A badminton player cognitive ability evaluation method and system

Abstract

The application discloses a badminton player cognitive ability evaluation method and system, and belongs to the technical field of the cross of artificial intelligence, computer vision and sports science, and comprises the following steps: receiving single or multi-channel synchronous video streams shot by an ordinary camera, generating clue observation period and reaction period video clips with a hitting moment as an alignment benchmark through hitting event detection, round clip and space-time standardization processing. The clips are input into a multi-stage cognitive reasoning network comprising a clue focusing network with knowledge guidance, an asynchronous neural causal discovery mechanism and a scene adaptive evaluation gate, so that the following steps are sequentially completed: key biomechanical clue extraction, clue and reaction causal correlation modeling and scene evaluation correction. After being structured and integrated, a complete evaluation report is output. The system does not require professional hardware and has high universality, and can cover training institutions, colleges and universities and individual enthusiasts at all levels, and breaks through the limitations of traditional surface motion analysis, and accurately quantifies and predicts senior cognitive abilities such as efficiency and decision delay.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of artificial intelligence, computer vision, and sports science, specifically to a method and system for assessing the cognitive abilities of badminton players. Background Technology

[0002] Currently, although artificial intelligence technology has been widely applied to sports analysis, its development path has obvious limitations. Early sports analysis mainly relied on specialized hardware, such as high-speed infrared cameras, inertial sensing suits, or eye trackers. These devices can provide high-precision data, but they are expensive and complex to deploy, limiting their use to laboratories or professional clubs and making it difficult to popularize them in ordinary training.

[0003] At the algorithmic level, most existing research focuses on surface-level motion analysis. Using computer vision technology to identify athletes' striking actions, predict ball trajectories, or statistically analyze movement speeds has become a common application. However, these methods are essentially observations and descriptions of motion outcomes, failing to answer the core question of why athletes make certain decisions. They miss the crucial factors determining victory or defeat—the athletes' higher-level cognitive processes such as anticipation, attention allocation, and tactical decision-making.

[0004] The root cause of this limitation lies in the lack of modeling capabilities for cognitive processes in current technologies. In high-speed sports like badminton, athletes' decisions rely on a cognitive chain that involves instantly observing cues, anticipating intentions, and reacting. Existing methods cannot quantify this process from ordinary video footage. They struggle to consistently capture subtle, predictive biomechanical cues from opponents, and even more so, they cannot establish interpretable causal models linking these cues, the athlete's attention allocation, and the final reaction. Consequently, the system cannot diagnose cognitive shortcomings, such as whether a mistake is due to failing to notice cues or misinterpreting them. Summary of the Invention

[0005] On the one hand, in order to solve the problems of the prior art, the present invention provides a cognitive ability assessment system for badminton players, comprising: The data input and preprocessing module is used to generate standardized video segments based on the badminton video stream, with the moment of the hit as the alignment reference. The standardized video segments include cue observation period video segments and reaction period video segments. A multi-stage cognitive reasoning network module, connected to the data input and preprocessing module, analyzes and processes the standardized video segments; the multi-stage cognitive reasoning network module includes a knowledge-guided cue-focusing network, an asynchronous neural causal discovery mechanism unit, and a knowledge-embedded scene adaptive evaluation gating unit, which are connected in sequence. The knowledge-guided cue-focusing network is used to receive video clips during the cue observation period and extract cues, outputting a sequence of key biomechanical cues related to the intention to hit the ball; the asynchronous neural causal discovery mechanism is used to simultaneously receive the key biomechanical cue sequence and the video clips during the reaction period, perform causal modeling, and obtain the relationship between cues and reactions; the knowledge-embedded scene-adaptive evaluation gating unit performs scene-based evaluation based on the received information about the relationship between cues and reactions. The results generation and output module is connected to the knowledge-guided cue-focusing network. It is used to receive the scenario-based evaluation results output by the scenario-adaptive evaluation gating system with knowledge embedding, and after structured integration, output the final cognitive ability evaluation report.

[0006] Furthermore, the knowledge-guided clue-focusing network adopts a hierarchical dual-branch architecture; The dual-branch architecture includes: a general spatiotemporal feature extraction branch and a conditional modulation branch; The input end of the general spatiotemporal feature extraction branch is connected to the output end of the data input and preprocessing module. It is used to receive the tensor corresponding to the video segment during the clue observation period, and to extract spatiotemporal features from the tensor using the pre-trained VideoSwinTransformerTiny model to output a general spatiotemporal feature map. The input of the conditional modulation branch receives the ball-hitting type label corresponding to the video segment during the cue observation period. This ball-hitting type label is synchronously acquired by the data input and preprocessing module during the ball-hitting event detection process and converted into an independent encoded vector. The conditional modulation branch generates modulation parameters that match the channel dimension of the general spatiotemporal feature map after the independent encoded vector is mapped by the embedding layer and transformed by the multilayer perceptron. The general spatiotemporal feature map is dynamically adjusted by conditional instance normalization. The knowledge-guided cue-focusing network also includes: a knowledge-guided attention module and a three-dimensional convolutional layer; The input end of the knowledge-guided attention module is connected to the output end of the conditional modulation branch to receive the adjusted feature map, and at the same time generate an ideal clue query vector by combining the clue regions annotated by experts, calculate the spatial attention heatmap, and selectively enhance the adjusted feature map. The input of the three-dimensional convolutional layer is connected to the output of the knowledge-guided attention module, which performs dimensionality reduction and temporal encoding on the enhanced feature map and outputs a structured clue sequence. This clue sequence is transmitted to the asynchronous neural causal discovery mechanism through the connection relationship within the multi-stage cognitive reasoning network module.

[0007] Furthermore, the input end of the asynchronous neural causal discovery mechanism is connected to the output end of the knowledge-guided cue-focusing network and the output end of the data input and preprocessing module, respectively. The asynchronous neural causal discovery mechanism includes: a behavioral encoder and a neural causal discovery layer; The input end of the behavior encoder is connected to the output end of the data input and preprocessing module. It receives the tensor corresponding to the reaction period video segment, processes the tensor using the PoseC3D network, extracts human key points and generates a pseudo 3D heat map, and outputs a dynamic behavior sequence representing the reaction action after spatiotemporal convolution. The input end of the neural causal discovery layer is a dual-input structure, which is connected to the output end of the knowledge-guided cue-focusing network and the output end of the behavior encoder, respectively. It receives the cue sequence and the behavior sequence, and internally performs probabilistic delayed alignment, causal attention calculation and quantitative index generation operations in sequence, including the following steps; A fully connected layer is used to predict the probabilistic delay distribution for the feature vector at each time step in the clue sequence. Differentiable sampling is achieved using reparameterization techniques, and the cue sequence and behavior sequence are aligned on the time axis by utilizing sampling delay. Attention scores are calculated using an attention mechanism that embeds causal temporal masks. The attention weight matrix is ​​obtained by Softmax normalization, aggregated to generate a causal strength matrix, and the average decision delay is calculated. The asynchronous neural causal discovery mechanism optimizes the model by combining loss functions and transmits the causal strength matrix and average decision delay as quantitative indicators to the scene adaptive evaluation gating of knowledge embedding. The combined loss function includes: behavior prediction loss, sparsity loss, decoupling loss and delay smoothing loss, and its output is connected to the input of the knowledge embedding scene adaptive evaluation gating.

[0008] Furthermore, the input end of the scene adaptive evaluation gating of the knowledge embedding is connected to the output end of the asynchronous neural causal discovery mechanism and the external scene information input source, respectively; The knowledge-embedded scene adaptive evaluation gating includes a scene encoder, an expert rule matrix, and a differentiable knowledge gating; The scene encoder receives a sequence of round context features at its input, which includes the real-time score difference, the grid positions of both players on the field, the round duration, and the type of the previous shot. The scene encoder processes this sequence using a single-layer gated recurrent unit network and outputs a tactical situation coding vector and a scene category prediction probability distribution. The expert rule matrix is ​​a predefined 4×5 fixed matrix. Its rows correspond to four tactical scenarios: active attack, long rally, passive defense, and net battle. Its columns correspond to five cognitive evaluation indicators: prediction efficiency, clue utilization rate, decision delay, anti-deception, and scenario adaptability. The matrix elements are the relative importance weights of the indicators in each scenario, which are used to provide evaluation logic constraints for differentiable knowledge gating. The input end of the differentiable knowledge gating is a dual-input structure, which is connected to the output end of the scene encoder and the output end of the asynchronous neural causal discovery mechanism, respectively. It receives the tactical situation encoding vector and the original cognitive evaluation index vector, and generates scene-adaptive weight vector and bias vector through two independent fully connected layers to adaptively correct the original cognitive evaluation index vector. The scene-adaptive evaluation gating of the knowledge embedding uses the weighted sum of the scene classification cross-entropy loss and the knowledge embedding loss as the total training loss, guiding the gating weights to align with the expert rule matrix. Its output is connected to the input of the result generation and output module, transmitting the corrected final cognitive ability score to this module.

[0009] On the other hand, this application provides a method for assessing the cognitive abilities of badminton players, the method comprising the following steps; S1. Data preparation and preprocessing: First, a structured badminton video dataset is constructed and labeled with event-level, region-level, and scene-level annotations. Then, the original videos are processed sequentially with hit event detection, segment trimming, spatiotemporal standardization, and data normalization to obtain standardized cue observation period video tensors and reaction period video tensors. S2. Clue Focusing Extraction: Input the video tensor of the clue observation period output in step S1 into the knowledge-guided clue focusing network, and combine it with the prior knowledge of the ball type obtained from the event-level annotation in step S1 to extract the key biomechanical cue sequence related to the ball hitting intention. S3. Causal relationship modeling: The key biomechanical cue sequence related to the hitting intention output in step S2 and the reaction period video tensor output in step S1 are input into the asynchronous neural causal discovery mechanism. First, the reaction period video tensor is processed by the behavior encoder in the mechanism to obtain the behavior sequence. Then, the causal strength and decision time delay between the key biomechanical cue sequence related to the hitting intention and the behavior sequence are quantified by the neural causal discovery layer. S4. Scenario-based evaluation and correction: Input the round context information obtained from the scenario-level annotation in step S1 into the scenario-adaptive evaluation gate of the knowledge embedding, and combine it with the predefined expert rule matrix to perform scenario-adaptive correction on the original evaluation index output in step S3. S5. Generate assessment results. The corrected assessment indicators output in step S4 are structured and integrated to output a cognitive ability assessment report.

[0010] Furthermore, the three-level annotation process described in step S1, which is sequentially associated, includes: The event-level labeling is the first-level labeling, the core of which is to accurately identify and mark the moment and type of each shot. The types of shots are divided into 8 categories, including forehand high clear, backhand high clear, forehand smash, backhand smash, forehand drop shot, backhand drop shot, forehand net shot, and backhand net shot. The shot type label is converted into an independent encoding vector for use in subsequent steps. The regional-level annotation is the second-level annotation. Based on the event-level annotation at the moment of impact, within a time window of 0.5 seconds before impact, the spatiotemporal location of key biomechanical cues that indicate the opponent's intention to hit the ball is marked. The key biomechanical cues include wrist pronation, elbow elevation, shoulder rotation, non-racket hand depression, and knee flexion and extension angle changes. The scenario-level annotation is the third level of annotation. Based on the complete match round, combined with the shot type of the event-level annotation and the clue information of the area-level annotation, the tactical situation of the round is divided into four categories: active attack, long rally, passive defense, and net battle.

[0011] Furthermore, the spatiotemporal standardization and data normalization processes described in step S1 are preprocessing operations performed sequentially, including: The spatiotemporal standardization process is performed after segment cropping. First, all cropped video segments are uniformly scaled to a resolution of 224×224 pixels. Then, the frame rate of the video segments during the cue observation period is upsampled to 32fps using a time interpolation algorithm, and the frame rate of the video segments during the reaction period is normalized to 10fps using a time interpolation algorithm. The data normalization process is performed after the spatiotemporal standardization process, which linearly transforms the pixel values ​​of each frame in the spatiotemporally standardized video segments from the integer range of [0,255] to the floating-point range of [0,1].

[0012] Furthermore, the probabilistic delay distribution described in step S3 is a Gaussian distribution. ,in The mean of the delay, The standard deviation is obtained by mapping the feature vector of each time step in the cue sequence through a fully connected layer; the causal temporal mask is a logical constraint embedded in the attention mechanism to ensure that the cause of the current behavior is found only from past cues, that is, the aligned cue time is no later than the behavior occurrence time.

[0013] Furthermore, the original cognitive assessment index vector mentioned in step S4 is calculated from the causal strength matrix and average decision delay output in step S3; The scenario adaptive correction process achieved through differentiable knowledge gating includes: generating scenario adaptive weight vectors and bias vectors based on the tactical situation encoding vectors output by the scenario encoder, adjusting the original cognitive assessment index vectors by element-wise multiplication, and obtaining the final cognitive ability score that conforms to the expert assessment logic.

[0014] Furthermore, it also includes system training steps; The system training steps adopt a two-stage training strategy of phased pre-training and end-to-end joint fine-tuning. The first stage is a phased pre-training process, in which the knowledge-guided cue-focusing network in step S2, the asynchronous neural causal discovery mechanism in step S3, and the scene encoder in step S4 are independently supervised pre-trained. The knowledge-guided cue-focusing network is trained with cue region localization as the training objective, the asynchronous neural causal discovery mechanism is trained with behavior prediction as the training objective, and the scene encoder is trained with scene classification as the training objective. The second stage is end-to-end joint fine-tuning, which connects the pre-trained knowledge-guided cue-focusing network, the asynchronous neural causal discovery mechanism, and the knowledge-embedded scene adaptive evaluation gating to form a complete system. The comprehensive cognitive ability score given by experts is introduced as a global supervision signal, and the entire system is jointly optimized through a weighted summation total loss function to ensure that each module works together to output accurate cognitive ability assessment results.

[0015] The beneficial effects of this invention are: The assessment can be completed using only videos captured by a regular RGB camera, completely breaking down the barriers of professional equipment and enabling training institutions, schools, and even individual enthusiasts at all levels to easily obtain high-level sports intelligence diagnostics. It breaks through the limitations of surface-level action analysis, directly and non-invasively diagnosing advanced cognitive abilities such as prediction and decision-making, and attention allocation, achieving a paradigm shift from behavioral observation to cognitive diagnosis. The reasoning process simulates the logic of expert coaches, and the assessment results include causal relationships and scenario-based evidence, which can clearly point out the cognitive root causes of errors and provide precise targeted intervention suggestions for training. The three core modules are based on in-depth modifications of knowledge in the badminton field and are tightly coupled to form a complete reasoning chain of clue extraction, causal modeling, and scenario assessment, with technical effects far exceeding the simple superposition of a single module. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the evaluation system process provided by the present invention; Figure 2 A schematic diagram of the knowledge-guided clue-focusing network process provided for this invention; Figure 3 A schematic diagram of the asynchronous neural causal discovery mechanism provided by this invention; Figure 4 A schematic diagram of the scene-adaptive evaluation gating process for knowledge embedding provided by the present invention; Figure 5 This is a schematic diagram of the multi-stage cognitive reasoning network module framework provided by the present invention. Detailed Implementation

[0017] 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.

[0018] Please see Figure 1-5 This invention provides a cognitive ability assessment system for badminton players, comprising: The data input and preprocessing module receives single or multiple synchronous video streams captured by a regular camera (camera frame rate set between 30 and 60 frames per second), and sequentially performs ball-hitting event detection, rally segment cropping, and standardization processing on these streams to form standardized video segments aligned to the moment of ball hitting. These standardized video segments include cue observation period video segments and reaction period video segments. The specific cropping rule is: based on the instant of the i-th ball hit... Centered on, the time span is... The video clip, i.e., the time interval ,in Defined as the "clue observation period," it is used to analyze the opponent's predictive clues; Defined as the "reaction period," it is used to analyze and evaluate the initial reaction actions of the receiving party.

[0019] Next, spatiotemporal normalization processing is performed. Spatially, all captured video clips are uniformly scaled to [a specific size]. Pixel resolution, and Here, are hyperparameters (constants), representing the height and width of the standardized video frame, respectively. This scaling operation is performed using a bilinear interpolation function. To ensure consistency in input space dimensions, different resampling strategies are employed to address the varying temporal granularity requirements of the two analysis phases. For the video during the "cue observation period," which needs to capture rapid and subtle biomechanical changes, a temporal interpolation algorithm (such as linear or cubic spline interpolation) is used to uniformly upsample its frame rate to [a specific value]. This process can be represented as: ; : Input variable (four-dimensional tensor), representing the original input video stream. Input parameter (time interval): Defines the start and end time range of the "clue observation period" corresponding to the i-th shot. Input parameter (scalar constant), representing the target frame rate, i.e., the frame rate that the processed cue-time video needs to achieve; : Function / operation, representing time interpolation and resampling functions. The output variable (four-dimensional tensor) is the standardized video tensor of the "clue observation period" after processing.

[0020] For videos of the "reaction period," the focus is primarily on the macroscopic movement trends of the body's initiation; therefore, their frame rate is standardized to [specific value]. ,Right now: ; Finally, data normalization is performed. For each frame of the processed video clip, its pixel values ​​are normalized from the original... Perform a linear transformation on the integer range and normalize it to... The floating-point range. The specific formula is: ;in, The pixel matrix representing the original frame. This represents the normalized pixel matrix. After the complete preprocessing workflow described above, each shot event... Ultimately, it is transformed into two standardized input tensors: a tensor for cue analysis. (in ), and tensors used for reaction analysis (in This prepares the data for the subsequent end-to-end cognitive analysis of the core neural network.

[0021] A multi-stage cognitive reasoning network module, connected to the data input and preprocessing module, analyzes and processes the standardized video segments; the multi-stage cognitive reasoning network module includes a knowledge-guided cue-focusing network, an asynchronous neural causal discovery mechanism unit, and a knowledge-embedded scene adaptive evaluation gating unit, which are connected in sequence. Among them, the multi-stage cognitive reasoning network module receives data input and standardized video segments transmitted by the preprocessing module at its input end. It contains three core modules connected in sequence: a knowledge-guided cue-focusing network, an asynchronous neural causal discovery mechanism, and a knowledge-embedded scene adaptive evaluation gating. The knowledge-guided cue-focusing network receives video clips from the cue observation period and extracts cues, outputting a sequence of key biomechanical cues related to the ball-hitting intention. Its output is connected to the input of the asynchronous neural causal discovery mechanism. The asynchronous neural causal discovery mechanism simultaneously receives the key biomechanical cue sequence, the reaction period video clips, and the reaction period video clips transmitted from the data input and preprocessing module. It performs causal modeling to obtain the relationship between cues and responses. Its output is connected to the input of the knowledge-embedded scene adaptive evaluation gating mechanism. The knowledge-embedded scene adaptive evaluation gating mechanism receives quantitative indicators output by the asynchronous neural causal discovery mechanism and performs scene-based evaluation in conjunction with external scene information. The results generation and output module, whose input end is connected to the output end of the multi-stage cognitive reasoning network module, is used to receive the scenario-based evaluation results (including five core cognitive indicators scores: prediction efficiency, cue utilization rate, decision delay, anti-deception ability, and scenario adaptability) output by the scenario-adaptive evaluation gating of knowledge embedding. After structural integration, it outputs a final cognitive ability evaluation report containing cognitive strengths analysis, weakness identification, and targeted training suggestions.

[0022] In some embodiments, the knowledge-guided clue-focusing network adopts a hierarchical dual-branch architecture; The dual-branch architecture includes: a general spatiotemporal feature extraction branch and a conditional modulation branch; The input of the general spatiotemporal feature extraction branch is connected to the output of the data input and preprocessing module, and is used to receive the video tensor of the cue observation period. (in For time frames, (For spatial dimensions), the VideoSwin Transformer Tiny model, pre-trained on a large action recognition dataset, is used to extract spatiotemporal features from this tensor. A hierarchical window attention mechanism is used to upsample and downsample the input video in the spatiotemporal dimension, outputting a general spatiotemporal feature map. (in The dimension after downsampling. This represents the number of feature channels (e.g., 768). It encodes the low-level appearance and mid-level motion information of the video; The input of the conditional modulation branch receives the shot type label corresponding to the video segment during the cue observation period. This shot type label is synchronously acquired by the data input and preprocessing module during the shot event detection process (including 8 categories: forehand high clear, backhand high clear, forehand smash, backhand smash, forehand drop shot, backhand drop shot, forehand net shot, and backhand net shot), and converted into an independent encoded vector. The conditional modulation branch maps the independent coded vector into a continuous vector through a learnable embedding layer. After undergoing a nonlinear transformation using two layers of multilayer perceptrons (MLPs), modulation parameters matching the channel dimensions of the general spatiotemporal feature map are generated: scaling vector. and bias vector The general spatiotemporal feature map is dynamically adjusted through conditional instance normalization, and the adjustment formula is as follows: (in This represents a channel-by-channel multiplication operation. This represents the instance normalization operation. This operation applies normalization to each sample along the channel dimension. Calculate the mean and variance separately, normalize, and then use... and (Scaling and offsetting are performed). This design enables the network to dynamically adjust the feature distribution according to the type of shot, thereby producing differentiated responses to similar poses under different intentions, establishing a relationship between high-level semantics and low-level visual patterns at the feature level. The knowledge-guided cue-focusing network also includes: a knowledge-guided attention module and a three-dimensional convolutional layer; The input of the knowledge-guided attention module is connected to the output of the conditional modulation branch to receive the adjusted feature map. Simultaneously, it generates an ideal clue query vector by combining the clue areas annotated by experts. (The generation method is as follows: for each shot type, extract all corresponding cue region features from the labeled data, and then average them after global average pooling to obtain the prototype query vector for that type.) For the query vector, with Spatial features of bonds Sum The spatial attention heatmap is calculated using the following formula: ; in The dimension of the key vector (usually equal to) ), The scaling factor is used to selectively enhance the adjusted feature map using a feature enhancement formula, which is as follows: ; ( It is a learnable scalar parameter used to control the intensity of attention modulation; Represents attention weights With feature map Element-wise multiplication (multiplying by each other) Broadcast to (Multiply the corresponding values ​​after matching the same dimensions). This operation selectively enhances the features of key areas. For example, it strengthens the elbow and wrist area when analyzing a smash, while focusing on the fingers and racket face when analyzing a drop shot, achieving precise focus on clues without the need for annotation during the testing phase. The input of the three-dimensional convolutional layer is connected to the output of the knowledge-guided attention module, which processes the enhanced feature map. After dimensionality reduction and temporal encoding using a lightweight 3D convolutional layer, a structured cue sequence is output. ,in For timing length, As a feature dimension of the clues, the clue sequence is transmitted to the asynchronous neural causal discovery mechanism through the connection relationship within the multi-stage cognitive reasoning network module.

[0023] In some embodiments, the input of the asynchronous neural causal discovery mechanism is connected to the output of the knowledge-guided cue-focusing network and the output of the data input and preprocessing module, respectively. The asynchronous neural causal discovery mechanism includes: a behavioral encoder and a neural causal discovery layer; The input of the behavior encoder is connected to the output of the data input and preprocessing module, and receives the reaction period video tensor. ,in The tensor is processed using a lightweight PoseC3D network as a behavior encoder: First, a pose estimation algorithm (such as HRNet) is used to extract the two-dimensional coordinates of 17 human keypoints from each frame. These keypoint coordinates are then stacked in temporal order to form a spatial dimension of [missing information]. A pseudo-3D heatmap with channel dimension representing the number of keypoints is generated, and after spatiotemporal convolution, a dynamic behavioral sequence representing the reaction action is output. ,(in For time steps, (as a feature dimension), this sequence sequentially captures the athlete's key behavioral stages from the moment of hitting the ball, including reaction initiation, center of gravity shift, footwork adjustment, direction selection, and movement formation. The neural causal discovery layer has a dual-input structure, connected to the output of the knowledge-guided cue-focusing network and the output of the behavior encoder, respectively, to receive the cue sequence. and behavioral sequence ,in, : The feature vector of each time step in the sequence. For time steps, For the feature dimension, probabilistic delayed alignment, causal attention calculation, and quantization index generation operations are performed sequentially, including the following steps; A fully connected layer is used for each time step in the clue sequence. eigenvectors Predicting probabilistic delay distribution The mean delay and standard deviation Through formula Get ( (This is ensured to be positive via an exponential linear unit); differentiable sampling is performed using a reparameterization technique, with the sampling formula being... ;in ~N(0,1)), used directly during inference. As a delay estimate; the cue sequence and the behavior sequence are aligned on the time axis using the sampling delay, and the cue sequence is then interpolated using a linear interpolation function. Resampling to behavioral sequences Aligned cue features are obtained on the same time grid. The alignment operation formula is: ;in, Clue feature sequence; : The prediction delay time for the t-th clue time step. : The index or time coordinate of the original clue time step. : The set of aligned time points. Target time grid (time step index of the behavior sequence).

[0024] Attention scores are calculated using an attention mechanism that embeds causal temporal masks, based on behavioral features. As a query, aligned clue features As keys and values, the first The action time step and the first Each clue time step, its attention score The calculation formula is as follows; ; in These are the query and key vectors, respectively. This is a learnable weight matrix. Let the dimension be the vector. This is the scaling factor.

[0025] It is a causal mask, defined as ; The effective time of the j-th clue on the aligned timeline. : The absolute time of the i-th action time step; Based on time difference Gaussian kernel ( (bandwidth parameter); attention weight matrix obtained after Softmax normalization. (The normalization formula is) The weight matrix is ​​aggregated to generate a causal strength matrix. (its elements) Indicates the first The first clue is related to the first (Causal contribution strength of each behavioral stage), and calculate the average decision delay. ; The asynchronous neural causal discovery mechanism optimizes the model by combining loss functions and transmits the causal strength matrix and average decision delay as quantitative indicators to the scene adaptive evaluation gating of knowledge embedding. The combined loss function includes: behavior prediction loss. (in, Mean squared error function, calculates the mean squared error between two inputs; Causality strength matrix; Aligned cue feature matrix; :True behavioral feature matrix), sparse loss (in, Regularization strength coefficient, used to control the relative importance of sparsity penalty in the total loss function; decoupling loss. (in, : Decoupling regularization strength coefficient, used to control the weight of this loss in the total loss. : The j-th column of the causal strength matrix A. Pearson correlation coefficient (used to measure the degree of linear correlation between two vectors) and delay smoothing loss (in, The smoothing regularization strength coefficient controls the relative importance of this loss in the total loss. (The prediction delay value at the t-th clue time step), the total loss formula is: (in (This is a hyperparameter used to balance the importance of various losses), and its output is connected to the input of the knowledge-embedded scene adaptive evaluation gating.

[0026] In some embodiments, the input end of the knowledge-embedded scene adaptive evaluation gating is connected to the output end of the asynchronous neural causal discovery mechanism and an external scene information input source, respectively. The knowledge-embedded scene adaptive evaluation gating includes a scene encoder, an expert rule matrix, and a differentiable knowledge gating; The scene encoder receives a sequence of round-based context features at its input. The sequence includes real-time score differences. The sequence includes the players' positions on the court grid (dividing the court into 6 zones), the duration of the rally, and the type of the previous shot. The scene encoder uses a single-layer gated cyclic unit network (GRU) to process this sequence, with the following processing formula: )(in for The hidden state at all times (For GRU network parameters), take the hidden state at the last time step. As a tactical situation coding vector Simultaneously, it is connected through a fully connected layer and the Softmax activation function. Perform classification and output the predicted probability distribution of scene categories. (Corresponding to four types of tactical scenarios); The expert rule matrix is ​​a predefined 4×5 fixed matrix R∈R4×5, and the rows of the matrix correspond to four typical competition scenarios: ; The list corresponds to 5 core cognitive assessment indicators: M= ； Matrix elements The relative importance weights of the indicators in each scenario (all weights have been normalized to satisfy) ,like This indicates that the anti-deception weight is extremely high in passive defense scenarios. (This indicates that the anti-deception weight is lower in proactive attack scenarios), and is used to provide evaluation logic constraints for differentiable knowledge gating; The input of the differentiable knowledge gating is a dual-input structure, connected to the output of the scene encoder and the output of the asynchronous neural causal discovery mechanism, respectively, to receive the tactical situation encoding vector. and the vector of original cognitive assessment indicators From the causal strength matrix and average decision delay (Calculated from statistical measures), and scene-adaptive weight and bias vectors are generated through two independent fully connected layers. The generation formula is as follows: (in For learnable parameters, (Using the Sigmoid activation function to compress the weights to the (0,1) interval); the original cognitive assessment index vector is adaptively corrected using element-wise multiplication to obtain the final cognitive ability score. The correction formula is Efinal. ; This represents element-wise multiplication. This operation is equivalent to learning a scenario-dependent linear transformer for each cognitive indicator, dynamically adjusting its value and offset to match the actual meaning and evaluation criteria of the indicator under different tactical backgrounds.

[0027] The scene-adaptive evaluation gating for knowledge embedding uses a weighted sum of scene classification cross-entropy loss and knowledge embedding loss as the total training loss, guiding the gating weights towards the expert rule matrix. The scene classification cross-entropy loss is... ;in, : The loss value for the scene classification task, used to measure the difference between the tactical scene category predicted by the scene encoder and the true label. The true label value of the c-th category, encoded using one-hot encoding; The scene encoder predicts the probability that the current sample belongs to the c-th scene category, and normalizes it using the Softmax function. (strue is the real scene label). The knowledge embedding loss is... ( These are the row vectors in the expert rule matrix that correspond to the real-world scenario. (representing the Frobenius norm of the matrix), the total loss formula is: ( To balance hyperparameters (used to balance scene classification accuracy and knowledge alignment strength), its output is connected to the input of the result generation and output module, and the corrected final cognitive ability score is then output. Transmitted to this module.

[0028] In addition, this application provides a method for assessing the cognitive abilities of badminton players, the method comprising the following steps; S1. Data Preparation and Preprocessing: First, a structured badminton video dataset was constructed (data was collected in a standard badminton singles court using 1 to 2 ordinary RGB cameras fixed at the back of the court's central axis, with a frame rate of 30-60fps, capturing training and match recordings of athletes at different skill levels). The dataset was then labeled with event-level, region-level, and scene-level annotations (the annotation work was completed by a team of experienced badminton coaches). Next, the original videos were processed sequentially with hit event detection, segment cropping, spatiotemporal standardization, and data normalization to obtain standardized cue observation period video tensors and reaction period video tensors. S2. Clue Focusing and Extraction: The video tensor of the clue observation period output from step S1 is input into a knowledge-guided clue focusing network. Combined with the prior knowledge of the shot type obtained from the event-level annotation in step S1, key biomechanical cue sequences related to the shot intention are extracted through bi-branch feature extraction, knowledge-guided attention enhancement, and 3D convolutional coding. ; S3. Causal relationship modeling: The key biomechanical cues related to the hitting intention output in step S2 and the reaction period video tensor output in step S1 are input into the asynchronous neural causal discovery mechanism. First, the behavior encoder in this mechanism performs key point extraction, pseudo-3D heat map generation and spatiotemporal convolution processing on the reaction period video tensor to obtain the behavior sequence B. Then, the neural causal discovery layer performs probabilistic delay alignment, causal attention calculation and quantification index generation operations to quantify the causal strength and decision time delay between the cues and the behavior sequence, and outputs the causal strength matrix C and the average decision delay τ. S4. Contextualized evaluation and correction: The round context information obtained from the scenario-level annotation in step S1 is input into the knowledge-embedded scenario adaptive evaluation gating. Combined with the predefined expert rule matrix R, a tactical situation encoding vector is generated by the scenario encoder. Then, the differentiable knowledge gating generates a scenario adaptive weight vector and a bias vector. The original cognitive evaluation index vector output in step S3 is adaptively corrected to obtain the final cognitive ability score that conforms to the expert evaluation logic. ; S5. Generate evaluation results. The corrected evaluation indicators (including five core indicators: prediction efficiency, cue utilization rate, decision delay, anti-deception ability, and scenario adaptability) output in step S4 are structured and integrated to output a cognitive ability evaluation report that includes cognitive strength analysis, weakness identification (such as "significant delay in recognizing feint cues during passive defense"), and targeted training suggestions.

[0029] In some embodiments, the three-level annotation in step S1 is a sequentially associated annotation process including: The event-level annotation is the first level of annotation, and its core is to accurately identify and mark the moment of each shot. (This time point serves as the reference origin for all subsequent time alignments) and shot type The shot types are divided into 8 categories, forming a set of forehand high clear, backhand high clear, forehand smash, backhand smash, forehand drop shot, backhand drop shot, forehand net shot, and backhand net shot. The shot type label is converted into an independent encoded vector. For use in subsequent steps; The region-level annotation is the second-level annotation, based on the event-level annotation at the moment of impact. In the 0.5-second time window before the shot Within the game, the annotator observes the opposing player (the cue provider) frame by frame, identifying and marking the spatiotemporal locations of key biomechanical cues that predict the player's intention to hit the ball. These key biomechanical cues include wrist pronation, elbow elevation, shoulder rotation, non-racket hand depression, and knee flexion / extension angle changes. Each cue j is recorded using a spatiotemporal bounding box and is defined as follows: ,(in , This indicates the time interval in the video where the clue appears. (This represents its two-dimensional spatial boundary in the corresponding frame). The scenario-level annotation is the third level of annotation. Based on a complete match rally, it combines the shot types from event-level annotations and the clues from area-level annotations to divide the tactical situation of the rally into four categories: proactive attack, long rallies, passive defense, and net play, forming a set. .

[0030] In some embodiments, the spatiotemporal standardization and data normalization processes described in step S1 are preprocessing operations performed sequentially, including: The spatiotemporal normalization process is performed after fragment pruning, first through a bilinear interpolation function. All cropped video clips were scaled up uniformly to Pixel resolution ensures consistent input space size; then, time interpolation algorithms (linear or cubic spline interpolation) are used to upsample the frame rate of video segments during the clue observation period. The corresponding formula is ),in, : Input variable (four-dimensional tensor), representing the original input video stream. Input parameter (time interval): Defines the start and end time range of the "clue observation period" corresponding to the i-th shot. Input parameter (scalar constant), representing the target frame rate, i.e., the frame rate that the processed cue-time video needs to achieve; : Function / operation, representing time interpolation and resampling functions. The output variable (a four-dimensional tensor) is the processed, standardized video tensor from the "cue observation period". The frame rate of the reaction period video segments is standardized using a temporal interpolation algorithm. The corresponding formula is ;in, : Input variable (four-dimensional tensor), representing the original input video stream. Input parameter (time interval): Defines the start and end time range of the "reaction period" corresponding to the i-th shot. : Input parameter (scalar constant), representing the target frame rate. The output variable (four-dimensional tensor) is the standardized "reaction period" video tensor after processing. : Function / operation, representing the time resampling function.

[0031] The data normalization process is performed after the spatiotemporal normalization process. It linearly transforms the pixel values ​​of each frame in the spatiotemporally normalized video segment from the integer range of [0, 255] to the floating-point range of [0, 1]. The specific formula is as follows: (in The pixel matrix representing the original frame. (Represents the normalized pixel matrix).

[0032] In some embodiments, the probabilistic delay distribution in step S3 is a Gaussian distribution. The feature vectors at each time step in the cue sequence are processed through fully connected layers. The mapping is obtained by mapping, and the mapping formula is [ ); The causal timing mask is a logical constraint embedded in the attention mechanism, used to ensure that the cause of the current behavior is sought only from past cues, i.e., the aligned cue time is no later than the behavior occurrence time. It is defined as follows: ; The attention score calculation also incorporates a Gaussian kernel term based on time difference. , ( (This is a bandwidth parameter) is used to encourage stronger associations between temporally adjacent cue-behavior pairs, which aligns with the principle that the influence of cues decays over time in cognitive processing.

[0033] In some embodiments, the original cognitive assessment index vector in step S4 The causal strength matrix C and average decision delay τ, etc., output from step S3, are calculated and cover five core cognitive indicators: prediction efficiency, clue utilization rate, decision delay, anti-deception, and scenario adaptability. The scene adaptive correction is achieved through differentiable knowledge gating, which includes: differentiable knowledge gating based on the tactical situation encoding vector output by the scene encoder. Scene-adaptive weight vectors are generated through two independent fully connected layers. and bias vector (The formula for generation is) ,in , The weight matrices of the two fully connected layers. Used to generate weight vectors Used to generate bias vectors. , : Bias vectors of two fully connected layers The Sigmoid function adjusts the original cognitive assessment index vector using element-wise multiplication. The adjustment formula is as follows: (in, : Vector of original cognitive assessment indicators. Element-wise multiplication (Hadamard product) yields the final cognitive ability score that conforms to the expert assessment logic. ; Knowledge embedding loss during the correction process (in, The row vectors corresponding to real-world scenarios in the expert rule matrix guide the gating weights toward the expert rule matrix, ensuring that the evaluation results conform to professional logic.

[0034] In some embodiments, the system training steps employ a two-stage training strategy of phased pre-training and end-to-end joint fine-tuning. The first stage is phased pre-training, which involves independent supervised pre-training of the knowledge-guided cue-focusing network in step S2, the asynchronous neural causal discovery mechanism in step S3, and the scene encoder in step S4. The knowledge-guided cue-focusing network uses cue region localization as its training objective, and its loss function is defined as the prediction heatmap. With real labeling The mean square error between them is given by the formula: This forces the network to accurately focus on key biomechanical areas under the guidance of knowledge about shot types; among them, : Batch size, k represents the k-th sample; The asynchronous neural causal discovery mechanism uses behavior prediction as the training objective, and for now only uses behavior prediction loss, the formula is as follows: (in, Initial prediction loss, used to pre-train the behavior prediction capabilities of the causal discovery module. The total number of time steps in the reaction behavior sequence. The feature dimension of the feature vector at each time step. The feature vector of the reaction behavior predicted by the model at the i-th time step. The true response behavior feature vector at the i-th time step enables the model to initially establish a mapping ability from cues to behavior; the scene encoder uses scene classification as its training objective and employs standard cross-entropy loss, as shown in the formula. This ensures that the encoder can accurately understand the overall tactical situation of the game; The second phase involves end-to-end joint fine-tuning, which connects the pre-trained knowledge-guided cue-focusing network, the asynchronous neural causal discovery mechanism, and the scene-adaptive evaluation gating of knowledge embedding into a complete system, and incorporates a comprehensive cognitive ability score from multiple expert coaches. As a global monitoring signal, the entire system is jointly optimized using a weighted summation total loss function, and the total loss formula is as follows: ,(in To balance hyperparameters, For the system's final prediction score With expert rating (mean squared error), to ensure that each module works together to output accurate cognitive ability assessment results; The training configuration is as follows: Optimizer: Using AdamW, weight decay is set to ; Learning rate scheduling: A cosine annealing strategy is used, with an initial learning rate of... This will make its descent smoother; Gradient clipping: The gradient norm threshold is set to 1.0 to prevent training instability.

[0035] The system's training configuration revolves around the goals of multi-module collaborative optimization and accurate quantification of cognitive capabilities. It adopts a two-stage training strategy of phased pre-training and end-to-end joint fine-tuning, coupled with an optimizer adapted to network characteristics, learning rate scheduling, and gradient constraint scheme.

[0036] The two-stage training strategy begins with phased pre-training, where independent supervised pre-training is conducted on the three core modules of the multi-stage cognitive reasoning network. The knowledge-guided cue-focusing network uses cue region localization as its training objective, employing the mean squared error between the predicted heatmap and the actual annotation as its loss function, forcing the network to accurately focus on biomechanical cue regions related to the intention to hit the ball. The asynchronous neural causal discovery mechanism uses behavior prediction as its training objective, temporarily using the behavior prediction loss function to initially establish the mapping relationship between cues and reaction behaviors. The scene encoder uses scene classification as its training objective, employing standard cross-entropy loss to ensure it can accurately identify four types of tactical scenarios, including proactive attacks and long rallies.

[0037] The second stage is end-to-end joint fine-tuning, which connects the three core modules after pre-training to form a complete system. It introduces the comprehensive cognitive ability scores given by multiple expert coaches as global supervision signals, and performs joint optimization of the entire system through a weighted summation total loss function, so that each module can work together to adapt and output accurate cognitive assessment results.

[0038] The optimizer configuration uses the AdamW optimizer with a weight decay coefficient set to 1e-4. This optimizer combines Adam's adaptive learning rate feature with the regularization effect of weight decay. It can dynamically adjust the learning rate according to the parameter update requirements of different modules such as the cue-focusing network and the causal discovery mechanism to achieve fast convergence. At the same time, it can suppress the overfitting risk of complex network structures through L2 regularization constraints, especially for complex components such as two-branch feature extraction and causal attention layers, effectively avoiding the model from overfitting on the training data.

[0039] The learning rate scheduling scheme adopts a cosine annealing learning rate scheduling strategy, with an initial learning rate set to 1e-4, which decreases smoothly according to a cosine curve with each training epoch. In the early stage of training, the higher initial learning rate supports each independent module to quickly explore the parameter space and adapt to the initial optimization goals such as cue localization and scene classification. In the later stage of training, the learning rate decreases smoothly to achieve fine-tuning of parameters, avoiding model oscillation due to an excessively high learning rate or convergence stagnation due to an excessively low learning rate. This perfectly adapts to the continuous value regression characteristics of cognitive evaluation indicators, resulting in higher accuracy and stability of the quantification results.

[0040] The gradient clipping constraint sets the gradient norm threshold to 1.0 to clip gradients during backpropagation. Addressing the gradient explosion problem that can arise from complex operations such as probabilistic delayed alignment and causal attention computation in asynchronous neural causal discovery mechanisms, this constraint effectively suppresses abnormal gradient growth, preventing model training crashes. Simultaneously, it ensures the stability of gradient propagation when multiple modules are connected in series, preventing gradient anomalies in a single module from affecting overall network optimization and making the training process smoother.

[0041] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cognitive ability assessment system for badminton players, characterized in that, include: The data input and preprocessing module is used to generate standardized video segments based on the badminton video stream, with the moment of the hit as the alignment reference. The standardized video segments include cue observation period video segments and reaction period video segments. A multi-stage cognitive reasoning network module is connected to the data input and preprocessing module to analyze and process the standardized video segments; the multi-stage cognitive reasoning network module includes a knowledge-guided cue-focusing network, an asynchronous neural causal discovery mechanism unit, and a knowledge-embedded scene adaptive evaluation gating unit connected in sequence. The knowledge-guided cue-focusing network is used to receive video clips during the cue observation period and extract cues, outputting a sequence of key biomechanical cues related to the ball-hitting intention; the asynchronous neural causal discovery mechanism is used to simultaneously receive the key biomechanical cue sequence and the video clips during the reaction period, perform causal modeling, and obtain the relationship between cues and reactions; the knowledge-embedded scene-adaptive evaluation gating unit performs scene-based evaluation based on the relationship information between the received cues and reactions. The result generation and output module is connected to the knowledge-guided clue-focusing network. It is used to receive the scenario-based evaluation results output by the knowledge-embedded scenario adaptive evaluation gating unit, and after structural integration, output the final cognitive ability evaluation report. The knowledge-guided clue-focusing network adopts a hierarchical dual-branch architecture; The dual-branch architecture includes: a general spatiotemporal feature extraction branch and a conditional modulation branch; The input end of the general spatiotemporal feature extraction branch is connected to the output end of the data input and preprocessing module. It is used to receive the tensor corresponding to the video segment during the clue observation period, and to extract spatiotemporal features from the tensor using the pre-trained VideoSwin Transformer Tiny model to output a general spatiotemporal feature map. The input of the conditional modulation branch receives the ball-hitting type label corresponding to the video segment during the cue observation period. This ball-hitting type label is synchronously acquired by the data input and preprocessing module during the ball-hitting event detection process and converted into an independent encoded vector. The conditional modulation branch generates modulation parameters that match the channel dimension of the general spatiotemporal feature map after the independent encoded vector is mapped by the embedding layer and transformed by the multilayer perceptron. The general spatiotemporal feature map is dynamically adjusted by conditional instance normalization. The knowledge-guided cue-focusing network also includes: a knowledge-guided attention module and a three-dimensional convolutional layer; The input end of the knowledge-guided attention module is connected to the output end of the conditional modulation branch to receive the adjusted feature map, and at the same time generate an ideal clue query vector by combining the clue regions annotated by experts, calculate the spatial attention heatmap, and selectively enhance the adjusted feature map. The input of the three-dimensional convolutional layer is connected to the output of the knowledge-guided attention module to perform dimensionality reduction and temporal encoding on the enhanced feature map and output a structured cue sequence. This cue sequence is transmitted to the asynchronous neural causal discovery mechanism through the connection relationship within the multi-stage cognitive reasoning network module. The input of the asynchronous neural causal discovery mechanism is connected to the output of the knowledge-guided cue-focusing network and the output of the data input and preprocessing module, respectively. The asynchronous neural causal discovery mechanism includes: a behavioral encoder and a neural causal discovery layer; The input end of the behavior encoder is connected to the output end of the data input and preprocessing module. It receives the tensor corresponding to the reaction period video segment, processes the tensor using the PoseC3D network, extracts human key points and generates a pseudo 3D heat map, and outputs a dynamic behavior sequence representing the reaction action after spatiotemporal convolution. The input of the neural causal discovery layer is a dual-input structure, connected to the output of the knowledge-guided cue-focusing network and the output of the behavior encoder, respectively. It receives the cue sequence and the behavior sequence, and internally performs probabilistic delayed alignment, causal attention calculation, and quantitative indicator generation operations in sequence, including the following steps: A fully connected layer is used to predict the probabilistic delay distribution for the feature vector at each time step in the clue sequence. Differentiable sampling is achieved using reparameterization techniques, and sampling delay is used to align the cue sequence and behavior sequence on the time axis; among which, t Indicates the time step of the clue sequence; Indicates the first t Predictive delay variables at each time step; This represents the mean delay of the predicted delay variable; The standard deviation of the predicted delay variable is represented by N; N represents a Gaussian distribution. Attention scores are calculated using an attention mechanism that embeds causal temporal masks. The attention weight matrix is ​​obtained by Softmax normalization, aggregated to generate a causal strength matrix, and the average decision delay is calculated. The asynchronous neural causal discovery mechanism optimizes the model by combining loss functions and transmits the causal strength matrix and average decision delay as quantitative indicators to the scene adaptive evaluation gating of knowledge embedding. The combined loss function includes: behavior prediction loss, sparsity loss, decoupling loss and delay smoothing loss, and its output is connected to the input of the knowledge embedding scene adaptive evaluation gate. The input end of the knowledge-embedded scene adaptive evaluation gating is connected to the output end of the asynchronous neural causal discovery mechanism and the external scene information input source, respectively. The knowledge-embedded scene adaptive evaluation gating includes a scene encoder, an expert rule matrix, and a differentiable knowledge gating; The scene encoder receives a sequence of round context features at its input, which includes the real-time score difference, the grid positions of both players on the field, the round duration, and the type of the previous shot. The scene encoder processes this sequence using a single-layer gated recurrent unit network and outputs a tactical situation coding vector and a scene category prediction probability distribution. The expert rule matrix is ​​a predefined 4×5 fixed matrix. The rows of the matrix correspond to four tactical scenarios: active attack, long rally, passive defense, and net battle. The columns correspond to five cognitive evaluation indicators: prediction efficiency, clue utilization rate, decision delay, anti-deception, and scenario adaptability. The matrix elements are the relative importance weights of the indicators in each scenario, which are used to provide evaluation logic constraints for differentiable knowledge gating. The input end of the differentiable knowledge gating is a dual-input structure, which is connected to the output end of the scene encoder and the output end of the asynchronous neural causal discovery mechanism, respectively. It receives the tactical situation encoding vector and the original cognitive evaluation index vector, and generates scene-adaptive weight vector and bias vector through two independent fully connected layers to adaptively correct the original cognitive evaluation index vector. The scene-adaptive evaluation gating of the knowledge embedding uses the weighted sum of the scene classification cross-entropy loss and the knowledge embedding loss as the total training loss, guiding the gating weights to align with the expert rule matrix. Its output is connected to the input of the result generation and output module, transmitting the corrected final cognitive ability score to this module.

2. A method for assessing the cognitive abilities of badminton players, characterized in that, The system for assessing the cognitive abilities of badminton players, as described in claim 1, is used, and the method includes the following steps: S1. Data preparation and preprocessing: First, a structured badminton video dataset is constructed and labeled with event-level, region-level, and scene-level annotations. Then, the original videos are processed sequentially with hit event detection, segment trimming, spatiotemporal standardization, and data normalization to obtain standardized cue observation period video tensors and reaction period video tensors. S2. Clue Focusing Extraction: Input the video tensor of the clue observation period output in step S1 into the knowledge-guided clue focusing network, and combine it with the prior knowledge of the ball type obtained from the event-level annotation in step S1 to extract the key biomechanical cue sequence related to the ball hitting intention. S3. Causal relationship modeling: The key biomechanical cue sequence related to the hitting intention output in step S2 and the reaction period video tensor output in step S1 are input into the asynchronous neural causal discovery mechanism. First, the reaction period video tensor is processed by the behavior encoder in the mechanism to obtain the behavior sequence. Then, the causal strength and decision time delay between the key biomechanical cue sequence related to the hitting intention and the behavior sequence are quantified by the neural causal discovery layer. S4. Scenario-based evaluation and correction: Input the round context information obtained from the scenario-level annotation in step S1 into the scenario-adaptive evaluation gate of the knowledge embedding, and combine it with the predefined expert rule matrix to perform scenario-adaptive correction on the original evaluation index output in step S3. S5. Generate assessment results. The corrected assessment indicators output in step S4 are structured and integrated to output a cognitive ability assessment report.

3. The method for assessing the cognitive abilities of badminton players according to claim 2, characterized in that, The three-level annotation process described in step S1, which involves sequential association, includes: The event-level labeling is the first-level labeling, the core of which is to accurately identify and mark the moment and type of each shot. The types of shots are divided into 8 categories, including forehand high clear, backhand high clear, forehand smash, backhand smash, forehand drop shot, backhand drop shot, forehand net shot, and backhand net shot. The shot type label is converted into an independent encoding vector for use in subsequent steps. The regional-level annotation is the second-level annotation. Based on the event-level annotation at the moment of impact, within a time window of 0.5 seconds before impact, the spatiotemporal location of key biomechanical cues that indicate the opponent's intention to hit the ball is marked. The key biomechanical cues include wrist pronation, elbow elevation, shoulder rotation, non-racket hand depression, and knee flexion and extension angle changes. The scenario-level annotation is the third level of annotation. Based on the complete match round, combined with the shot type of the event-level annotation and the clue information of the area-level annotation, the tactical situation of the round is divided into four categories: active attack, long rally, passive defense, and net battle.

4. The method for assessing the cognitive abilities of badminton players according to claim 2, characterized in that, The spatiotemporal standardization and data normalization processes described in step S1 are preprocessing operations performed sequentially, including: The spatiotemporal standardization process is performed after segment cropping. First, all cropped video segments are uniformly scaled to a resolution of 224×224 pixels. Then, the frame rate of the video segments during the cue observation period is upsampled to 32fps using a time interpolation algorithm, and the frame rate of the video segments during the reaction period is normalized to 10fps using a time interpolation algorithm. The data normalization process is performed after the spatiotemporal standardization process, which linearly transforms the pixel values ​​of each frame in the spatiotemporally standardized video segments from the integer range of [0,255] to the floating-point range of [0,1].

5. The method for assessing the cognitive abilities of badminton players according to claim 2, characterized in that, The probabilistic delay distribution mentioned in step S3 is a Gaussian distribution. ,in The mean of the delay, The standard deviation is obtained by mapping the feature vector of each time step in the cue sequence through a fully connected layer; the causal temporal mask is a logical constraint embedded in the attention mechanism to ensure that the cause of the current behavior is found only from past cues, that is, the aligned cue time is no later than the behavior occurrence time.

6. The method for assessing the cognitive abilities of badminton players according to claim 2, characterized in that, The original cognitive assessment index vector mentioned in step S4 is calculated from the causal strength matrix and average decision delay output in step S3; The process of scene adaptive correction achieved through differentiable knowledge gating includes: differentiable knowledge gating generates scene adaptive weight vectors and bias vectors based on the tactical situation encoding vectors output by the scene encoder, and adjusts the original cognitive assessment index vectors by element-wise multiplication to obtain the final cognitive ability score that conforms to the expert assessment logic.

7. The method for assessing the cognitive abilities of badminton players according to claim 2, characterized in that, It also includes system training steps; The system training steps adopt a two-stage training strategy of phased pre-training and end-to-end joint fine-tuning. The first stage is a phased pre-training process, in which the knowledge-guided cue-focusing network in step S2, the asynchronous neural causal discovery mechanism in step S3, and the scene encoder in step S4 are independently supervised pre-trained. The knowledge-guided cue-focusing network is trained with cue region localization as the training objective, the asynchronous neural causal discovery mechanism is trained with behavior prediction as the training objective, and the scene encoder is trained with scene classification as the training objective. The second stage is end-to-end joint fine-tuning, which connects the pre-trained knowledge-guided cue-focusing network, the asynchronous neural causal discovery mechanism, and the knowledge-embedded scene adaptive evaluation gating to form a complete system. The comprehensive cognitive ability score given by experts is introduced as a global supervision signal, and the entire system is jointly optimized through a weighted summation total loss function to ensure that each module works together to output accurate cognitive ability assessment results.

Images