Event tactical mining method and system based on multi-granularity visual perception and timing causal reasoning

Abstract

The present application relates to a kind of event tactics mining method and system based on multi-granularity visual perception and timing causal inference.The method comprises: obtaining the event video stream of multi-view acquisition and pre-processing, to obtain standardized video frame data;Multi-granularity perception data is obtained using group equivalent convolution network, capsule network and slot attention mechanism;Optimal transport matching method is used to realize the alignment of object identity between adjacent frames, combined with neural differential equation for continuous space-time modeling, to construct dynamic space-time atlas;Neural point process is used for tactical timing modeling, and the tactic rule constraint is carried out using differentiable logic network, and the causal relationship between tactic behaviors is mined using causal graph discovery algorithm, to obtain tactic semantic result;Uncertainty estimation is carried out using deep set inference, and encapsulated as standardized interface, to output tactic mining result;Tactics information in event video can be effectively extracted, and high-level reasoning and analysis are carried out.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning. Background Technology

[0002] With the digital development of sports events, traditional event analysis heavily relies on professionals manually watching videos, marking player movements, and analyzing tactical combinations. This approach suffers from inefficiency, strong subjectivity, limited analytical dimensions, and difficulty in capturing dynamic spatiotemporal relationships, making it unable to meet the real-time analysis needs of high-intensity, fast-paced events. Existing video analysis technologies mostly remain at a low level of information extraction, lacking a deep understanding of spatiotemporal changes during the game. Traditional intelligent video analysis technologies can achieve basic functions such as player detection, trajectory tracking, and motion recognition, primarily based on traditional computer vision-based target detection and tracking, deep learning-based motion recognition, and simple spatiotemporal data analysis. Specifically, traditional computer vision-based target detection and tracking utilizes algorithms such as OpenCV, YOLO, or SORT to locate and select basic targets like players and balls in the event video; deep learning-based motion recognition uses convolutional neural networks or recurrent neural networks to classify and identify specific actions in the video; and simple spatiotemporal data analysis involves constructing spatiotemporal grids or basic graph structures to analyze player running heatmaps or the overall formation distribution of a team.

[0003] However, traditional intelligent analysis technologies for sports videos mostly use ordinary convolutional neural networks, which lack visual perception capabilities; they perform feature matching and trajectory stitching based on discrete time frames, resulting in weak spatiotemporal modeling capabilities and a lack of continuous modeling capabilities for the spatiotemporal evolution of sports events; and they can only identify basic actions such as passing and shooting, lacking in-depth reasoning on tactical triggering rules, logical constraints, and causal relationships; they often only provide tactical recognition results.

[0004] Therefore, traditional intelligent analysis technology for sports videos, which relies solely on ordinary convolutional networks for shallow visual feature extraction, struggles to accurately analyze multi-granular visual information from complex sports environments, often resulting in poor tactical reasoning logic and low reliability of output results. Summary of the Invention

[0005] Based on this, in order to solve the above-mentioned technical problems, a method and system for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning is provided. This method can effectively extract tactical information from competition videos and perform high-level reasoning and analysis, providing intelligent support for competition commentary, audience experience, and coach tactical review.

[0006] A method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning, the method comprising:

[0007] Acquire event video streams captured from multiple perspectives, and preprocess the event video streams to obtain standardized video frame data;

[0008] By utilizing group equivariant convolutional networks, capsule networks, and slot attention mechanisms, feature extraction and object recognition are performed on the standardized video frame data to obtain multi-granularity perceptual data that fuses geometric equivariant features, hierarchical pose features, and semantic instantiation features.

[0009] For the multi-granularity sensing data, the optimal transmission matching method is used to align the identities of objects between adjacent frames. Continuous spatiotemporal modeling is performed in combination with neural differential equations to construct a dynamic spatiotemporal graph containing nodes, edges, and a weighted adjacency matrix.

[0010] For the dynamic spatiotemporal graph, tactical temporal modeling is performed using neural point processes, tactical rule constraints are performed using differentiable logic networks, and causal graph discovery algorithms are used to mine causal relationships between tactical behaviors to obtain tactical semantic results.

[0011] Based on the tactical semantic results, an interpretable evidence subgraph is generated, and uncertainty estimation is performed using deep set inference. The results are then encapsulated into a standardized interface and the tactical mining results are output.

[0012] In one embodiment, a multi-view captured video stream of the event is acquired, and the video stream is preprocessed to obtain standardized video frame data, including:

[0013] Multi-view video streams of the event are captured by synchronous cameras deployed in various locations of the event venue, and the original frame sequence is obtained after decoding and timestamp alignment.

[0014] The original frame sequence is denoised by Gaussian blur, and key frames containing key actions and tactical events are extracted from the denoised frame sequence.

[0015] The brightness and contrast of the keyframes are adjusted to obtain standardized video frame data.

[0016] In one embodiment, a group-variable convolutional network, a capsule network, and a slot attention mechanism are used to extract features and identify objects from the standardized video frame data, resulting in multi-granularity perceptual data fused from geometrically variable features, hierarchical pose features, and semantic instantiation features, including:

[0017] The standardized video frame data is input into a group equivariant convolutional network, and convolutional kernels are defined on continuous groups to extract features, thereby obtaining geometrically equivariant features.

[0018] The geometric equilateral features are reshaped into capsule vectors, and the dynamic routing algorithm of the capsule network is used to capture the target hierarchical relationship and output the hierarchical pose features.

[0019] The hierarchical pose features are input into the slot attention mechanism, and the features are iteratively aggregated through learnable slot variables to realize object instantiation and output semantic instantiation features.

[0020] The geometric isomorphic features, hierarchical pose features, and semantic instantiation features are weighted and fused to obtain multi-granularity perception data.

[0021] In one embodiment, for the multi-granularity sensing data, an optimal transmission matching method is used to align object identities between adjacent frames. Continuous spatiotemporal modeling is performed using neural differential equations to construct a dynamic spatiotemporal graph containing nodes, edges, and a weighted adjacency matrix, including:

[0022] Based on the multi-granularity sensing data, the object set in adjacent frames is extracted, and the matching cost between each object in adjacent frames is calculated to construct a matching cost matrix.

[0023] Regularization is introduced and the Sinkhorn algorithm is used to iteratively solve the matching cost matrix to obtain the optimal transmission matrix, thereby achieving object identity alignment between adjacent frames.

[0024] Determine the object state after identity alignment, and construct a neural differential equation with the object state as input and the state change rate as output;

[0025] Graph nodes are created based on the aligned objects, graph edges are established according to the spatial distance and semantic relationship between nodes, and weights representing the closeness of tactical cooperation are dynamically calculated through graph attention mechanism to construct a weighted adjacency matrix;

[0026] The neural differential equation is used to continuously update the node state and synchronously adjust the edge connection relationship and weighted adjacency matrix to generate a dynamic spatiotemporal graph that evolves over time.

[0027] In one embodiment, for the dynamic spatiotemporal graph, tactical temporal modeling is performed using neural point processes, tactical rule constraints are applied using differentiable logic networks, and causal graph discovery algorithms are used to mine causal relationships between tactical behaviors to obtain tactical semantic results, including:

[0028] Tactical events are extracted from the dynamic spatiotemporal graph, and the sequence of historical tactical events is encoded using neural point processes. The conditional strength function that changes over time is learned to predict the instantaneous probability of future tactical events.

[0029] Logical rules are constructed based on preset tactical knowledge, and these logical rules are embedded in a differentiable logic network. The differentiable logic network is used to impose rule constraints on the tactical reasoning process and correct the output logical patterns.

[0030] Based on the dynamic spatiotemporal graph, a causal graph discovery algorithm is used to learn a structured causal graph that represents the causal relationship between tactical behaviors by optimizing the objective function and imposing a directed acyclic graph constraint.

[0031] The instantaneous probability, logical rules, and structured causal graph are integrated to generate tactical semantic results.

[0032] In one embodiment, an interpretable evidence subgraph is generated based on the tactical semantic results, and uncertainty estimation is performed using deep ensemble inference, including:

[0033] For the logical patterns and structured causal graphs in the tactical semantic results, key nodes and paths are selected by combining temporal correlation, and interpretable evidence subgraphs are generated.

[0034] Construct independent sub-models with different initial weights, input the tactical semantic results into each independent sub-model, and obtain the tactical prediction mean and variance based on the output of each independent sub-model;

[0035] Based on the mean and variance of the tactical predictions, an uncertainty estimation index is obtained, thus completing the uncertainty estimation.

[0036] In one embodiment, the method further includes:

[0037] The tactical semantic results, interpretable evidence subgraphs, and uncertainty estimation indices are encapsulated into a structured data format;

[0038] The encapsulated data is output through a standardized interface to generate tactical mining results.

[0039] In one embodiment, the standardized interface includes a data service layer, an inference service layer, and a visualization service layer, wherein:

[0040] The data service layer is used for interface calls, request routing, and result caching.

[0041] The inference service layer is used to perform real-time tactical inference and uncertainty estimation.

[0042] The visualization service layer is used to render the evidence subgraph into a visual graphic.

[0043] A sports event tactics mining system based on multi-granularity visual perception and temporal causal reasoning, the system comprising:

[0044] The data acquisition and preprocessing module is used to acquire event video streams from multiple perspectives and preprocess the event video streams to obtain standardized video frame data.

[0045] The multi-granularity visual perception module is used to extract features and identify objects from the standardized video frame data by utilizing group isovariant convolutional networks, capsule networks and slot attention mechanisms, to obtain multi-granularity perception data that fuses geometric isovariant features, hierarchical pose features and semantic instantiation features.

[0046] The spatiotemporal graph construction module is used to align the identities of objects between adjacent frames using the optimal transmission matching method on the multi-granularity sensing data, and to perform continuous spatiotemporal modeling by combining neural differential equations to construct a dynamic spatiotemporal graph containing nodes, edges and weighted adjacency matrices.

[0047] The tactical semantic reasoning and mining module is used to perform tactical temporal modeling using neural point processes, tactical rule constraints using differentiable logic networks, and mining causal relationships between tactical behaviors using causal graph discovery algorithms to obtain tactical semantic results from the dynamic spatiotemporal graph.

[0048] The application service interface module is used to generate an interpretable evidence subgraph based on the tactical semantic results, perform uncertainty estimation using deep set inference, encapsulate it into a standardized interface, and output the tactical mining results.

[0049] The aforementioned method and system for mining sports tactics based on multi-granularity visual perception and temporal causal reasoning acquires standardized video frame data through multi-view video acquisition and preprocessing. It combines group-variable convolutional networks, capsule networks, and slot attention mechanisms to achieve multi-granularity visual perception, improving the robustness and accuracy of object recognition and feature extraction in complex sports scenarios. Optimal transmission matching is used to align object identities across frames, and neural differential equations are combined to construct a dynamic spatiotemporal graph, enabling continuous and accurate characterization of object spatiotemporal trajectories and interaction relationships. Through neural point processes, differentiable logic networks, and causal graph discovery algorithms, deep reasoning is performed on the dynamic spatiotemporal graph, improving the accuracy, logic, and causal interpretability of tactical pattern recognition. Finally, interpretable evidence subgraphs are generated and uncertainty is estimated. Standardized interface output results are encapsulated, ensuring output reliability while enabling visualized traceability and cross-platform real-time access to the tactical analysis process. This effectively extracts tactical information from sports videos and performs high-level reasoning and analysis, providing intelligent support for sports commentary, audience experience, and coach tactical review. Attached Figure Description

[0050] Figure 1 This is an application environment diagram of a competition tactic mining method based on multi-granularity visual perception and temporal causal reasoning in one embodiment;

[0051] Figure 2 This is a flowchart illustrating a competition tactics mining method based on multi-granularity visual perception and temporal causal reasoning in one embodiment.

[0052] Figure 3This is a flowchart illustrating the tactical semantic reasoning and mining process in one embodiment;

[0053] Figure 4 This is a structural block diagram of a competition tactics mining system based on multi-granularity visual perception and temporal causal reasoning in one embodiment;

[0054] Figure 5 This is a schematic diagram of the architecture of a competition tactics mining system based on multi-granularity visual perception and temporal causal reasoning in one embodiment;

[0055] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] The event tactic mining method based on multi-granularity visual perception and temporal causal reasoning provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the application environment includes computer device 110. Computer device 110 can acquire multi-view captured video streams of the event and preprocess them to obtain standardized video frame data. Computer device 110 can utilize group equivariant convolutional networks, capsule networks, and slot attention mechanisms to extract features and identify objects in the standardized video frame data, obtaining multi-granularity perceptual data fused from geometric equivariant features, hierarchical pose features, and semantic instantiation features. For multi-granularity perceptual data, computer device 110 can use optimal transmission matching methods to align object identities between adjacent frames, combine neural differential equations for continuous spatiotemporal modeling, and construct a dynamic spatiotemporal graph containing nodes, edges, and a weighted adjacency matrix. For the dynamic spatiotemporal graph, computer device 110 can use neural point processes for tactical temporal modeling, use differentiable logic networks for tactical rule constraints, and use causal graph discovery algorithms to mine causal relationships between tactical behaviors, obtaining tactical semantic results. Computer device 110 can generate interpretable evidence subgraphs based on the tactical semantic results, use deep ensemble inference for uncertainty estimation, encapsulate it into a standardized interface, and output the tactical mining results. Among them, computer equipment 110 may include, but is not limited to, various personal computers, laptops, smartphones, robots, unmanned aerial vehicles, tablets, and other devices.

[0058] In one embodiment, such as Figure 2 As shown, a method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning is provided, including the following steps:

[0059] Step 202: Acquire the event video streams captured from multiple perspectives and preprocess the event video streams to obtain standardized video frame data.

[0060] Computer equipment can acquire raw, compressed video streams of the event captured in real time by multiple cameras, and after preprocessing, obtain a standard-sized, high-quality sequence of image frames that can be used for subsequent processing.

[0061] In one embodiment, the provided method for mining event tactics based on multi-granularity visual perception and temporal causal reasoning may further include a data preprocessing process, specifically including: acquiring multi-view event video streams through synchronous cameras deployed at various locations in the event venue, obtaining the original frame sequence after decoding and timestamp alignment; performing Gaussian blur denoising on the original frame sequence, and extracting key frames containing key actions and tactical events from the denoised frame sequence; adjusting the brightness and contrast of the key frames to obtain standardized video frame data.

[0062] Multiple high-definition cameras can be deployed at various locations within the competition venue, including cameras at the four corners, sidelines, and overhead views, to simultaneously capture the raw video stream of the event from different angles and perspectives. Examples include shots from the four corners of the venue or from an aerial viewpoint. Each high-definition camera can connect and communicate with computer equipment, allowing the computer to capture continuous, independent real-time video streams of the event.

[0063] After receiving multi-view video streams of a sporting event, computer equipment can decode, denoise, extract keyframes, and enhance images from these streams. By optimizing the data processing, it ensures that the video data meets high-quality standards so that subsequent modules can process it efficiently. Video streams are typically in compressed formats, and they first need to be decoded to convert them into raw image frames. The decoded video frames are then processed using denoising algorithms to remove interfering components from the image, such as compression noise and the effects of low light, thereby enhancing image quality.

[0064] Specifically, the computer equipment can decode the compressed video stream output by the camera in real time, restoring it to the original RGB image frame sequence, eliminating data loss caused by compression encoding, and restoring the original color, brightness, and detail information of the image, thus obtaining the unprocessed original video frame sequence. Since the startup time and network transmission latency of different cameras may vary, the computer equipment can first extract the display timestamps of each video stream, calculate the time offset of each signal, and use sliding window buffering or frame dropping strategies to align multiple video streams on the timeline, ensuring that images from different perspectives at the same moment correspond precisely.

[0065] After acquiring the video, computer equipment can use keyframe extraction technology to select the most representative frames from the video, reducing subsequent computation while retaining information about key actions and events. Simultaneously, image enhancement is performed, including adjustments to brightness and contrast, to ensure image clarity and readability, providing clear and standardized video frame data for subsequent analysis. The video sequence is defined as: V = {v1, v2, ..., vn}; where each frame vi is a video frame after decoding, denoising, and enhancement processing.

[0066] Step 204: Using group isovariant convolutional networks, capsule networks and slot attention mechanisms, feature extraction and object recognition are performed on standardized video frame data to obtain multi-granularity perceptual data that integrates geometric isovariant features, hierarchical pose features and semantic instantiation features.

[0067] Computer devices can achieve multi-granularity visual feature parsing from the geometric layer, structural layer to the semantic layer through the collaborative structure of group isovariant convolutional networks (E(2)-CNN / G-CNN), capsule networks (CapsNet), and slot attention mechanisms. This complementary approach, combining spatial invariance, hierarchical representation, and object instantiation, enables high-precision recognition of participating objects, action details, and scene semantics in complex competition environments.

[0068] In one embodiment, a method for mining event tactics based on multi-granularity visual perception and temporal causal reasoning may further include a multi-granularity visual perception process. This process specifically includes: inputting standardized video frame data into a group equilateral convolutional network; defining convolutional kernels on continuous groups for feature extraction to obtain geometrically equilateral features; reshaping the geometrically equilateral features into capsule vectors; and using the capsule network's dynamic routing algorithm to capture target hierarchical relationships and output hierarchical pose features; inputting the hierarchical pose features into a slot attention mechanism; iteratively aggregating features through learnable slot variables to achieve object instantiation and output semantic instantiation features; and weightedly fusing the geometrically equilateral features, hierarchical pose features, and semantic instantiation features to obtain multi-granularity perception data.

[0069] Computer devices can take pre-processed, standardized video frames with uniform size, color, and format as input, and process them through group-variable convolutional networks, capsule networks, and slot attention mechanisms to obtain multi-granularity perceptual data.

[0070] Specifically, computer equipment can input standardized video frame data into a group equivariant convolutional network (E(2)-CNN / G-CNN) for feature extraction. The group equivariant convolutional network defines convolution kernels on a continuous group G, i.e., a continuous group composed of rotation, translation, and reflection, and performs multi-scale convolution operations on the image. The output features ensure equivariance under changes in viewpoint, lens translation, and slight rotation, while simultaneously extracting low-level geometric features to form a geometrically equivariant feature map, achieving a geometrically robust low-level representation. The computation process can be defined as follows: Where x is the convolution kernel function, vi is the input frame, and the output is a geometrically equivariant feature map. The low-level representation effectively maintains the geometric consistency of structures such as field markings and area boundaries under different camera perspectives.

[0071] After geometrically isovariant feature extraction, the data is fed into a Capsule Network (CapsNet) for hierarchical structure parsing. The Capsule Network encodes the pose and existence of features in vector form and captures the hierarchical relationships between local components and the overall target through a dynamic routing mechanism. Specifically, the computer device can feed the geometric feature map output by group isovariant convolution into the Capsule Network. Lower-level capsules encode local region features, outputting vectors containing position, angle, and scale information; then, through a dynamic routing mechanism, lower-level capsules are aggregated into higher-level capsules to express hierarchical relationships. The computation process can be defined as follows: Where ui is the output of the lower-level capsule, Wij is the weight matrix, cij is the routing coefficient, and vj is the output of the higher-level capsule. Through the capsule structure, the system can simultaneously express the spatial hierarchy, attitude changes, and motion characteristics of the target, providing a structured representation for subsequent spatiotemporal modeling.

[0072] Subsequently, the output hierarchical pose features are fed into the slot attention mechanism to achieve object-level semantic aggregation. Slot attention uses a learnable slot variable zk to represent each independent object in the frame, and dynamically aggregates relevant features through an iterative attention update process. Mathematically, this can be expressed as: Here, αik represents the attention weight, xi represents the input feature, and zk represents the slot representation. Each slot corresponds to an independent object in the video frame, thus achieving object separation and semantic instantiation.

[0073] Through the aforementioned three-layer algorithm structure, computer devices can achieve visual feature parsing at different granular levels: low-level features possess geometric equivariance, mid-level features possess hierarchical structural expression capabilities, and high-level features possess object-level semantic separation capabilities. Computer devices can concatenate or weightedly fuse geometric equivariance features, hierarchical pose features, and semantic instantiation features from the same frame along the channel dimension to form a unified feature tensor. This feature tensor simultaneously contains low-level geometric robustness, mid-level structural pose, and high-level object instance information, ultimately outputting multi-granularity perceptual data for subsequent spatiotemporal map construction. That is, the final output is a comprehensive representation composed of geometric, structural, and semantic features: Mi={F eo, F aps, F lot}; where F eo represents geometric equivariance, F aps represents hierarchical pose features, F lot represents semantic instantiation features, which serve as subsequent inputs to provide high-precision, multi-layered visual information support for tactical semantic reasoning.

[0074] Step 206: For multi-granularity sensing data, the optimal transmission matching method is used to align the identities of objects between adjacent frames. Continuous spatiotemporal modeling is performed in conjunction with neural differential equations to construct a dynamic spatiotemporal graph containing nodes, edges, and a weighted adjacency matrix.

[0075] Computer equipment can fuse multi-granularity sensing data to establish a continuous and evolvable spatiotemporal semantic graph, thereby accurately depicting the spatiotemporal interaction relationships and dynamic evolution patterns among participants in the competition.

[0076] In one embodiment, a method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning may further include a spatiotemporal graph construction process. The specific process includes: extracting object sets from adjacent frames based on multi-granularity perception data, calculating the matching cost between objects in adjacent frames, and constructing a matching cost matrix; introducing regularization and using the Sinkhorn algorithm to iteratively solve the matching cost matrix to obtain the optimal transmission matrix, achieving object identity alignment between adjacent frames; determining the state of the aligned objects, and constructing a neural differential equation with object state as input and state change rate as output; creating graph nodes based on the aligned objects, establishing graph edges according to the spatial distance and semantic relationship between nodes, and dynamically calculating the weights representing the closeness of tactical cooperation through a graph attention mechanism to construct a weighted adjacency matrix; continuously updating node states using the neural differential equation, and synchronously dynamically adjusting the edge connection relationships and the weighted adjacency matrix to generate a dynamic spatiotemporal graph that evolves over time.

[0077] Computer devices can employ a joint modeling approach combining Optimal Transport (OT) and Neural Ordinary Differential Equation (Neural ODE) to perform object matching and correspondence calculation across time frames through the optimal transport mechanism. This ensures the consistency of identity and trajectory continuity of all participating objects (players, referees, balls, etc.) under continuous temporal conditions.

[0078] Specifically, the computer device can extract all detected objects from two adjacent frames of multi-granularity sensor data, forming a source domain object set and a target domain object set. For objects in these two sets, a matching cost is calculated. This matching cost comprehensively considers the object's geometric equivariance features, semantic instantiation features, and appearance features. It also considers feature similarity and spatial Euclidean distance; a smaller distance and higher feature similarity result in a lower cost, indicating a higher probability that the two objects belong to the same entity. Next, the computer device can organize the matching costs of all object pairs into a two-dimensional matching cost matrix.

[0079] To find the globally optimal matching scheme, this embodiment introduces the Optimal Transport (OT) theory. Specifically, the computer device can use the Sinkhorn algorithm with entropy regularization to iteratively solve the cost matrix. Let the object sets of two adjacent time frames be x = {xi}. 1 and y={yj} =1, defining its matching cost matrix. The regularized optimal transfer matrix obtained by the Sinkhorn approximation is: ;in, Let u(a, b) be the optimal matching matrix, representing the set of probabilities satisfying the row and column edge distribution constraints, and ε be the entropy regularization parameter. The matching process achieves consistent temporal alignment of objects, thus providing accurate node mapping relationships for the construction of the spatiotemporal graph. Specifically, after iterative convergence of the Sinkhorn algorithm, the optimal transfer matrix is ​​obtained, where each element represents the probability of a source domain object being assigned to a target domain object. Through the optimal transfer matrix, precise alignment of object identities between adjacent frames is achieved.

[0080] After determining the identity alignment, the state information of each object at the current moment, including spatial position, motion speed, posture features and interaction intensity, is used to construct a neural differential equation with the object state as input and the state change rate as output. The neural differential equation parameterizes the state evolution law through a learnable neural network to characterize the continuous motion and dynamic changes of the object over time.

[0081] After object matching is completed, the computer device can construct a time-continuous spatiotemporal graph based on the matching results. In the spatiotemporal graph Gt=(vt, Et), nodes vt represent participants and key events, and edges Et represent spatial or semantic relationships between objects. The interaction weights between nodes are obtained by weighting feature similarity with spatial distance. To characterize the dynamic features evolving over time, this embodiment introduces a Neural ODE to continuously model the state of the spatiotemporal graph. The evolution equation is defined as: Where zt represents the hidden state of the graph at time t; The neural function is learnable, and the parameter θ is optimized through inverse integration. By numerically solving the above equation over the time interval, continuous modeling of object motion trajectories, event triggering, and interaction patterns can be achieved, thereby obtaining a dynamic representation with high temporal resolution. That is, in this embodiment, the computer device can create spatiotemporal graph nodes based on objects that have completed identity alignment. The node attributes include object identity, spatial coordinates, motion parameters, posture features, and timestamps. The computer device can establish graph edges based on the spatial distance, feature similarity, and semantic association between nodes, and introduce a graph attention mechanism to dynamically calculate the association weights of each edge. The weight represents the tightness of tactical cooperation between players, thus constructing a complete weighted adjacency matrix.

[0082] Next, the computer equipment can continuously iterate and update the node states using neural differential equations, reflecting the continuous evolution of the object's position, motion, and attitude in real time. Simultaneously, it dynamically adjusts the connectivity of graph edges and the weight values ​​in the weighted adjacency matrix based on changes in node states, ultimately generating a dynamic spatiotemporal graph that fully characterizes the event process and continuously evolves over time. Specifically, in actual modeling, to improve the physical consistency of the spatiotemporal graph, structural constraints and energy conservation terms can be introduced simultaneously during node updates. The temporal update rule for nodes is defined as follows: Where xi(t) represents the node position or state vector, and Ni(τ) represents its neighborhood node set. Through continuous integration, long-term temporal dependencies between objects can be captured without relying on discrete time steps, achieving a natural description of the overall tactical evolution.

[0083] Ultimately, the spatiotemporal graph output by the module not only includes nodes and their dynamic attributes, but also interaction weights and association strengths that change over time, defined as: Gt=(vt, Et, At); where At represents the weighted adjacency matrix of the graph at time t, recording the association strength between objects.

[0084] Step 208: For the dynamic spatiotemporal graph, tactical temporal modeling is performed using neural point processes, tactical rule constraints are performed using differentiable logic networks, and causal graph discovery algorithms are used to mine causal relationships between tactical behaviors to obtain tactical semantic results.

[0085] Computer devices can combine Neural Hawkes Process (NHP), Differentiable Logic Network (DLN), and Causal Graph Discovery (CGD) algorithms to identify tactical trigger patterns, model rule-level semantic constraints, and analyze policy-level causal relationships, thereby forming a self-explanatory tactical reasoning system.

[0086] In one embodiment, a method for mining event tactics based on multi-granularity visual perception and temporal causal reasoning may further include a tactical semantic reasoning process. This process specifically includes: extracting tactical events from a dynamic spatiotemporal graph and encoding historical tactical event sequences using neural point processes; learning time-varying conditional strength functions to predict the instantaneous probability of future tactical events; constructing logical rules based on pre-defined tactical knowledge; embedding these rules into a differentiable logic network; imposing rule constraints on the tactical reasoning process and correcting the output logical patterns through the differentiable logic network; learning a structured causal graph representing the causal relationships between tactical behaviors by optimizing the objective function and imposing directed acyclic graph constraints based on the dynamic spatiotemporal graph; and fusing instantaneous probabilities, logical patterns, and structured causal graphs to generate tactical semantic results.

[0087] Computer devices can extract tactical events from dynamic spatiotemporal graphs, including key actions such as player positioning, passing, pressing, positional switching, shooting, and tackling, as well as information such as the time of occurrence, participating parties, and spatial location. After extracting the tactical events, the computer device can encode historical tactical event sequences using neural point processes. Specifically, the computer device can organize past tactical events into an event sequence according to chronological order, with each event containing a type label and the time of occurrence, forming a historical tactical event sequence. The computer device can characterize the temporal dependencies between events using a time decay kernel function, adaptively learn a conditional strength function that dynamically changes over time, and predict the instantaneous probability of future tactical events of various types, thereby achieving temporal modeling of tactical evolution trends.

[0088] Specifically, computer equipment can perform time series modeling of the triggering patterns of event nodes in a spatiotemporal graph. To characterize the dynamic dependencies of tactical events, a neural point process model is used to define the conditional intensity function of event category c at time t: Where μc is the basic strength term, αc, ci represent the activation weights between event types, and gβ(τ) is the time decay kernel function; the model captures the time delay dependencies between different types of tactical events (such as advancing, switching positions, pressing, etc.) through an adaptively parameterized recursive structure, thereby forming the temporal probability distribution of tactical triggers. The log-likelihood of the event trigger sequence is defined as: By maximizing LNHP, the model can learn the distribution of trigger intensity of tactical events, and achieve time-series modeling of the formation and transition patterns of tactics.

[0089] Building upon temporal modeling, to ensure logical consistency and rule interpretability in tactical reasoning, computer devices can incorporate Differentiable Logic Networks (DLNs) to impose rule constraints on tactical reasoning, ensuring that the reasoning results conform to known tactical patterns. Based on professional competition experience and a tactical knowledge base, computer devices can construct first-order logical rules that include triggering conditions, execution actions, and result feedback. These logical rules are then transformed into differentiable tensor operations and embedded into the differentiable logic network. Rule constraints are applied during model reasoning, and the network output is corrected through a logical loss function, filtering out reasoning results that do not conform to tactical common sense and outputting reasonable and consistent tactical logical patterns.

[0090] That is, during network training and inference, differentiable logic networks perform real-time verification and correction of the output of neural point processes. Specifically, the differentiable logic mechanism embeds traditional logic rules into the neural network structure in a differentiable form, realizing rule expression and constraint propagation through logic tensors. The rule loss function is defined as: Llogic wr.max(0,Bodyr(Gt)—Headr(Gt) Where r represents the logical rule number, Bodyr(Gt) and Headr(Gt) represent the continuity scores of the rule premise and conclusion, respectively, and wr is the rule confidence weight. Through logical regularization constraints, computer devices can maintain the consistency between tactical triggering conditions and tactical results in the neural space, enabling the model to follow known tactical rules during reasoning.

[0091] To reveal directional causal dependencies between tactical actions, computing devices can employ Causal Graph Discovery (CGD) algorithms to adaptively learn the directional dependencies between graph nodes. The device can construct a joint optimization objective function including prediction loss and a sparse regularization term. Through iterative optimization and the imposition of directed acyclic graph constraints, it automatically learns the directional dependencies between nodes, generating a structured causal graph that characterizes the direct causal links, causal strength, and confidence levels between tactical actions, thus revealing the deep driving relationships between them.

[0092] Specifically, the computer device can optimize the adjacency matrix A by sparsifying it and applying directed acyclic constraints, with the objective defined as: Where A represents the causal adjacency matrix of tactical behavior to be learned, which is used to characterize the causal weights between nodes; Lpred(A) is the prediction loss based on the adjacency structure, which is used to measure the degree of fit between the structural model and the observed data; λ is the L1 norm regularization term, used to induce sparsity in the adjacency matrix; λ is the sparsity regularization coefficient, used to adjust the balance between prediction accuracy and structural sparsity; constraint term. Ensure that the causal graph represented by A is a directed acyclic graph (DAG), where, `trace` is the trace operator for matrices, `e` is the matrix exponentiation operator, `∘` is the Hadamard product, and `d` is the total number of nodes in the causal graph. Through this mechanism, computer devices can automatically discover the directional influence relationships between tactical and reactive behaviors, enabling causal analysis and reasoning at the strategy level.

[0093] After completing the three-dimensional modeling, the computer device can fuse the instantaneous probabilities output by the neural point process, the logical rules output by the differentiable logic network, and the structured causal graph output by the causal graph discovery algorithm to generate tactical semantic results. The computer device can perform feature-level fusion of the instantaneous occurrence probabilities of tactical events output by the neural point process, the tactical logical rules output by the differentiable logic network, and the structured causal graph output by the causal graph discovery algorithm, integrating temporal evolution features, logical constraint features, and causal dependency features to form a unified high-dimensional semantic representation. Ultimately, this generates tactical semantic results that can accurately characterize tactical patterns, combat strategies, and formation changes.

[0094] Specifically, the computer device can integrate neural point processes, differentiable logic networks, and causal graph learning to form an end-to-end optimization system. The overall objective function is defined as: Ltotal = λ1LNHP + λ2Llogic + λ3LCGD; where λ1, λ2, and λ3 are multi-task balancing coefficients. Through joint training, the computer device can simultaneously learn the probabilistic mechanisms, logical rules, and causal structures of tactical formation at the temporal, rule-based, and causal levels, achieving self-explanatory and highly robust tactical semantic reasoning. The final output tactical semantic result is expressed as: ; where Φ eq represents the time-triggered features based on neural point processes, Φ ogic is a characteristic of symbolic reasoning under differentiable logic constraints, Φ ausal is the policy dependency structure obtained from causal graph learning.

[0095] In one embodiment, the process of obtaining tactical semantic results in a competition tactical mining method based on multi-granularity visual perception and temporal causal reasoning is as follows: Figure 3 As shown, the spatiotemporal graph is used as input, and temporal modeling is performed through neural point processes. The neural point processes extract event sequences from the spatiotemporal graph, learn the temporal dependencies between events, and output event temporal features. Next, rule constraints are applied through a differentiable logic network, which mainly receives the temporal features output by the neural point processes, transforms the preset professional tactical rules into differentiable mathematical operations, imposes constraints on the reasoning process, filters out predictions that do not conform to tactical common sense, corrects the output logic, and generates tactical rule reasoning results. Then, causal mining is performed through causal graph learning, which mainly receives the rule reasoning results of the differentiable logic network, automatically learns the causal relationships between tactical behaviors through algorithms, and finally outputs tactical strategy predictions. Finally, by integrating temporal features, rule constraints, and causal relationships, interpretable tactical patterns and adversarial strategies are output.

[0096] Step 210: Generate an interpretable evidence subgraph based on the tactical semantic results, use deep set inference to estimate uncertainty, encapsulate it into a standardized interface, and output the tactical mining results.

[0097] Computer devices can combine evidential subgraph and uncertainty estimation techniques to provide decision-making basis and confidence intervals while outputting tactical identification results, thereby enhancing interpretability and reliability.

[0098] In one embodiment, a method for mining event tactics based on multi-granularity visual perception and temporal causal reasoning may further include an uncertainty estimation process. This process includes: for logical patterns and structured causal graphs in the tactical semantic results, selecting key nodes and paths based on temporal correlations to generate interpretable evidence subgraphs; constructing independent sub-models with different initial weights, inputting the tactical semantic results into each independent sub-model, obtaining the tactical prediction mean and variance based on the output of each independent sub-model; and obtaining uncertainty estimation indices based on the tactical prediction mean and variance to complete the uncertainty estimation.

[0099] After obtaining the tactical semantic results, the computer device can first analyze the logical patterns and structured causal graphs within the tactical semantic results, and then filter out key nodes and paths based on temporal correlation to generate an interpretable evidence subgraph. For the logical patterns and structured causal graphs contained in the tactical semantic results, the computer device can use temporal correlation indicators to filter nodes and paths in the dynamic spatiotemporal graph, prioritizing the retention of key nodes and key interaction paths with high causal strength, strong logical consistency, and close temporal correlation, while eliminating irrelevant or weakly correlated redundant information, generating a concise interpretable evidence subgraph with clear causal explanation capabilities.

[0100] Next, the computer equipment can construct multiple independent sub-models with the same network structure but different initial weights, forming a model ensemble system. The same tactical semantic result is input in parallel into each independent sub-model, and each sub-model outputs an independent tactical prediction result. The outputs of all sub-models are combined to calculate the tactical prediction mean and variance. Then, based on the obtained tactical prediction mean and variance, the confidence level of the model in the current scenario is quantified, and an uncertainty estimation index is constructed. A large variance is marked as high uncertainty, and a small variance is marked as low uncertainty, thus completing the uncertainty estimation.

[0101] Specifically, the computer device can receive composite outputs from the tactical semantic reasoning and mining module. Furthermore, the most representative explanatory subgraphs are generated using an evidence graph analysis algorithm. The generation process of the evidence subgraphs aims to maximize tactical relevance and causal consistency, and is defined as follows: Where wij is the edge weight, which is obtained by weighting causality strength, logical consistency, and temporal relevance. This represents the subgraph structure that best explains tactical behavior at time t. Through this subgraph, computer devices can display the evidence path for tactical formation at the interface layer, enabling a visual explanation of tactical decisions.

[0102] When outputting tactical results, the computer equipment employs an uncertainty estimation method based on deep ensemble inference to quantify the confidence intervals of the model predictions. Let the model set... For k independent sub-models, the mean and variance of the tactical prediction distribution are defined as follows: ;in, t represents the final tactical prediction value, σ This indicates the quantification of uncertainty in the output; enabling result labeling or human-machine collaborative review when the confidence level of tactical judgment is low, thereby improving the reliability of decision-making.

[0103] In one embodiment, a competition tactical mining method based on multi-granularity visual perception and temporal causal reasoning may further include a process of generating the final tactical mining results. The specific process includes: encapsulating tactical semantic results, interpretable evidence subgraphs, and uncertainty estimation indicators into a structured data format; and outputting the encapsulated data through a standardized interface to generate tactical mining results.

[0104] After completing tactical semantic reasoning, interpretable evidence subgraph generation, and uncertainty estimation, the computer device can encapsulate these three core data types into a unified structured data format. That is, the computer device can structure and standardize the multi-dimensional tactical information output from tactical semantic reasoning and data mining, encapsulating it into a standardized application interface to achieve real-time integration with external systems. The interface output includes not only tactical type and strategy prediction results, but also corresponding evidence subgraphs, confidence intervals, and logical consistency indicators. The output structure is defined as follows: in, Indicates tactical prediction label, Indicates the prediction confidence interval. For interpretable evidence subgraphs, Logical consistency is scored. The interface adopts a unified API communication protocol, supports dual-channel interaction of REST and WebSocket, and ensures that tactical information can be transmitted in real time to the commentary system, audience interaction terminal, and coach decision-making platform.

[0105] By combining interpretable evidence subgraphs with uncertainty estimation mechanisms, standardized outputs are achieved while providing verifiable evidence for tactical identification and risk self-assessment capabilities, significantly improving interpretability, stability, and engineering deployability.

[0106] In one embodiment, the standardized interface includes a data service layer, an inference service layer, and a visualization service layer, wherein: the data service layer is used for interface calls, request routing, and result caching; the inference service layer is used for performing real-time tactical inference and uncertainty estimation; and the visualization service layer is used for rendering evidence subgraphs into visual graphics.

[0107] In this embodiment, the computer device can support a multi-layer service encapsulation mechanism, including a data service layer, an inference service layer, and a visualization service layer. The data service layer is responsible for interface calls and result caching; the inference service layer provides real-time inference results and confidence information; and the visualization service layer presents tactical evidence and logical paths in a graphical form, providing users with a traceable and interpretable tactical analysis experience.

[0108] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0109] In one embodiment, such as Figure 4 As shown, a sports event tactics mining system based on multi-granularity visual perception and temporal causal reasoning is provided, including: a data acquisition and preprocessing module 410, a multi-granularity visual perception module 420, a spatiotemporal graph construction module 430, a tactical semantic reasoning and mining module 440, and an application service interface module 450, wherein:

[0110] The data acquisition and preprocessing module 410 is used to acquire the event video stream from multiple perspectives and preprocess the event video stream to obtain standardized video frame data.

[0111] The multi-granularity visual perception module 420 is used to extract features and recognize objects from standardized video frame data by utilizing group isovariant convolutional networks, capsule networks and slot attention mechanisms, to obtain multi-granularity perception data that fuses geometric isovariant features, hierarchical pose features and semantic instantiation features.

[0112] The spatiotemporal graph construction module 430 is used to align the identities of objects between adjacent frames by using the optimal transmission matching method for multi-granularity sensing data, and to perform continuous spatiotemporal modeling by combining neural differential equations to construct a dynamic spatiotemporal graph containing nodes, edges and weighted adjacency matrices.

[0113] The tactical semantic reasoning and mining module 440 is used to perform tactical temporal modeling using neural point processes, tactical rule constraints using differentiable logic networks, and mining causal relationships between tactical behaviors using causal graph discovery algorithms to obtain tactical semantic results from dynamic spatiotemporal graphs.

[0114] Application service interface module 450 is used to generate interpretable evidence subgraphs based on tactical semantic results, perform uncertainty estimation using deep set inference, and encapsulate them into a standardized interface to output tactical mining results.

[0115] In one embodiment, the data acquisition and preprocessing module 410 is further configured to acquire multi-view video streams of the event through synchronous cameras deployed at various locations in the event venue, obtain the original frame sequence after decoding and timestamp alignment; perform Gaussian blur denoising on the original frame sequence, and extract key frames containing key actions and tactical events from the denoised frame sequence; adjust the brightness and contrast of the key frames to obtain standardized video frame data.

[0116] In one embodiment, the multi-granularity visual perception module 420 is further configured to input standardized video frame data into a group equilateral convolutional network, define convolution kernels on the continuous group for feature extraction, and obtain geometric equilateral features; reshape the geometric equilateral features into capsule vectors, and use the dynamic routing algorithm of the capsule network to capture the target hierarchical relationship and output hierarchical pose features; input the hierarchical pose features into a slot attention mechanism, iteratively aggregate features through learnable slot variables to realize object instantiation and output semantic instantiation features; and perform weighted fusion of geometric equilateral features, hierarchical pose features and semantic instantiation features to obtain multi-granularity perception data.

[0117] In one embodiment, the spatiotemporal graph construction module 430 is further configured to extract the object set in adjacent frames based on multi-granularity sensing data, calculate the matching cost between each object in adjacent frames, and construct a matching cost matrix; introduce regularization and use the Sinkhorn algorithm to iteratively solve the matching cost matrix to obtain the optimal transmission matrix, thereby achieving object identity alignment between adjacent frames; determine the state of the aligned objects, and construct a neural differential equation with the object state as input and the state change rate as output; create graph nodes based on the aligned objects, establish graph edges according to the spatial distance and semantic relationship between nodes, and dynamically calculate the weights representing the closeness of tactical cooperation through a graph attention mechanism to construct a weighted adjacency matrix; continuously update the node state using the neural differential equation, and synchronously dynamically adjust the connection relationship of the edges and the weighted adjacency matrix to generate a dynamic spatiotemporal graph that evolves over time.

[0118] In one embodiment, the tactical semantic reasoning and mining module 440 is further configured to extract tactical events from a dynamic spatiotemporal graph, encode historical tactical event sequences using neural point processes, learn time-varying conditional strength functions to predict the instantaneous probability of future tactical events, construct logical rules based on preset tactical knowledge, embed the logical rules into a differentiable logic network, impose rule constraints on the tactical reasoning process through the differentiable logic network, and correct the output logical patterns; based on the dynamic spatiotemporal graph, employ a causal graph discovery algorithm to learn a structured causal graph representing the causal relationships between tactical behaviors by optimizing the objective function and imposing directed acyclic graph constraints; and fuse instantaneous probabilities, logical patterns, and structured causal graphs to generate tactical semantic results.

[0119] In one embodiment, the application service interface module 450 is further used to filter out key nodes and paths from the logical patterns and structured causal graphs in the tactical semantic results, combined with temporal correlation, and generate an interpretable evidence subgraph; construct independent sub-models with different initial weights, input the tactical semantic results into each independent sub-model, obtain the tactical prediction mean and variance based on the output of each independent sub-model; and obtain uncertainty estimation index based on the tactical prediction mean and variance to complete the uncertainty estimation.

[0120] In one embodiment, the application service interface module 450 is further configured to encapsulate tactical semantic results, interpretable evidence subgraphs, and uncertainty estimation indicators into a structured data format; and output the encapsulated data through a standardized interface to generate tactical mining results.

[0121] In one embodiment, the standardized interface includes a data service layer, an inference service layer, and a visualization service layer, wherein: the data service layer is used for interface calls, request routing, and result caching; the inference service layer is used for performing real-time tactical inference and uncertainty estimation; and the visualization service layer is used for rendering evidence subgraphs into visual graphics.

[0122] In one embodiment, such as Figure 5 As shown, a sports event tactical mining system based on multi-granularity visual perception and temporal causal reasoning is provided, including: a data acquisition and preprocessing module, a multi-granularity visual perception module, a spatiotemporal graph construction module, a tactical semantic reasoning and mining module, and an application service interface module, demonstrating the process from raw video streams to final tactical prediction results.

[0123] Specifically, such as Figure 5As shown, the original match video stream is input to the data acquisition and preprocessing module, which performs the following operations: decodes the video stream to restore the original image frames; uses Gaussian blur to denoise and eliminate image noise and compression artifacts; extracts keyframes, filters redundant images, and retains core action frames relevant to tactics; enhances the image by unifying brightness and contrast to improve image quality; and finally outputs a standardized video stream to provide input data with a unified format and clear image quality for subsequent processing. The standardized video stream is then input to the multi-granularity visual perception module, which performs the following operations: uses group isovariant convolution to extract viewpoint-robust geometric features to adapt to camera rotation and translation changes; uses capsule networks to capture target-level pose features to accurately identify player limb movements and states; uses slot attention to achieve instantiation and separation of objects such as players, balls, and referees, outputting independent object features; and finally outputs multi-granularity perception data that integrates geometric isovariant features, target recognition results, and action recognition information. Inputting multi-granularity perception data into the spatiotemporal graph construction module, this module performs the following operations: using optimal transmission matching to align object identities between adjacent frames, resolving player trajectory drift and identity confusion issues; using neural differential equations to model the continuous spatiotemporal evolution of object states, constructing a dynamic spatiotemporal graph containing nodes, edges, and a weighted adjacency matrix; and finally outputting the dynamic spatiotemporal graph, which fully records the position, trajectory, and interaction relationships of objects during the match. Inputting node and edge relationship data from the dynamic spatiotemporal graph into the tactical semantic reasoning and mining module, this module performs the following operations: modeling the temporal patterns of tactical events through neural point processes to predict the instantaneous probability of events; embedding tactical rules through differentiable logic to impose logical constraints on the reasoning process, ensuring the results conform to professional understanding; mining causal relationships between tactical behaviors through causal graph learning, constructing a structured causal graph; and finally outputting tactical pattern recognition results and tactical strategy prediction information. Input the tactical pattern and prediction results to the application service interface module; the application service interface module performs the following operations: generate an interpretable evidence subgraph to visually display the core basis of the tactical judgment; estimate uncertainty to quantify the confidence level of the prediction results; and finally output the final tactical prediction results, which can be directly used in scenarios such as event commentary and coach review.

[0124] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0125] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0126] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a competition tactic mining method based on multi-granularity visual perception and temporal causal reasoning.

[0127] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a competition tactics mining method based on multi-granularity visual perception and temporal causal reasoning.

[0128] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0129] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning, characterized in that, The method includes: Acquire event video streams captured from multiple perspectives, and preprocess the event video streams to obtain standardized video frame data; By utilizing group equivariant convolutional networks, capsule networks, and slot attention mechanisms, feature extraction and object recognition are performed on the standardized video frame data to obtain multi-granularity perceptual data that fuses geometric equivariant features, hierarchical pose features, and semantic instantiation features. For the multi-granularity sensing data, the optimal transmission matching method is used to align the identities of objects between adjacent frames. Continuous spatiotemporal modeling is performed in combination with neural differential equations to construct a dynamic spatiotemporal graph containing nodes, edges, and a weighted adjacency matrix. For the dynamic spatiotemporal graph, tactical temporal modeling is performed using neural point processes, tactical rule constraints are performed using differentiable logic networks, and causal graph discovery algorithms are used to mine causal relationships between tactical behaviors to obtain tactical semantic results. Based on the tactical semantic results, an interpretable evidence subgraph is generated, and uncertainty estimation is performed using deep set inference. The results are then encapsulated into a standardized interface and the tactical mining results are output.

2. The method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning according to claim 1, characterized in that, Acquire multi-view captured video streams of the event, and preprocess the video streams to obtain standardized video frame data, including: Multi-view video streams of the event are captured by synchronous cameras deployed in various locations of the event venue, and the original frame sequence is obtained after decoding and timestamp alignment. The original frame sequence is denoised by Gaussian blur, and key frames containing key actions and tactical events are extracted from the denoised frame sequence. The brightness and contrast of the keyframes are adjusted to obtain standardized video frame data.

3. The method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning according to claim 1, characterized in that, Using group-variable convolutional networks, capsule networks, and slot attention mechanisms, feature extraction and object recognition are performed on the standardized video frame data to obtain multi-granularity perceptual data fused from geometrically variable features, hierarchical pose features, and semantic instantiation features, including: The standardized video frame data is input into a group equivariant convolutional network, and convolutional kernels are defined on continuous groups to extract features, thereby obtaining geometrically equivariant features. The geometric equilateral features are reshaped into capsule vectors, and the dynamic routing algorithm of the capsule network is used to capture the target hierarchical relationship and output the hierarchical pose features. The hierarchical pose features are input into the slot attention mechanism, and the features are iteratively aggregated through learnable slot variables to realize object instantiation and output semantic instantiation features. The geometric isomorphic features, hierarchical pose features, and semantic instantiation features are weighted and fused to obtain multi-granularity perception data.

4. The method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning according to claim 1, characterized in that, For the aforementioned multi-granularity sensing data, an optimal transmission matching method is used to align object identities between adjacent frames. Continuous spatiotemporal modeling is then performed using neural differential equations to construct a dynamic spatiotemporal graph containing nodes, edges, and a weighted adjacency matrix, including: Based on the multi-granularity sensing data, the object set in adjacent frames is extracted, and the matching cost between each object in adjacent frames is calculated to construct a matching cost matrix. Regularization is introduced and the Sinkhorn algorithm is used to iteratively solve the matching cost matrix to obtain the optimal transmission matrix, thereby achieving object identity alignment between adjacent frames. Determine the object state after identity alignment, and construct a neural differential equation with the object state as input and the state change rate as output; Graph nodes are created based on the aligned objects, graph edges are established according to the spatial distance and semantic relationship between nodes, and weights representing the closeness of tactical cooperation are dynamically calculated through graph attention mechanism to construct a weighted adjacency matrix; The neural differential equation is used to continuously update the node state and synchronously adjust the edge connection relationship and weighted adjacency matrix to generate a dynamic spatiotemporal graph that evolves over time.

5. The method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning according to claim 1, characterized in that, For the aforementioned dynamic spatiotemporal graph, tactical temporal modeling is performed using neural point processes, tactical rule constraints are implemented using differentiable logic networks, and causal graph discovery algorithms are used to mine causal relationships between tactical behaviors, yielding tactical semantic results, including: Tactical events are extracted from the dynamic spatiotemporal graph, and the sequence of historical tactical events is encoded using neural point processes. The conditional strength function that changes over time is learned to predict the instantaneous probability of future tactical events. Logical rules are constructed based on preset tactical knowledge, and these logical rules are embedded in a differentiable logic network. The differentiable logic network is used to impose rule constraints on the tactical reasoning process and correct the output logical patterns. Based on the dynamic spatiotemporal graph, a causal graph discovery algorithm is used to learn a structured causal graph that represents the causal relationship between tactical behaviors by optimizing the objective function and imposing a directed acyclic graph constraint. The instantaneous probability, logical rules, and structured causal graph are integrated to generate tactical semantic results.

6. The method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning according to claim 5, characterized in that, Based on the aforementioned tactical semantic results, an interpretable evidence subgraph is generated, and uncertainty estimation is performed using deep ensemble inference, including: For the logical patterns and structured causal graphs in the tactical semantic results, key nodes and paths are selected by combining temporal correlation, and interpretable evidence subgraphs are generated. Construct independent sub-models with different initial weights, input the tactical semantic results into each independent sub-model, and obtain the tactical prediction mean and variance based on the output of each independent sub-model; Based on the mean and variance of the tactical predictions, an uncertainty estimation index is obtained, thus completing the uncertainty estimation.

7. The method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning according to claim 6, characterized in that, The method further includes: The tactical semantic results, interpretable evidence subgraphs, and uncertainty estimation indices are encapsulated into a structured data format; The encapsulated data is output through a standardized interface to generate tactical mining results.

8. The method for mining competition tactics based on multi-granularity visual perception and temporal causal reasoning according to claim 1, characterized in that, The standardized interface includes a data service layer, an inference service layer, and a visualization service layer, wherein: The data service layer is used for interface calls, request routing, and result caching. The inference service layer is used to perform real-time tactical inference and uncertainty estimation. The visualization service layer is used to render the evidence subgraph into a visual graphic.

9. A sports event tactics mining system based on multi-granularity visual perception and temporal causal reasoning, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire event video streams from multiple perspectives and preprocess the event video streams to obtain standardized video frame data. The multi-granularity visual perception module is used to extract features and identify objects from the standardized video frame data by utilizing group isovariant convolutional networks, capsule networks and slot attention mechanisms, to obtain multi-granularity perception data that fuses geometric isovariant features, hierarchical pose features and semantic instantiation features. The spatiotemporal graph construction module is used to align the identities of objects between adjacent frames using the optimal transmission matching method on the multi-granularity sensing data, and to perform continuous spatiotemporal modeling by combining neural differential equations to construct a dynamic spatiotemporal graph containing nodes, edges and weighted adjacency matrices. The tactical semantic reasoning and mining module is used to perform tactical temporal modeling using neural point processes, tactical rule constraints using differentiable logic networks, and mining causal relationships between tactical behaviors using causal graph discovery algorithms to obtain tactical semantic results from the dynamic spatiotemporal graph. The application service interface module is used to generate an interpretable evidence subgraph based on the tactical semantic results, perform uncertainty estimation using deep set inference, encapsulate it into a standardized interface, and output the tactical mining results.

Images