused by sudden changes in tactical intentions. When players perform non-linear and discontinuous actions such as sudden stops, changes of direction, and reverse runs, the prediction results relying solely on general physical models will deviate significantly from their actual movements, leading to a decrease in the accuracy of tracking algorithms, and even problems such as identity loss or trajectory interruption. Therefore, existing technologies have limitations in the accuracy and robustness of tracking when dealing with high-intensity, high-dynamic basketball game scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an automatic method for generating player position maps in basketball games. This method solves the problem that existing technologies, which use a single, general motion model, cannot accurately describe the nonlinear motion of athletes driven by complex tactical intentions, resulting in insufficient tracking accuracy and robustness when handling highly dynamic tactical actions.

[0006] To achieve the above objectives, the present invention provides a method for automatically generating a basketball game player position map, comprising the following steps: The input competition video is initialized, and each athlete is detected and identified in the initial frame of the competition video. A dynamic kinematic posture model is instantiated for each athlete, and its individual kinematic constraint parameters are calibrated. Based on the state of all the athletes at the previous moment, analyze and identify the global tactical context of the current game and the tactical role played by each athlete. Based on the identified tactical role, a state transition function is adaptively selected for each athlete from a preset motion model library, and the athlete's state is predicted based on the state transition function to calculate the prior state at the current moment. Visual observation data related to each athlete is extracted from the current video frame, and the prior state is corrected using the visual observation data to calculate the posterior state of each athlete at the current moment. Extract the foot contact position coordinates in the standard court coordinate system from the posterior state of each athlete, generate the athlete position map of the current frame, and output it.

[0007] The state vector of the dynamic kinematic posture model includes: The athlete's position coordinates in the standard court coordinate system; The athlete's kinematic state, which includes velocity and acceleration components; Full-body 3D pose parameter vectors constructed based on a parametric human body model; The orientation of the athlete's main body in three-dimensional space, expressed as a unit quaternion.

[0008] The individualized kinematic constraint parameters include: The athlete's height; The athlete's limb segment length; The athletes' historical maximum speed and maximum acceleration statistics.

[0009] The steps of analyzing and identifying the global tactical context of the current match and the tactical role played by each athlete include: Based on the states of all the athletes at the previous moment, a dynamic topology graph containing node features and edge features is constructed. The dynamic topology graph is input into a pre-trained graph neural network model for inference. The global tactical context is output through the graph-level classifier of the graph neural network model, and the tactical role of each athlete is output through the node-level classifier of the graph neural network model.

[0010] The motion model library includes at least: A general inertial model used to describe normal running conditions; A high-dynamic sprint model used to describe fast-paced or cut-and-run characters; and a positional constraint model used to describe cover or stationary characters.

[0011] The step of predicting the athlete's state based on the state transition function is implemented using unscented Kalman filtering, specifically including: Based on the athlete's posterior state estimate mean and posterior error covariance matrix at the previous moment, a set of Sigma points is generated; Substitute the Sigma points into the selected state transition function for propagation to obtain a set of propagated Sigma points; The propagation Sigma point set is weighted and statistically analyzed to calculate the prior state at the current moment. The prior state includes the mean of the prior state estimate and the prior error covariance matrix.

[0012] The visual observation data includes: Two-dimensional joint coordinate data of the athlete are extracted from the current video frame using a pose estimation algorithm; The athlete's human body contour mask data is extracted from the current video frame using an instance segmentation network.

[0013] The step of correcting the prior state using the visual observation data includes: Construct an observation model that maps the state space to the observation space; The prior state is projected onto the observation model to obtain the predicted observation; The difference between the actual visual observation data and the predicted observation is calculated as information; Calculate the Kalman gain, and use the Kalman gain and the information to correct the prior state to obtain the posterior state.

[0014] The specific steps for extracting the foot contact position coordinates in the standard court coordinate system from the posterior state of each athlete are as follows: Predefined position coordinate components are extracted directly from the state vector of the posterior state. The position coordinate components are defined as the vertical projection of the athlete onto the standard court plane.

[0015] The step of generating and outputting the athlete position map for the current frame further includes: The three-dimensional pose parameters, body orientation, and tactical role contained in the posterior state are encapsulated into a standardized data structure and serialized for output.

[0016] This invention provides a method for automatically generating a player position map for a basketball game. It has the following beneficial effects: 1. By introducing a tactical intent recognition module, this invention can identify the tactical roles played by athletes based on their interactions, and adaptively select a matching state transition function from a motion model library accordingly. This combination of high-level semantic tactical judgment and low-level mathematical model prediction enables the system to accurately predict the nonlinear motion trajectory of athletes when executing different tactical intentions such as fast breaks and screens, thereby significantly improving the accuracy of tracking.

[0017] 2. This invention employs a dynamic kinematic posture model that includes three-dimensional posture parameters and kinematic constraints. In the state update stage, it utilizes an unscented Kalman filter framework to fuse multimodal visual observation data such as two-dimensional joints and human body contours extracted from the video. This projection matching and correction mechanism from the three-dimensional model to the two-dimensional observation ensures that the final estimated athlete state is not only accurately positioned on the two-dimensional plane, but its three-dimensional posture also conforms to human body structure and kinematic laws, enhancing the physical realism of the state estimation.

[0018] 3. In the final data output module, this invention not only extracts the foot contact position coordinates from the posterior state to generate a traditional two-dimensional point map, but also encapsulates the contained three-dimensional posture parameters, body orientation, and identified tactical roles into structured data and outputs it serially. This enables the output of this invention to go beyond simple position information, providing rich semantic data and providing direct data support for downstream advanced tactical analysis, virtual scene reconstruction, or technical and tactical statistics. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system architecture diagram of the present invention.

[0020] Among them, 10 is the initialization module; 20 is the tactical intent recognition module; 30 is the status prediction module; 40 is the status update module; and 50 is the data output module. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example: Please see the appendix Figure 1 - Appendix Figure 2This invention provides a method for automatically generating a basketball game player position map, comprising the following steps: S100. Initialize the input competition video, detect and identify each athlete in the initial frame of the video, instantiate a unique dynamic kinematic posture model for each athlete, and calibrate their individual kinematic constraint parameters. S200: Based on the state of all athletes in the previous moment, analyze and identify the global tactical context of the current game and the tactical role played by each athlete. S300. Based on the identified tactical role, adaptively select a state transition function for each athlete from the preset motion model library, and predict the athlete's state based on the function to calculate the prior state at the current moment. S400: Extract visual observation data related to each athlete from the current video frame, and use the observation data to correct the prior state obtained in step S300, and calculate the posterior state of each athlete at the current moment. S500: Extract the coordinates of each athlete's foot contact position in the standard court coordinate system from the posterior state of each athlete, generate the athlete position map of the current frame and output it.

[0023] See attached document Figure 2 An automatic generation system for basketball game player position maps, which may include: an initialization module 10, a tactical intent recognition module 20, a status prediction module 30, a status update module 40, and a data output module 50.

[0024] The initialization module 10 is used to receive the competition video data and perform field calibration to establish the mapping relationship between the image coordinate system and the field coordinate system. The initialization module 10 is also used to detect athletes in the initial stage of the video and establish a unique dynamic kinematic posture model instance for each detected athlete. This process includes initial calibration of individualized kinematic constraint parameters such as the athlete's height and limb proportions.

[0025] The tactical intent recognition module 20 is connected to the output of the initialization module 10. The tactical intent recognition module 20 receives the complete state information of all athletes at the previous moment, and calculates the macro tactical pattern on the field and the specific role of each athlete in the tactical pattern through the built-in analysis logic.

[0026] The state prediction module 30 has its input end connected to the output end of the tactical intent recognition module 20. Based on the tactical pattern and role information provided by the tactical intent recognition module 20, the state prediction module 30 matches and selects the most suitable nonlinear state transition function from the preset motion model library for each athlete. Subsequently, under the nonlinear filtering framework, the state prediction module 30 uses the selected state transition function to perform state propagation on the athlete's state kinematic posture model, thereby predicting its prior state at the current moment, including position, velocity, posture and orientation.

[0027] The state update module 40 has its input connected to the output of the state prediction module 30 and the video data source. The state update module 40 first extracts visual observation information related to the athlete's predicted position from the current video frame. Then, the state update module 40 compares this actual visual observation information with the observation information predicted based on the prior state, calculates the difference between them, and calculates the Kalman gain under the filtering framework to correct the prior state. This process finally outputs a more accurate posterior state of the athlete after visual information correction.

[0028] The data output module 50 is connected to the output of the state update module 40. The data output module 50 extracts the foot contact coordinates representing the precise position of each athlete on the court from the final posterior state vector of each athlete, integrates the coordinates of all athletes, and generates and outputs a two-dimensional point map of the current frame.

[0029] The system performs initialization processing and model building of video data, which may specifically include the following steps: Establishment of the relationship between field location and coordinate mapping.

[0030] The system receives a game video stream as input, selects a clear frame from the video as a reference image, and identifies court feature points in the reference image. These feature points include, but are not limited to, the intersection of the sidelines, the endpoint of the free throw line, or the vertex of the three-point line. The system matches the coordinates of these feature points in the image's pixel coordinate system with their corresponding coordinates in the physical coordinate system of a standard basketball court plan. Based on at least four pairs of matching points, the system calculates the homography matrix using a direct linear transformation algorithm, denoted as [matrix name missing]. The homography matrix Define the pixel coordinates of the image plane Physical coordinates of the court plane The mapping relationship is known in the field. As for the specific algorithm implementation for field calibration, those skilled in the art can use existing camera calibration technology or key point detection technology based on deep learning. It will not be elaborated here.

[0031] Initial testing and identification assignment for athletes.

[0032] The initial frames of the video are processed using a pre-trained object detection model to detect all human targets in the scene. The object detection model can employ a convolutional neural network architecture. The system assigns a unique identifier (ID) to each detected independent bounding box, denoted as . ,in , This represents the total number of athletes detected.

[0033] Instantiation and state definition of dynamic kinematic posture models.

[0034] For each athlete assigned an ID The system instantiates a dynamic kinematic posture model. The core of this model lies in defining a high-dimensional, unified state vector to fully represent the athlete's physical presence in a continuous time series. ,athlete state vector The definition is as follows: in, This indicates the coordinates of the athlete's foot's center of contact with the ground in a standard stadium coordinate system. This coordinate directly corresponds to the final generated point map data; This represents the kinematic state of the athlete in the field coordinate system; This represents a vector of three-dimensional posture parameters of the athlete's entire body, which is constructed based on a parametric human model; It indicates the orientation of the athlete's main body in three-dimensional space.

[0035] Calibration of individualized kinematic constraint parameters.

[0036] While instantiating the dynamic kinematic pose model, the system calculates and stores a set of individualized kinematic constraint parameters. As the athlete This parameter set is a unique static property; it does not update iteratively over time, but it is invoked as a constraint in state prediction and observation correction. Defined as: in, For the athlete's height; It is a vector containing the lengths of limb segments; and These are the athlete's historical maximum speed and maximum acceleration statistics, respectively.

[0037] The specific calibration process is as follows: The system uses a monocular 3D human pose estimation algorithm to process the athlete in the initial frame. From the image region, its normalized 3D mesh or skeleton in the camera coordinate system is recovered, and combined with the obtained homography matrix. Assuming ground plane constraints, the normalized 3D skeleton is restored to a 3D skeleton with true physical scale using perspective projection principles, thereby calculating the athlete's true physical height. and limb length Initial maximum speed and acceleration It can be preset to the limits of human movement and updated and tightened based on actual observations during subsequent tracking. This step ensures that the model is not just a general rigid body, but a digital avatar with specific biological characteristics, thus providing a strong geometric constraint basis when occlusion or confusion occurs later.

[0038] In step S200, the system performs tactical semantic reasoning based on the global state information of the previous time step and configures the corresponding prediction model. This process may specifically include the following steps: Construct a feature graph of multi-agent interactions across the entire field.

[0039] The system reads all athletes at any given time. posterior state vector set To capture the spatial interactions between athletes, the system constructs a dynamic topology graph. , where the node set Each node in the vector corresponds to an athlete, and the node feature vector... Includes the athlete's position, speed, posture parameters, and orientation, edge set Representing the interactive connections between athletes, the system calculates the interaction between any two athletes. and If the Euclidean distance between the points is less than a preset interaction threshold, a connection edge is established, and the edge features are defined. The relative position vector between the two is encoded. and relative velocity vector .

[0040] Joint reasoning based on tactical context and role category.

[0041] The constructed dynamic topology graph The input is fed into a pre-trained graph neural network model for inference. This model aggregates neighborhood information through a message-passing mechanism, updates the hidden state of each node, and contains two parallel output layers. One is a graph-level classifier used to output the macro-tactical context of the entire field at the current moment. The other is a node-level classifier, used to output the results for each athlete. Specific roles under this tactic Macro-tactical context The value space belongs to a predefined set. These correspond to tactical modes such as regular running, one-on-one, pick-and-roll, and fast break.

[0042] Character Category The value space of belongs to the set These correspond to tactical roles such as ball handler, screener, cutter, and stationary player. Through the above reasoning process, the system maps the underlying physical state to a higher-level tactical semantic description, formally represented as: in, The parameter is The graph neural network mapping function, for the specific network structure design and training method of the graph neural network, can be implemented by those skilled in the art using existing architectures such as graph convolutional networks or graph attention networks.

[0043] Construction of motion model library and retrieval of specific models.

[0044] The system has a pre-built library of motion models containing descriptions of various specific motion behaviors. Each element in the library This represents a specific state transition function used to describe the motion patterns under different tactical roles. (Model library) Mathematical models should include at least the following types: The first category is the general inertial model. The first type is suitable for normal running conditions, employing a constant velocity or constant acceleration model based on Langevin dynamics, allowing for larger process noise to accommodate random changes in direction; the second type is a high-dynamic sprint model. The first type is suitable for fast-paced or air-cutting characters. This model imposes a strong constraint on the velocity direction, aligning it with the body's orientation, and assumes that acceleration remains at a high level. The third type is the position-constrained model. This model is suitable for screeners or stationary pitchers. It assumes that the velocity approaches zero and imposes stability constraints on the posture parameters, limiting large joint movements.

[0045] Dynamic matching and instantiation of motion models.

[0046] Based on the output tactical context and roles The system uses table lookup or decision tree logic to retrieve data from the model library. For each athlete Find the best matching state transition function The selection logic is as follows: For example, when Identified as a pick-and-roll combination and When acting as a cover, the system selects a position constraint model. As a predictive model for the athlete's current moment; when When the player is in possession of the ball, either a general inertial model or a more maneuverable model should be chosen. The selected function... This will be directly loaded into the subsequent Kalman filter prediction step, so that the subsequent state prediction is no longer a blind physical extrapolation, but an intelligent deduction based on the prior knowledge of tactical intentions.

[0047] In step S300, the system uses the selected dynamic motion model to iteratively predict the state of each athlete. This process may specifically include the following steps: Generation of Sigma point sets.

[0048] For any athlete, obtain their position at the previous moment. The posterior state estimate mean and posterior error covariance matrix are used. Since the motion model involved in this invention has highly nonlinear characteristics, in order to accurately convey the statistical characteristics of the state, an unscented transformation method is adopted. Based on the state mean of the previous time step, the system generates a set of deterministic sampling points, namely Sigma points, symmetrically around the mean using the Cholesky decomposition results of the error covariance matrix. The total number of Sigma points depends on the dimension of the state vector. If the dimension of the state vector is L, the number of Sigma points generated is 2L+1. Each Sigma point represents a possible state instance of the system, and the weighted statistical characteristics of this set of points are completely consistent with the posterior distribution of the previous time step.

[0049] Nonlinear state propagation based on a specific motion model.

[0050] Each generated Sigma point is substituted into the state transition function selected for the athlete in step S200 for calculation. This step simulates the evolution of each possible state instance over time under given tactical intentions and physical constraints. Specifically, the state transition function calculates the predicted state at the next moment based on the position, velocity, attitude, and orientation components in the input Sigma point, combined with the athlete's individualized parameters.

[0051] Aggregate calculation of prior state mean and covariance.

[0052] After all Sigma points have propagated, the system performs weighted statistical analysis on this new set of propagated Sigma points. First, using preset mean weighting coefficients, the system performs a weighted summation on all propagated Sigma points, and the result is the current time step. The system calculates the prior state estimate mean, which represents the system's best prediction of the athlete's state based on tactical intent and physical laws before introducing visual observations in the current frame. Then, the system calculates the prior error covariance matrix, uses preset covariance weighting coefficients to calculate the deviation product between each propagation Sigma point and the prior state mean, and sums all the products with weights, while adding the process noise covariance matrix. The resulting prior error covariance matrix reflects the degree of uncertainty of the predicted state. This prior state mean and covariance matrix will be passed to the subsequent state update module as the prior input for the Bayesian filtering update step.

[0053] In step S400, the system uses the visual information of the current video frame to correct the prediction state. This process may specifically include the following steps: Extraction of multimodal visual observation data.

[0054] The system obtains the current time. The video frame images, and for each athlete Extract observation vectors describing its visual features. This observation vector contains two key types of data: The first type is two-dimensional joint coordinate data. The system uses a convolutional neural network to estimate the pose of the image and outputs the coordinates of human key points (such as shoulders, elbows, knees, ankles, etc.) in the pixel coordinate system. and its confidence level .

[0055] The second category is human body contour mask data. The system uses an instance segmentation network to extract pixel-level binary masks of athletes, or calculates the moment features of the mask as observation components, ultimately constructing an observation vector. It is a concatenation of the two types of data mentioned above. If some key points are occluded or the confidence level is lower than the threshold, they are marked as invalid in the observation vector and will not participate in the subsequent residual calculation.

[0056] Predictive observation generation based on physical models.

[0057] To compare the high-dimensional prior state with the two-dimensional visual observation, the system needs to construct an observation model. The state space is mapped to the observation space, and a virtual projection operation is performed on the state instance corresponding to each Sigma point using the propagation Sigma point set generated in step S300: First, based on the three-dimensional pose parameters at the Sigma point. and individualized morphological parameters The vertices of a three-dimensional human body mesh are reconstructed using a parametric human body model; Secondly, based on the court location in the Sigma point and orientation The reconstructed 3D mesh is transformed to its corresponding position in the world coordinate system; finally, the camera intrinsic parameter matrix obtained during the site calibration stage is used. The system uses the extrinsic parameter matrix to project the vertices of the 3D mesh onto the 2D image plane. Through this process, the system calculates the expected key point coordinates and contour projection of each Sigma point on the image, which constitutes the prediction observation set.

[0058] Calculate the predicted observation mean and covariance.

[0059] The generated predicted observation set is weighted statistically, and the mean of the predicted observation vector is calculated using the mean weights of the unscented transform. Using covariance weights, the autocovariance matrix of the predicted observations is calculated. This matrix reflects the uncertainty of the observation prediction. Simultaneously, the cross-covariance matrix between the state vector and the observation vector is calculated. This matrix describes the correlation between state variables and observed variables, in the calculation At this time, the observation noise covariance matrix needs to be superimposed. This is used to characterize the inherent measurement error of visual detection algorithms.

[0060] Posterior state correction and update.

[0061] The system calculates the actual observation vector. Compared with the predicted observation mean The difference between them, i.e., the information vector Then, the Kalman gain matrix is ​​calculated. : Finally, the prior state is corrected using Kalman gain to obtain the posterior state estimate at the current time step. : Simultaneously update the posterior error covariance matrix. : The posterior state This is the optimal state estimate that integrates tactical priors, physical constraints, and visual evidence. It serves as the starting point for the prediction of the next moment and is used to generate the final point map. This update mechanism ensures that even under partial occlusion or high detection noise, the system can still maintain stable tracking by relying on model prediction and tactical context.

[0062] In step S500, the system parses, maps, and formats the updated athlete status, which may specifically include the following steps: Precise calculation and extraction of foot contact point.

[0063] The system reads the output of each athlete in step S400. At the present moment posterior state vector According to the definition of a state vector, the system directly extracts the first two components, namely the position vector, from this high-dimensional vector. Because during the model building and updating process, this position vector The coordinates are always defined as the center point of the athlete's foot contact with the ground in the standard two-dimensional field coordinate system and have been corrected by multimodal observation. Therefore, the coordinates directly reflect the athlete's position on the real physical field without the need for perspective transformation from image pixels to physical coordinates. For athletes in a jumping or airborne state, the system calculates their projected position on the horizontal plane as the contact point based on the physical kinematic equations, or maintains their horizontal position at the moment of take-off until landing. The specific strategy is determined according to the preset tactical analysis requirements.

[0064] Point map generation and visualization rendering.

[0065] The system constructs a virtual standard basketball court plan model, whose dimensions are consistent with the actual court size, and extracts the position coordinates of all athletes. Mapped onto this virtual plane model, the system generates visual markers for each athlete and labels each marker with its identification identifier. Furthermore, the system can extract the velocity components from the state vector. The instantaneous movement direction and speed of each athlete are drawn on the point map in the form of arrow vectors. The virtual plane model containing the position points, ID identifiers and velocity vectors is rendered into an image format or vector graphic format, thus generating the athlete point map of the current frame.

[0066] Structured output of the complete semantic state vector.

[0067] In addition to the two-dimensional point map, the system also serializes and outputs the rich three-dimensional semantic information contained in the DKPM model to support advanced tactical analysis. The system traverses the posterior state vectors of all athletes and extracts the three-dimensional posture parameters. and body orientation quaternion This will include tags for position, speed, attitude, orientation, and tactical role. It is encapsulated into a standardized data structure that clearly records where the athlete is, which direction they are facing, their posture, and what tactical role they are performing. The output data stream can be transmitted to a downstream tactical analysis server via a network interface for reconstructing a 3D competition scene and generating running heatmaps.

Claims

1. A method for automatically generating a point graph of a basketball game player, characterized in that, Includes the following steps: The input competition video is initialized, and each athlete is detected and identified in the initial frame of the competition video. A dynamic kinematic posture model is instantiated for each athlete, and its individual kinematic constraint parameters are calibrated. Based on the state of all the athletes at the previous moment, analyze and identify the global tactical context of the current game and the tactical role played by each athlete. Based on the identified tactical role, a state transition function is adaptively selected for each athlete from a preset motion model library, and the athlete's state is predicted based on the state transition function to calculate the prior state at the current moment. Visual observation data related to each athlete is extracted from the current video frame, and the prior state is corrected using the visual observation data to calculate the posterior state of each athlete at the current moment. Extract the foot contact position coordinates in the standard court coordinate system from the posterior state of each athlete, generate the athlete position map of the current frame, and output it.

2. The method of claim 1, wherein, The state vector of the dynamic kinematic posture model includes: The athlete's position coordinates in the standard court coordinate system; The athlete's kinematic state, which includes velocity and acceleration components; Full-body 3D pose parameter vectors constructed based on a parametric human body model; The orientation of the athlete's main body in three-dimensional space, expressed as a unit quaternion.

3. The method for automatically generating a basketball game player position map according to claim 1, characterized in that, The individualized kinematic constraint parameters include: The athlete's height; The athlete's limb segment length; The athletes' historical maximum speed and maximum acceleration statistics.

4. The method for automatically generating a basketball game player position map according to claim 1, characterized in that, The steps of analyzing and identifying the global tactical context of the current match and the tactical role played by each athlete include: Based on the states of all the athletes at the previous moment, a dynamic topology graph containing node features and edge features is constructed. The dynamic topology graph is input into a pre-trained graph neural network model for inference. The global tactical context is output through the graph-level classifier of the graph neural network model, and the tactical role of each athlete is output through the node-level classifier of the graph neural network model.

5. The method for automatically generating a basketball game player position map according to claim 1, characterized in that, The motion model library includes at least: A general inertial model used to describe normal running conditions; A high-dynamic sprint model used to describe fast-paced or cut-and-run characters; and a positional constraint model used to describe cover or stationary characters.

6. The method for automatically generating a basketball game player position map according to claim 1, characterized in that, The step of predicting the athlete's state based on the state transition function is implemented using unscented Kalman filtering, specifically including: Based on the athlete's posterior state estimate mean and posterior error covariance matrix at the previous moment, a set of Sigma points is generated; Substitute the Sigma points into the selected state transition function for propagation to obtain a set of propagated Sigma points; The propagation Sigma point set is weighted and statistically analyzed to calculate the prior state at the current moment. The prior state includes the mean of the prior state estimate and the prior error covariance matrix.

7. The method for automatically generating a basketball game player position map according to claim 1, characterized in that, The visual observation data includes: Two-dimensional joint coordinate data of the athlete are extracted from the current video frame using a pose estimation algorithm; The athlete's human body contour mask data is extracted from the current video frame using an instance segmentation network.

8. The method for automatically generating a basketball game player position map according to claim 1, characterized in that, The step of correcting the prior state using the visual observation data includes: Construct an observation model that maps the state space to the observation space; The prior state is projected onto the observation model to obtain the predicted observation; The difference between the actual visual observation data and the predicted observation is calculated as information; Calculate the Kalman gain, and use the Kalman gain and the information to correct the prior state to obtain the posterior state.

9. The method for automatically generating a basketball game player position map according to claim 1, characterized in that, The specific steps for extracting the foot contact position coordinates in the standard court coordinate system from the posterior state of each athlete are as follows: Predefined position coordinate components are extracted directly from the state vector of the posterior state. The position coordinate components are defined as the vertical projection of the athlete onto the standard court plane.

10. The method for automatically generating a basketball game player position map according to claim 1, characterized in that, The step of generating and outputting the athlete position map for the current frame further includes: The three-dimensional pose parameters, body orientation, and tactical role contained in the posterior state are encapsulated into a standardized data structure and serialized for output.

Images