Game event prediction method, device, and equipment and storage medium

Abstract

The application relates to the technical field of video processing, and discloses a game event prediction method, device and equipment and a storage medium. The game event prediction method comprises the following steps: receiving a game video to be predicted, wherein the game video comprises a game live video or a game video recording; video frame sampling is performed on each time point in the game video respectively to obtain a time-sequenced video frame sequence corresponding to each time point; a structured text prompt in a predefined format is constructed, and the video frame sequence corresponding to each time point and the structured text prompt are taken as model input data respectively; the model input data is input into a pre-trained game event prediction model for processing, and the output comprises the burst probability of the type and position of various game events in the future. The application improves the understanding of the model for the key content and time-sequenced content of the video, and realizes the forward-looking prediction capability of the video content.

Description

Technical Field

[0001] This invention relates to the field of video processing technology, and in particular to a method, apparatus, device, and storage medium for predicting game events. Background Technology

[0002] With the rise of online games, their competitiveness and strategic elements have increasingly attracted the attention of players and the esports industry. In games, key events such as team fights, ganking, and tower pushing are core factors that directly influence the course of the battle and determine victory or defeat. To facilitate the prediction of these key in-game events and provide business decision-making support for applications such as AI tactical coaches, intelligent commentators, and live streaming platforms, real-time battle analysis of game videos using artificial intelligence technology has become a hot research topic in the industry.

[0003] Currently, existing multimodal large language models are mainly based on general natural image and text data for description and question answering tasks. When multimodal large language models are directly applied to online game video analysis, it is difficult to effectively capture the deep dynamic temporal dependencies and tactical intentions between consecutive frames, which makes it unable to meet the needs of in-depth understanding and accurate prediction of game content in game battle analysis. Summary of the Invention

[0004] The main objective of this invention is to provide a game event prediction method, apparatus, device, and storage medium, aiming to solve the technical problem that existing multimodal large models are unable to effectively capture the deep dynamic temporal dependencies and tactical intentions between consecutive frames when analyzing online game videos.

[0005] The first aspect of this invention provides a game event prediction method, the game event prediction method comprising: Receive game videos to be predicted, including live game videos or game recordings; Video frames are sampled at each time point in the game video to obtain a time-series video frame sequence corresponding to each time point. Construct a predefined formatted structured text prompt, and use the video frame sequence corresponding to each time point and the structured text prompt together as model input data; The input data of the model is fed into a pre-trained game event prediction model for processing, and the output includes the probability of the occurrence of various types and locations of future game events.

[0006] Optionally, in a first implementation of the first aspect of the present invention, the step of sampling video frames at each time point in the game video to obtain a time-series video frame sequence corresponding to each time point includes: Using each time point in the game video as a base point, extract a video segment of a preset duration before each time point; Based on a preset video frame sampling strategy, video frames are sampled for each video segment to obtain a time-series video frame sequence with the same number of frames corresponding to each time point.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the structured text prompt includes task instructions, task definitions, task analysis guidelines, and model output formats; the task analysis guidelines are used to guide model analysis, and the model output formats are used to indicate the probability of game event outbreaks with event type and spatial location orientation in the model output.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the training method used for the game event prediction model includes: Acquire game recordings of multi-player battles; Video frames are sampled at each time point in the game recording to obtain a time-series video frame sequence corresponding to each time point; Construct a predefined format of structured text prompts, and use the video frame sequence corresponding to each time point and the structured text prompts together as training samples; Locate the game events that occur in the game recording, and construct sample labels corresponding to each time point based on the difference between the occurrence time of each game event and each time point; Based on the training samples and sample labels corresponding to each time point, the multimodal large model is trained to obtain a trained game event prediction model.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, the step of locating game events occurring in the game recording and constructing sample labels corresponding to each time point based on the difference between the occurrence time of each game event and each time point includes: Obtain the game battle logs corresponding to the game recordings; Based on the game battle log, the time and location of all game events are located in the game recording; Based on the occurrence time of each game event, an event prediction time window is set for each game event before the occurrence time. Using the difference between the occurrence time and the time point within the event prediction time window as the independent variable, a time decay probability function is constructed to represent the probability of a game event occurring. For each time point within the event prediction time window, positive sample labels corresponding to multiple event types and location orientations are generated based on the time decay probability function; For each time point outside the event prediction time window, generate a corresponding negative sample label with multiple event types and location pointers.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, the step of training the multimodal large model based on the training samples and sample labels corresponding to each time point to obtain a trained game event prediction model includes: The training samples and sample labels corresponding to each time point are sequentially input into the pre-trained multimodal large model; The game event is predicted using the multimodal large model, and the output includes the probability of the occurrence of various types and locations of the game events in the future. Based on the sample label and the outbreak probability corresponding to each time point, a target loss function is constructed, which includes a language modeling loss function and a predicted value regression loss function. Based on the target loss function, the model parameters of the multimodal large model are fine-tuned to obtain a trained game event prediction model.

[0011] A second aspect of the present invention provides a game event prediction device, the game event prediction device comprising: A receiving module is used to receive game videos to be predicted, including live game videos or game recordings. The second sampling module is used to sample video frames at each time point in the game video to obtain a time-series video frame sequence corresponding to each time point. The input construction module is used to construct a predefined format of structured text prompts, and uses the video frame sequence corresponding to each time point and the structured text prompts together as model input data; The prediction module is used to process the input data of the model into a pre-trained game event prediction model and output the probability of the occurrence of various types and locations of future game events.

[0012] Optionally, in a first implementation of the second aspect of the present invention, the second sampling module is specifically used for: Using each time point in the game video as a base point, extract a video segment of a preset duration before each time point; Based on a preset video frame sampling strategy, video frames are sampled for each video segment to obtain a time-series video frame sequence with the same number of frames corresponding to each time point.

[0013] Optionally, in a second implementation of the second aspect of the present invention, the structured text prompt includes task instructions, task definitions, task analysis guidelines, and model output formats; the task analysis guidelines are used to guide model analysis, and the model output formats are used to indicate the probability of game event outbreaks with event type and spatial location orientation in the model output.

[0014] Optionally, in a third implementation of the second aspect of the present invention, the game event prediction device further includes: The training module is used to acquire game recordings of multi-player game battles; sample video frames at each time point in the game recordings to obtain a time-series video frame sequence corresponding to each time point; construct a predefined formatted structured text prompt, and use the video frame sequence corresponding to each time point and the structured text prompt together as training samples; locate game events occurring in the game recordings, and construct sample labels corresponding to each time point based on the difference between the occurrence time of each game event and each time point; train a multimodal large model based on the training samples and sample labels corresponding to each time point to obtain a trained game event prediction model.

[0015] Optionally, in a fourth implementation of the second aspect of the present invention, the training module is specifically used for: Obtain the game battle logs corresponding to the game recordings; Based on the game battle log, the time and location of all game events are located in the game recording; Based on the occurrence time of each game event, an event prediction time window is set for each game event before the occurrence time. Using the difference between the occurrence time and the time point within the event prediction time window as the independent variable, a time decay probability function is constructed to represent the probability of a game event occurring. For each time point within the event prediction time window, positive sample labels corresponding to multiple event types and location orientations are generated based on the time decay probability function; For each time point outside the event prediction time window, generate a corresponding negative sample label with multiple event types and location pointers.

[0016] Optionally, in a fifth implementation of the second aspect of the present invention, the training module is further configured to: The training samples and sample labels corresponding to each time point are sequentially input into the pre-trained multimodal large model; The game event is predicted using the multimodal large model, and the output includes the probability of the occurrence of various types and locations of the game events in the future. Based on the sample label and the outbreak probability corresponding to each time point, a target loss function is constructed, which includes a language modeling loss function and a predicted value regression loss function. Based on the target loss function, the model parameters of the multimodal large model are fine-tuned to obtain a trained game event prediction model.

[0017] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the computer device to execute the above-described game event prediction method.

[0018] A sixth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described game event prediction method.

[0019] The game event prediction method provided by this invention is based on the core idea of ​​using an innovative "input-output-training" closed-loop design to "adapt" and "domesticate" existing multimodal large language models, enabling them to have professional game timing understanding and prediction capabilities.

[0020] (1) The input of the multimodal large language model is modified by using discretized temporal information as the model input. Specifically, video frames are sampled at each time point in the game recording to obtain the temporal video frame sequence corresponding to each time point. This method can transform a continuous video stream (such as the past 20 seconds) into an ordered static image sequence (such as 10 keyframes). This processing method is not simply video compression, but cleverly transforms the continuous temporal dependency problem into a problem of comparing and associating "ordered snapshot sequences". This is in line with the current mainstream multimodal large language model architecture for processing multiple images, forcing the model to no longer look at each frame in isolation, but to analyze and compare the changes between different video frames (such as changes in hero position and skill status), thereby capturing the dynamic tactical intentions that need to be understood across frames.

[0021] (2) By constructing predefined structured text prompts, the model is guided to perform implicit thought chain reasoning. For example, when the model is instructed to pay attention to "the movement trajectory of the hero icon on the minimap" and "the change of the ultimate skill cooldown indicator", it will stimulate the large model to think: "Because hero A's trajectory is from the jungle to the lane, and hero B's ultimate skill has just finished cooling down, a gank event is likely to occur next." This guidance breaks down a vague tactical task into a series of specific and observable temporal changes, thereby improving the large model's ability to understand the dynamic relationships and tactical intentions between consecutive frames. In addition, the structured text prompts clearly list the information that needs to be focused on, so that when the large model understands the image, the strong instructional text prompts force the large model's computing resources and attention to those UI elements that are small but have extremely high information density. This greatly makes up for the large model's shortcomings in professional fields, enabling it to accurately interpret structured information that has a decisive impact on the battle.

[0022] (3) Structured text prompts not only guide the model input but also strictly define the model output format: a JSON object containing probability values. This design transforms an open-ended task of predicting the future into a well-defined, structured output format with explicit requirements. The model is no longer allowed to generate vague, unreliable natural language descriptions but must output quantified probability data that can be directly parsed by the program. This fundamentally eliminates the generation of illusionary responses from large models and provides a reliable data interface for downstream automated systems.

[0023] (4) Fine-tune the multimodal large model using video frame sequences, structured prompts, and sample labels. During closed-loop training, the model establishes a strong correlation between specific visual temporal patterns and specific structured JSON outputs. Through learning from massive amounts of data, the model learns the reasoning ability from "observing the past" to "predicting the future," ultimately achieving the ability to predict key events in various games. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of an embodiment of the multimodal large model training method in this invention; Figure 2 This is a schematic diagram of one embodiment of the game event prediction method in this invention; Figure 3 This is a schematic diagram of one embodiment of the game event prediction device in this invention; Figure 4 This is a schematic diagram of one embodiment of the computer device in this invention. Detailed Implementation

[0025] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the multimodal large model training method in this invention includes: 101. Obtain game recordings of multi-player battles; In this embodiment, multi-player game battles refer to multiple players controlling their own game characters (such as heroes, soldiers, etc.) to battle against game characters controlled by opposing players. Game recordings refer to recordings of the entire game process from the start to the end.

[0027] In this embodiment, the game recording includes not only a video stream but also a live chat text stream. The video stream contains all the information from the game screen, such as hero positions, skill effects, minimap status, hero health and mana bars, and economic and equipment information. These are the most direct and objective basis for judging the battle situation. The live chat text stream provides real-time feedback from different viewers watching the game at the same moment.

[0028] 102. Sample video frames at each time point in the game recording to obtain a time-series video frame sequence corresponding to each time point; In this embodiment, to transform a continuous, high-dimensional video stream into a discretized multi-image input containing clear temporal relationships that can be efficiently processed by a multimodal large model, temporal video slicing is introduced. Through temporal video slicing, the continuous, complex game video stream is transformed into a structured visual input sequence containing rich temporal cues. This adapts to the current mainstream multimodal large model architecture that receives multiple static images as input, and discretizes the continuous temporal information, forcing the model to focus on and compare changes between frames.

[0029] In an optional embodiment, step 102 above further includes: 1021. Using each time point in the game recording as a base point, extract a video segment of a preset duration before each time point; This optional embodiment does not analyze the entire game replay. Instead, it defines an analysis window (e.g., 20 seconds) with the current time point t as the base point and a fixed length of T. For each second (or smaller time granularity) of the game replay, a 20-second video segment is extracted for analysis. For example, if the goal is to analyze the game at 05:15, the video segment to be extracted would be [04:55, 05:15]. The size of this analysis window should not only be able to capture the evolution of tactical layouts but also avoid information overload and excessive computational costs.

[0030] 1022. Based on a preset video frame sampling strategy, video frames are sampled for each video segment to obtain a time-series video frame sequence with the same number of frames corresponding to each time point.

[0031] Considering that too many video frames in a 20-second video clip might exceed the input length limit of a multimodal large model, it is necessary to perform video frame sampling on each video clip.

[0032] For example, a uniform sampling strategy can be used. For a 20-second window, with a sampling frame count N=10 frames and a sampling interval of 20 seconds / (10-1) ≈ 2.22 seconds, the specific sampling time points are t-20s, t-17.78s, t-15.56s, …, t. To obtain richer information, keyframes can also be sampled.

[0033] For example, first define key events as triggers for keyframes, such as: any hero's health dropping by more than 15%; any hero releasing their ultimate skill (detectable by UI changes in the skill bar); hero icons clustering on the minimap (more than 3 heroes within close proximity at screen coordinates). Within a 20-second analysis window, first perform basic uniform sampling at a low density (e.g., one frame every 5 seconds, for a total of 4 frames). Simultaneously, scan the game log within this window (if available) or analyze the video stream in real time using a lightweight CV model to detect the occurrence times of the aforementioned key events. Collect 1-2 frames within 1 second before and after each key event. Then merge the base frames and keyframes and sort them by time. If the total exceeds N (e.g., 10 frames), prioritize retaining keyframes or uniformly thinning them out. This sampling strategy ensures the model captures moments of tactical shifts, such as the initiation of ganks or the triggering of team fights, significantly improving the relevance and density of the input information.

[0034] The sampled N frames [Image_1, Image_2, ..., Image_N] need to be organized into a well-defined, ordered sequence for processing by the multimodal large language model. To enable the model to understand the order of the frames, temporal information needs to be added when the images are input into the model.

[0035] Image serialization: Arrange N frames of images into a list in chronological order.

[0036] Positional encoding: Borrowing from the idea of ​​the Transformer model, a learnable temporal positional embedding is assigned to each frame of the image. For example, the positional embedding corresponding to the i-th frame is PE_i. Inside the model, the image feature F_i is added to PE_i to obtain the feature F'_i with temporal information = F_i + PE_i.

[0037] Timestamps in text prompts: Time information can also be implicitly added to the structured text prompts constructed, such as "This is the first screenshot (T-20s), this is the second screenshot (T-18s)...", to further enhance the model's time sequence awareness.

[0038] 103. Construct a predefined format of structured text prompts, and use the video frame sequence corresponding to each time point and the structured text prompts together as training samples; In this embodiment, to transform a vague prediction task into a well-defined "fill-in-the-blank" question with clear guidance and output format, structured text prompts are introduced. A complete structured text prompt consists of the following four core parts: (1) Task instructions, used to inform the large model of its role and to declare its task. By setting the role of "professional game analyst," the knowledge about games, strategies, and logical reasoning learned by the multimodal large model during the pre-training phase is activated. This clear task declaration outlines the core objectives that the model needs to accomplish. For example, the following task instructions are given in structured text prompts: "You are a professional multiplayer online game battle analyst. Your task is to predict the probability of a key event occurring in different areas within the next 10-15 seconds, based on the 8 game screenshots below arranged in chronological order." (2) Task definition: This is used to inform the large model to provide precise and quantifiable definitions for terms such as "roaming and ganking" and "teamfight," which have specific meanings in the game context, thus avoiding the large model using vague definitions from its general knowledge base. For example, the following task definition is given in the structured text prompt: Gank: Refers to a surprise attack launched by 1-2 heroes against a lone hero. Teamfight: Refers to a battle of 3v3 or larger. Objective: Refers to a battle fought between two sides around the Tyrant or Lord.

[0039] (3) Task analysis guidelines are used to guide model analysis and address the issue of multimodal large language models being insensitive to key UI information in video frames. These guidelines guide the large model to actively seek out and interpret structured information that has a decisive impact on the battle but is easily overlooked. For example, emphasizing the "ultimate skill cooldown indicator" directly prompts the model to make tactical inferences such as "Hero A's ultimate skill is ready, while Hero B's is not, so Hero A may be more aggressive." For example, the following task analysis guidelines are provided in the structured text prompts: For example, please pay close attention to the following trends in the screenshots: (1) Minimap: the movement trajectory, concentration and disappearance / appearance of hero icons (entering bushes); (2) Hero status bar: changes in health, mana and most importantly, the ultimate skill cooldown indicator; (3) Battle situation: the relative positions of heroes, whether they have entered bushes, the attack range of turrets, etc.

[0040] (4) Model output format, used to constrain the output of large models, specifically indicating the probability of game events occurring with event type and spatial location in the model output, realizing the transformation of model output from fuzzy text to structured data. By enforcing a strict JSON format for output, the model output can be directly parsed and used by downstream programs (such as real-time early warning systems and data analysis platforms). Quantifying the prediction results can provide data support for decision-making. A model output can simultaneously cover multiple event types and multiple spatial regions, greatly improving the model prediction efficiency. For example, the following JSON format model output is given in the structured text prompt: { "gank_probability": {"top_lane": <float>, "mid_lane": <float>, "bot_lane": <float>}, "teamfight_probability": {"river_area": <float>, "base_area": <float>},"objective_contest_probability": { "dragon_pit": <float>, "baron_pit": <float>}} For each video frame sequence obtained in step 102 at time point t, it is paired with the structured Prompt constructed above to form a complete training sample. For example, the input format for multimodal input fusion is: the input part of a training sample is a tuple: (video frame sequence, structured text prompt). This tuple is processed into a unified token sequence when input into the model. For example, [Image_Token_1][Image_Token_2]...[Image_Token_N][Text_Token_1][Text_Token_2].... The image token represents the image features extracted by the visual encoder, and the text token is the digital representation of the structured Prompt.

[0041] Step 103 constructs a combination of a "question" (video frame + prompt) and an "answer template" (JSON format) for each time point of the game recording. This is not just a simple prompting project, but a deep task reconstruction that cleverly transforms the complex temporal prediction problem into a text generation task under strong constraints that M multimodal large language models are good at handling.

[0042] 104. Locate the game events that occur in the game recording, and construct sample labels corresponding to each time point based on the difference between the occurrence time of each game event and each time point; This step is used for game event localization based on weak supervision, constructing high-quality dynamic probability labels. This solves the problem of providing smooth probability signals of "future events are about to occur" for large models even without manual annotation. This embodiment does not require manual viewing of game recordings to annotate events; instead, it uses a script to automatically scan game recordings or their associated structured logs to locate game events.

[0043] In an optional embodiment, step 104 above further includes: 1041. Obtain the game battle log corresponding to the game recording; The core idea of ​​this optional embodiment is to use reliable and easily accessible game logs as anchor points to deduce the time and location of key events in the game. Preferably, a pre-written automated scanning script is used to scan massive amounts of game recordings and their corresponding game logs.

[0044] 1042. Based on the game battle log, locate the time and location of all game events in the game recording; MOBA game replay files contain parsable, chronologically ordered game logs. By analyzing these logs, the time and location of game events can be determined.

[0045] (1) Gank event location Triggering conditions: For example, within a time window of Δt=2 seconds, a sharp drop in player_A's health is detected (e.g., a drop of more than 40%), while the positions of players_B and_C (enemy heroes) change from a distance to near player_A.

[0046] Event Location: The moment when Player_A's health begins to drop drastically, T_event, is defined as the outbreak point of the Gank event. The event location, L_event, is Player_A's current map position (e.g., top lane).

[0047] (2) Teamfight event location Triggering conditions: For example, within a time window of Δt=3 seconds, it is detected that the health of at least 6 different heroes from both sides (in a 3v3 battle) has decreased significantly.

[0048] Event location: The moment T_event, when the first hero's health drops, is defined as the team fight breakout point. L_event is determined by calculating the center point of all participating heroes' positions (e.g., the river area).

[0049] (3) Objective event location Triggering conditions: For example, in "Honor of Kings", the game detects that the Tyrant or Lord's health has started to be damaged by enemy heroes.

[0050] Event positioning: The moment T_event, when a neutral monster first takes damage from a hero, is defined as the trigger point for a resource contention event. L_event is the Dragon Pit or Baron Pit.

[0051] By using scripted scanning, the time point T_peak (i.e., T_event) and location L_event of all key game events can be automatically, efficiently, and accurately located from massive amounts of video recordings.

[0052] 1043. Based on the occurrence time of each game event, set an event prediction time window located before the occurrence time for each game event; After locating game events in the game recording, corresponding sample labels need to be generated for each time point t in the training data. To facilitate the generation of sample labels, for each located game event in the game recording, an event prediction time window is set based on the occurrence time of the game event, located before the occurrence time.

[0053] For example, consider a top lane gank event occurring in T_gank. A prediction time window W is set, such as 0 to 30 seconds, representing the 0 to 30 seconds before the top lane gank event. The significance of setting this prediction time window is that, within this window period, an ideal model should be able to predict the upcoming top lane gank event.

[0054] 1044. Using the difference between the occurrence time and the time point within the event prediction time window as the independent variable, construct a time decay probability function to represent the probability of a game event occurring. This optional embodiment does not use hard labels of "0" or "1", but instead constructs a probability curve that changes smoothly over time as the sample label to match the process of events brewing and erupting in game battles. The sample label is preferably in JSON format, i.e., target JSON.

[0055] The time decay probability function f(Δt) in this optional embodiment preferably adopts the half-life form of a Gaussian function. The range of the function f(Δt) is [0, 1], and it monotonically increases as Δt decreases (i.e., time t gets closer and closer to T_peak).

[0056] f(Δt) = exp( - (Δt) 2 / (2 * σ 2 ) ) Where: Δt = T_peak - t: represents the time difference between the current time t and the event outbreak time T_peak. σ is a configurable time scale hyperparameter, for example, σ = 8 seconds. It is used to control the length of the "precursor period" of the game event and the steepness of the probability curve. The larger σ is, the flatter the probability curve, and the longer the model needs to react in advance.

[0057] At time t, the value of `gank_probability.top_lane` in the target JSON is f(T_peak - t). The probability values ​​for all other events remain 0.0 (or a very small smoothed value).

[0058] For example: at t = T_peak - 15s, Δt = 15s, f(15) might be approximately 0.1. The probability of top_lane in the target JSON is 0.1. At t = T_peak - 8s, Δt = 8s, f(8) might be approximately 0.6. The probability of top_lane in the target JSON is 0.6. At t = T_peak - 1s, Δt = 1s, f(1) might be approximately 0.99. The probability of top_lane in the target JSON is 0.99. For all time points in the game replay that do not belong to any event precursor window, all probability values ​​in the target JSON are 0.0.

[0059] 1045. For each time point within the event prediction time window, based on the time decay probability function, generate positive sample labels corresponding to multiple event types and location orientations; 1046. For each time point outside the event prediction time window, generate a negative sample label corresponding to multiple event types and location points.

[0060] For example, assuming a Gank event occurs in the top lane at T_peak = 10:30, the dynamic probability label is generated for each time point t within the time window [T_peak - 30s, T_peak], i.e. [10:00, 10:30]. JSON format is preferred.

[0061] For example, for t = 10:15 (i.e., T_peak - 15s): Δt = 15s, then f(15) = exp(-(15)) 2 / (2 * 8 2 )) = exp(-225 / 128)≈exp(-1.7578)≈0.172, at this time, the target JSON tag is: {"gank_probability":{"top_lane":0.172,"mid_lane":0.0,"bot_lane":0.0},"teamfight_probability":{"river_area":0.0,"base_area":0.0}, "objective_contest_probability":{"dragon_pit":0.0,"baron_pit": 0.0}} For t = 10:22 (i.e., T_peak - 8s): Δt = 8s, then f(8) = exp(- (8) 2 / (2 * 8 2 )) = exp(-64 / 128) = exp(-0.5)≈0.607, at this time, the target JSON tag is: {"gank_probability":{"top_lane":0.607,"mid_lane":0.0,"bot_lane":0.0},"teamfight_probability":{"river_area":0.0,"base_area":0.0}, "objective_contest_probability":{"dragon_pit":0.0,"baron_pit":0.0}} For other time points, any time point that does not belong to any event precursor window (e.g., T_peak + 5s) will have all probability values ​​of 0.0 in its target JSON. That is, the sample labels corresponding to time points that belong to the event precursor window will be used as positive sample labels, while the sample labels corresponding to time points that do not belong to the event precursor window will be used as negative sample labels.

[0062] In this embodiment, a precise, smooth, and forward-looking target JSON label is generated for each time point t. This label not only tells the model "what will happen in the future," but more importantly, it tells the model "how likely this event is to happen in the future," which provides a high-quality supervision signal for the model to learn its time series prediction capabilities.

[0063] 105. Based on the training samples and sample labels corresponding to each time point, train the multimodal large model to obtain a trained game event prediction model.

[0064] In this embodiment, based on the training samples obtained in step 103 and the sample labels obtained in step 104, an open-source multimodal large model is used as the base model for model training, ultimately resulting in a trained game event prediction model. Through instruction fine-tuning and an innovative composite loss function, the general multimodal large model is trained into a professional game analysis tool. The base model used for training is an open-source multimodal large model.

[0065] In an optional embodiment, step 105 above further includes: 1051. Input the training samples and sample labels corresponding to each time point into the pre-trained multimodal large model in sequence; The complete model input for model training consists of two parts: (1) Input context: video frame sequence + structured prompts, such as ([Image_1, Image_2, ...,Image_N], Structured_Prompt).

[0066] (2) Target label: A target JSON string containing dynamic probability values.

[0067] The following example, using the training sample at time t=10 minutes and 22 seconds, illustrates the model training process.

[0068] Suppose that in the training data stream, the current sample to be processed corresponds to the time point t=10 minutes and 22 seconds of a certain game.

[0069] (a) Preparing the input context: Video frame sequence: According to the time-slicing strategy in step 102, a 20-second video window [10:02, 10:22] is extracted, and N=10 static images are sampled from it. These 10 images are arranged in chronological order as [Image_1, Image_2,..., Image_10]. Suppose that these 10 frames capture a key dynamic: the enemy hero in the top lane begins to disappear from the vision, while the friendly jungler hero is moving from the river to the top lane, and at the same time, the cooldown indicator of the friendly top laner hero's ultimate skill has just lit up.

[0070] Structured Text Prompt: Loads the standard Prompt template designed in step 103, which defines the task instructions, task definition, task analysis guidelines, and model output format.

[0071] (ii) Construction of model input: Choose an open-source pre-trained multimodal large model as the base model. The model architecture supports mixed input of multiple images and text.

[0072] Image encoding: Ten frames of images [Image_1, ..., Image_10] are sequentially passed through a multimodal large-scale visual encoder, and each image is converted into one or more visual token sequences.

[0073] Text encoding: The structured text Prompt is converted into a sequence of text tokens using a tokenizer.

[0074] Sequence concatenation: According to the model's input format requirements, the visual token sequence and the text token sequence are concatenated into a complete input sequence. This sequence may begin with a special start token, followed by an image token, and then a text token.

[0075] (III) Input process: During the training loop, this constructed complete input sequence is fed into the multimodal large model. At this point, the model's parameters are still the initial parameters from the pre-training phase (or parameters that have undergone partial fine-tuning). The model will then begin forward propagation computation based on this input.

[0076] The model input process is not simply about "throwing" the video into the model. Instead, it uses a combination of "temporal slices" and "structured prompts" to provide the model with a highly structured, information-rich, and clearly defined context, laying the foundation for generating high-quality prediction results.

[0077] 1052. Predict game events using the multimodal large model, and output the probability of occurrence of various types and locations of future game events; This step is the process by which the model infers and generates based on the input. Its core lies in how the model transforms its multimodal understanding into a structured, probabilistic predictive output.

[0078] Once the input sequence from step 1051 is fed into the model, the multi-layered Transformer structure within the multimodal large model begins to function. Its self-attention mechanism allows the features of each frame to interact with the features of other frames, as well as with each word in the text Prompt.

[0079] Cross-modal association: The model learns the correspondence between "the small map in the text prompt" and "the small square area in the upper left corner of the image frame".

[0080] Temporal relationship modeling: By comparing the changes in the hero icon positions from Image_1 to Image_10, the model can understand the dynamic concept of "movement trajectory".

[0081] Semantic understanding and reasoning: Combining the definition of "Gank" in the Prompt with the visual cue of "ultimate cooldown indicator", the model may form an implicit reasoning chain internally: "Because the enemy hero disappears, and our jungler and ultimate hero approach at the same time, the probability of a gank occurring in the top lane is very high."

[0082] After complex internal calculations, the model ultimately generates a text sequence as output. Due to the extremely strict output format requirements given in the Prompt, and the requirement to "not add any additional interpretations," the model is trained to only generate text conforming to JSON format.

[0083] For example, the model output for t=10 minutes and 22 seconds is as follows: For an input at t=10 minutes and 22 seconds, after forward propagation, assuming the model output is pred_json_str ={"gank_probability":{"top_lane":0.58,"mid_lane":0.05,"bot_lane":0.02},"teamfight_probability":{"river_area":0.15,"base_area":0.01}, "objective_contest_probability":{"dragon_pit":0.10,"baron_pit":0.00}} The model's output, `pred_json_str`, represents its preliminary prediction of the battle situation within the next 10-15 seconds. The output format is fixed and can be reliably parsed by downstream programs. Probabilistic numerical values ​​quantify the likelihood of key game events occurring. It also predicts the probability of ganks, team fights, and resource contention occurring in different areas. This output format transforms the model's predictive capabilities into signals that can be directly used for automated decision-making. The model output is no longer vague natural language (e.g., "a fight might break out in the top lane"), but rather deterministic, parsable, probabilistic, and multi-dimensional structured data.

[0084] 1053. Based on the sample label and the outbreak probability corresponding to each time point, construct a target loss function, which includes a language modeling loss function and a predicted value regression loss function; Using only standard language modeling loss is insufficient to guarantee the accuracy of the predicted values ​​for large models. Therefore, a double loss function was designed: L_total = L_lm + α * L_reg.

[0085] Assume that during model training, the input to the model is a sequence of video frames and a structured prompt, and the target output of the model is target_json.

[0086] (1) Language modeling loss L_lm (cross-entropy loss), used to ensure that the text generated by the model is in grammatically correct JSON format. The calculation process is as follows: a. The model performs forward propagation, generating a text sequence pred_text, for example, pred_text = {"gank_probability":{"top_lane":0.508,"mid_lane":0.0,"bot_lane":0.0},"teamfight_probability":{"river_area":0.0,"base_area":0.0}, "objective_contest_probability":{"dragon_pit":0.0,"baron_pit":0.0}}.

[0087] b. Tokenize pred_text and the corresponding target tag target_json (also considered as a string) to obtain the token sequence pred_tokens and target_tokens.

[0088] c. Using the standard cross-entropy loss function, calculate the difference between the probability distribution of each token predicted by the model and the actual token.

[0089] (2) The regression loss L_reg (mean squared error loss) is used to supervise and improve the accuracy of the predicted values ​​of large models. Calculation process: a. Parse all probability values ​​in real time from the text pred_text generated by the model.

[0090] b. Obtain the corresponding real value p_true from the target tag target_json.

[0091] c. Calculate the mean squared error (MSE) between all predicted and actual values.

[0092] (3) The weighting factor α is a hyperparameter used to balance the importance of L_lm and L_reg. For example, α can be set to 0.7, indicating that more emphasis is placed on the accuracy of numerical prediction. The value of α needs to be determined through experiments (such as grid search) to maximize the prediction accuracy while ensuring the correct JSON format.

[0093] 1054. Based on the target loss function, fine-tune the model parameters of the multimodal large model to obtain a trained game event prediction model.

[0094] In this embodiment, all (training samples, sample labels) generated in steps 103 and 104 are paired and integrated into a large training dataset. The dataset contains a large number of positive samples (event precursor windows) and negative samples (calm periods) to avoid model bias. During model training, the large model will adjust its parameters based on the calculated loss L_total so that it can perform better when encountering similar situations in the future.

[0095] (1) The training parameters are configured as follows: Optimizer: Choose the AdamW optimizer, which works very well for Transformer class models.

[0096] Learning rate: A small learning rate can be set. Since fine-tuning is being done on a pre-trained model that is already very strong, a small learning rate can avoid destroying its original knowledge.

[0097] Batch size: refers to the number of samples used to calculate the gradient of model parameter updates in a single training iteration. This is set based on the GPU memory size.

[0098] Training rounds: Fine-tuning usually doesn't require too many rounds; 3-5 epochs are enough to see results.

[0099] (2) Backpropagation and parameter update Calculating gradients: After obtaining the total loss L_total of the model, the training framework (such as PyTorch or TensorFlow) automatically executes the backpropagation algorithm. This algorithm calculates the gradient (partial derivative) of L_total with respect to each trainable parameter (weight and bias) in the model. The gradient indicates the direction and magnitude by which each parameter should be adjusted to reduce the loss.

[0100] Optimizer Update: An optimizer, such as AdamW, is used to update the model's parameters based on the calculated gradients. The update formula is: New parameters = Old parameters - learning_rate * Gradient. The learning rate is a hyperparameter that controls the step size for each update.

[0101] (3) Model fine-tuning process For example, the value of L_reg (e.g., 0.01) contributes significantly to L_total because the model's predicted 0.58 differs from the actual 0.607. Backpropagation calculates how this discrepancy is caused by the combined effects of millions or even billions of parameters within the model. The AdamW optimizer fine-tunes these parameters. For instance, it might slightly strengthen the connection weights between "ultimate cooldown indicator on" and "increased gank probability," while weakening some unrelated connections between "river champion movement" and "top lane gank." After this update, if the model sees a similar input sequence again, it is more likely to generate an output with a top_lane probability closer to 0.607.

[0102] (4) Iterative training and model convergence The above process (steps 1051-1054) is not performed only once for a single sample, but rather through multiple rounds of iterative training using a massive amount of pre-constructed sample data pairs (input context, target JSON). Through repeated "prediction-scoring-learning" cycles, the model gradually learns to establish strong correlations between specific visual temporal patterns (such as hero disappearance, ultimate skill readiness, location gathering) and specific structured JSON outputs (such as a surge in probability in a certain area). It no longer simply imitates text, but truly internalizes the reasoning ability from visual observation to probabilistic prediction.

[0103] Training stops when the model's loss on the validation set no longer decreases, or when the prediction performance reaches a preset standard. The model parameters obtained at this point are the final "trained game event prediction model". This model has been successfully trained from a general-purpose multimodal large model into a professional multiplayer online game battle situation analyzer with forward-looking prediction capabilities. It can be deployed in practical applications to receive real-time game streams and output high-quality battle situation predictions.

[0104] Through training with massive amounts of data, the large model eventually learned to establish strong correlations between specific visual temporal patterns (such as heroes moving into bushes or ultimate abilities completing their cooldown) and specific structured JSON outputs (such as an increased probability of ganking in a certain area), thereby internalizing predictive capabilities.

[0105] In this embodiment, through the detailed implementation of the five steps 101-105 above, a general multimodal large language model was successfully "transformed" and "adapted" into a professional multiplayer online game battle situation analyzer with forward-looking predictive capabilities through a closed-loop "input-output-training" design. The fusion of "temporal slicing and structured prompts" at the input end solves the problems of long temporal dependencies and the omission of key information. The "probabilistic JSON" format at the output end achieves a leap from fuzzy description to quantitative prediction. The "weakly supervised dynamic labels" and "composite loss function" at the training end end endow the model with accurate numerical prediction capabilities. This embodiment creatively replaces "cumbersome construction" with "intelligent guidance," representing a more efficient and advanced AI application paradigm, fully demonstrating how to cleverly utilize and enhance the potential of existing large models to solve complex problems in specific domains.

[0106] See Figure 2 One embodiment of the game event prediction method of the present invention includes: 201. Receive the game video to be predicted, wherein the game video includes live game video or game recording; In this embodiment, the game recording refers to the recording of a live game video, and the live game video supports bullet comments. Event prediction can be performed on either the recorded game video or the live game video itself.

[0107] 202. Sample video frames at each time point in the game video to obtain a time-series video frame sequence corresponding to each time point; In this embodiment, to transform a continuous, high-dimensional video stream into a discretized multi-image input containing clear temporal relationships that can be efficiently processed by a multimodal large model, temporal video slicing is introduced. Through temporal video slicing, the continuous, complex game video stream is transformed into a structured visual input sequence containing rich temporal cues. This adapts to the current mainstream multimodal large model architecture that receives multiple static images as input, and discretizes the continuous temporal information, forcing the model to focus on and compare changes between frames.

[0108] 203. Construct a predefined formatted structured text prompt, and use the video frame sequence corresponding to each time point and the structured text prompt together as model input data; In this embodiment, to transform a vague prediction task into a well-defined "fill-in-the-blank" question with clear guidance and output format, structured text prompts are introduced. A complete structured text prompt consists of the following four core parts: (1) Task instructions, used to inform the large model of its role and to declare its task. By setting the role of "professional game analyst", the knowledge about games, strategies, and logical reasoning learned by the multimodal large model during the pre-training phase is activated. This clear task declaration defines the core objectives that the model needs to accomplish.

[0109] (2) Task definition, which is used to inform the large model to give precise and quantitative definitions for words such as "roaming and catching people" and "team battle" that have specific meanings in the game context, so as to avoid the large model using the vague definitions in its general knowledge base.

[0110] (3) Task analysis guide, used to guide model analysis to solve the problem of multimodal large models being insensitive to key UI information in video frames. Guide the large model to actively seek and interpret structured information that has a decisive impact on the battle but is easily overlooked. For example, emphasizing "ultimate skill cooldown indicator" directly prompts the model to make tactical reasoning such as "Hero A's ultimate skill is ready, while Hero B's is not, so Hero A may be more aggressive".

[0111] (4) Model output format, used to constrain the output of large models, specifically indicates the probability of game events occurring with event type and spatial location in the model output, realizing the transformation of model output from fuzzy text to structured data. By enforcing a strict JSON format for output, the model output can be directly parsed and used by downstream programs (such as real-time early warning systems and data analysis platforms). Quantifying the prediction results can provide data support for decision-making. A single model output can simultaneously cover multiple event types and multiple spatial regions, greatly improving the model prediction efficiency.

[0112] 204. Input the model input data into the game event prediction model for processing, and output the probability of the occurrence of various types and locations of future game events. The game event prediction model is obtained based on a multimodal large model training method.

[0113] In this embodiment, the video frame sequence and structured text prompts are input into the trained game event prediction model for processing, and the model can output the probability of the outbreak of various types and locations of future game events.

[0114] For example, suppose we predict the probability of various game events occurring within the next 10-15 seconds at time t=10 minutes and 22 seconds. The model output is assumed to be as follows: {"gank_probability":{"top_lane":0.58,"mid_lane":0.05,"bot_lane":0.02},"teamfight_probability":{"river_area":0.15,"base_area":0.01}, "objective_contest_probability":{"dragon_pit":0.10,"baron_pit":0.00}} The meanings of the above model outputs are as follows: (1) The probability of a gank in the top lane within the next 10-15 seconds is 0.58, the probability of a gank in the mid lane is 0.05, and the probability of a gank in the bottom lane is 0.02. (2) The probability of a team fight occurring on the riverbank within the next 10-15 seconds is 0.15, and the probability of a team fight occurring at the base is 0.01. (3) The probability of a dragon pit resource contest occurring within the next 10-15 seconds is 0.1, and the probability of a lord resource contest occurring is 0.

[0115] This embodiment innovates the model input (temporal video frame sequence + structured prompts) and the model structured output (probability of occurrence of key events in each game), thereby improving the multimodal large model's understanding of key video content and temporal video content, and realizing the ability to predict video content in advance.

[0116] Please see Figure 3 One embodiment of the game event prediction device in this invention includes: The receiving module 301 is used to receive the game video to be predicted, the game video including live game video or game recording; The second sampling module 302 is used to sample video frames at each time point in the game video to obtain a time-series video frame sequence corresponding to each time point. The input construction module 303 is used to construct a predefined format structured text prompt, and uses the video frame sequence corresponding to each time point and the structured text prompt together as model input data; The prediction module 304 is used to input the model input data into the game event prediction model for processing, and output the probability of the outbreak of various types and locations of future game events, wherein the game event prediction model is obtained based on the multimodal large model training method described above.

[0117] Optionally, in one embodiment, the second sampling module 302 is specifically used for: Using each time point in the game video as a base point, extract a video segment of a preset duration before each time point; Based on a preset video frame sampling strategy, video frames are sampled for each video segment to obtain a time-series video frame sequence with the same number of frames corresponding to each time point.

[0118] Optionally, in one embodiment, the structured text prompt includes task instructions, task definition, task analysis guide, and model output format; the task analysis guide is used to guide model analysis, and the model output format is used to indicate the probability of game event outbreaks with event type and spatial location.

[0119] Optionally, in one embodiment, the game event prediction device further includes: Training module 305 is used to acquire game recordings of multi-player game battles; sample video frames at each time point in the game recordings to obtain a time-series video frame sequence corresponding to each time point; construct a predefined formatted structured text prompt, and use the video frame sequence corresponding to each time point and the structured text prompt together as training samples; locate game events occurring in the game recordings, and construct sample labels corresponding to each time point based on the difference between the occurrence time of each game event and each time point; train a multimodal large model based on the training samples and sample labels corresponding to each time point to obtain a trained game event prediction model.

[0120] Optionally, in one embodiment, the training module 305 is specifically used for: Obtain the game battle logs corresponding to the game recordings; Based on the game battle log, the time and location of all game events are located in the game recording; Based on the occurrence time of each game event, an event prediction time window is set for each game event before the occurrence time. Using the difference between the occurrence time and the time point within the event prediction time window as the independent variable, a time decay probability function is constructed to represent the probability of a game event occurring. For each time point within the event prediction time window, positive sample labels corresponding to multiple event types and location orientations are generated based on the time decay probability function; For each time point outside the event prediction time window, generate a corresponding negative sample label with multiple event types and location pointers.

[0121] Optionally, in one embodiment, the training module 305 is further configured to: The training samples and sample labels corresponding to each time point are sequentially input into the pre-trained multimodal large model; The game event is predicted using the multimodal large model, and the output includes the probability of the occurrence of various types and locations of the game events in the future. Based on the sample label and the outbreak probability corresponding to each time point, a target loss function is constructed, which includes a language modeling loss function and a predicted value regression loss function. Based on the target loss function, the model parameters of the multimodal large model are fine-tuned to obtain a trained game event prediction model.

[0122] Since the embodiments of the device part correspond to the embodiments of the above method, please refer to the above method embodiments for the introduction of the game event prediction device provided by the present invention. The present invention will not be described again here, but it has the same beneficial effects as the above game event prediction method.

[0123] Figure 4 This is a schematic diagram of the structure of a computer device 500 provided in an embodiment of the present invention. The computer device 500 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 510 (e.g., one or more processors) and a memory 520, and one or more storage media 530 (e.g., one or more mass storage devices) for storing application programs 533 or data 532. The memory 520 and storage media 530 can be temporary or persistent storage. The program stored in the storage media 530 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the computer device 500. Furthermore, the processor 510 may be configured to communicate with the storage media 530 and execute the series of instruction operations in the storage media 530 on the computer device 500.

[0124] Computer device 500 may also include one or more power supplies 540, one or more wired or wireless network interfaces 550, one or more input / output interfaces 560, and / or one or more operating systems 531, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 4 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0125] The present invention also provides a computer device, the computer device including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor performs the steps of the game event prediction method described in the above embodiments. The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the computer performs the steps of the game event prediction method described in the above embodiments.

[0126] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0127] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.< / float> < / float> < / float> < / float> < / float> < / float> < / float>

Claims

1. A method for predicting game events, characterized in that, The game event prediction method includes: Receive game videos to be predicted, including live game videos or game recordings; Video frames are sampled at each time point in the game video to obtain a time-series video frame sequence corresponding to each time point. Construct a predefined format of structured text prompts, and use the video frame sequence corresponding to each time point and the structured text prompts together as model input data; The input data of the model is fed into a pre-trained game event prediction model for processing, and the output includes the probability of the occurrence of various types and locations of future game events.

2. The game event prediction method according to claim 1, characterized in that, The step of sampling video frames at each time point in the game video to obtain a time-series video frame sequence corresponding to each time point includes: Using each time point in the game video as a base point, extract a video segment of a preset duration before each time point; Based on a preset video frame sampling strategy, video frames are sampled for each video segment to obtain a time-series video frame sequence with the same number of frames corresponding to each time point.

3. The game event prediction method according to claim 1, characterized in that, The structured text prompts include task instructions, task definitions, task analysis guidelines, and model output formats; the task analysis guidelines are used to guide model analysis, and the model output formats are used to indicate the probability of game events occurring with event types and spatial location orientations.

4. The game event prediction method according to any one of claims 1-3, characterized in that, The training methods used in the game event prediction model include: Acquire game recordings of multiplayer battles; Video frames are sampled at each time point in the game recording to obtain a time-series video frame sequence corresponding to each time point; Construct a predefined format of structured text prompts, and use the video frame sequence corresponding to each time point and the structured text prompts together as training samples; Locate the game events that occur in the game recording, and construct sample labels corresponding to each time point based on the difference between the occurrence time of each game event and each time point; Based on the training samples and sample labels corresponding to each time point, the multimodal large model is trained to obtain a trained game event prediction model.

5. The game event prediction method according to claim 4, characterized in that, The process of locating game events occurring in the game recording and constructing sample labels corresponding to each time point based on the difference between the occurrence time of each game event and each time point includes: Obtain the game battle logs corresponding to the game recordings; Based on the game battle log, the time and location of all game events are located in the game recording; Based on the occurrence time of each game event, an event prediction time window is set for each game event before the occurrence time. Using the difference between the occurrence time and the time point within the event prediction time window as the independent variable, a time decay probability function is constructed to represent the probability of a game event occurring. For each time point within the event prediction time window, positive sample labels corresponding to multiple event types and location orientations are generated based on the time decay probability function; For each time point outside the event prediction time window, generate a corresponding negative sample label with multiple event types and location pointers.

6. The game event prediction method according to claim 4, characterized in that, The step of training a multimodal large model based on the training samples and sample labels corresponding to each time point to obtain a trained game event prediction model includes: The training samples and sample labels corresponding to each time point are sequentially input into the pre-trained multimodal large model; The game event is predicted using the multimodal large model, and the output includes the probability of the occurrence of various types and locations of the game events in the future. Based on the sample label and the outbreak probability corresponding to each time point, a target loss function is constructed, which includes a language modeling loss function and a predicted value regression loss function. Based on the target loss function, the model parameters of the multimodal large model are fine-tuned to obtain a trained game event prediction model.

7. A game event prediction device, characterized in that, The game event prediction device includes: A receiving module is used to receive game videos to be predicted, including live game videos or game recordings. The second sampling module is used to sample video frames at each time point in the game video to obtain a time-series video frame sequence corresponding to each time point. The input construction module is used to construct a predefined format of structured text prompts, and uses the video frame sequence corresponding to each time point and the structured text prompts together as model input data; The prediction module is used to process the input data of the model into a pre-trained game event prediction model and output the probability of the occurrence of various types and locations of future game events.

8. The game event prediction device according to claim 7, characterized in that, The game event prediction device also includes: The training module is used to acquire game recordings of multi-player game battles; sample video frames at each time point in the game recordings to obtain a time-series video frame sequence corresponding to each time point; construct a predefined formatted structured text prompt, and use the video frame sequence corresponding to each time point and the structured text prompt together as training samples; locate game events occurring in the game recordings, and construct sample labels corresponding to each time point based on the difference between the occurrence time of each game event and each time point; train a multimodal large model based on the training samples and sample labels corresponding to each time point to obtain a trained game event prediction model.

9. A computer device, characterized in that, The computer device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the computer device to execute the game event prediction method as described in any one of claims 1-6.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements the game event prediction method as described in any one of claims 1-6.

Images