Game scene element detection and situation awareness method and system based on VLM technology

Abstract

This invention discloses a method and system for element detection and situational awareness in game scenarios based on VLM technology, belonging to the field of artificial intelligence technology. The method includes: combining a map background image to obtain dense and non-dense areas of elements in the current game interaction interface; recursively segmenting the dense areas until the segmentation result is a non-dense area or no larger than a minimum size setting; identifying elements in the non-dense areas and the dense areas no larger than the minimum size setting, and constructing a knowledge graph based on these elements; and performing reasoning based on the knowledge graph to obtain the situational awareness result of the current game interaction interface. This invention can overcome the key bottleneck from pixel perception to semantic decision-making, meeting the urgent needs of modern large-scale interactive systems for real-time, accurate, and intelligent situational awareness.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for detecting elements and recognizing the situation in a game scenario based on VLM (Visual Language Model) technology. Background Technology

[0002] With the development of information technology, large-scale interactive systems, such as command and control systems and digital twin cities, commonly adopt two-dimensional or three-dimensional map interfaces as the core human-computer interaction carrier. These interfaces integrate symbolic representations of terrain, administrative divisions, spatial coordinates, and various participating entities (such as vehicles, personnel, and facilities), collectively forming a comprehensive situational map reflecting the system's operational status. The types, quantities, distributions, and relationships of participating entities contain rich semantic information, which is crucial for assessing system scale, interaction intensity, and behavioral intentions. Therefore, automatically and accurately identifying various geographical elements and participating entities from complex system interface images, and further understanding the situational information they represent, is of great significance for achieving real-time monitoring of system status, supporting decision-making, and intelligent game theory.

[0003] Currently, the technology for identifying interface elements from images mainly relies on the following three types of methods.

[0004] 1. Rule-based image processing methods: These methods typically utilize traditional image processing algorithms such as color space (e.g., HSV) segmentation, edge detection, and template matching to identify specific icons or regions. However, their recognition rules are highly dependent on preset interface visual styles (e.g., fixed colors and shape templates). When the system software updates the interface theme, or when the user uses custom entity symbols, the recognition accuracy of such methods drops significantly, lacking the necessary adaptability and robustness.

[0005] 2. Deep Learning-Based Object Detection Models: These models employ general-purpose object detection models such as Faster R-CNN and the YOLO series, enabling them to automatically learn and detect elements within an interface. However, these methods fall under supervised learning, requiring the collection and labeling of massive amounts of training data specific to the system interface to train a dedicated detector. This results in high model costs and poor transferability. More importantly, these purely visual models are unimodal, capable of processing only pixel features and unable to correlate and understand semantic information closely related to entities on the interface, such as text labels and status descriptions, leading to incomplete entity identification.

[0006] 3. Manual input or semi-automatic annotation: Operators directly observe the interface and manually record or annotate entity information. When facing large-scale, dynamic interactive scenarios with complex entity relationships, this method suffers from low efficiency, high subjectivity, fatigue, and the tendency to miss key information, making it difficult to meet the needs for real-time and accurate situational awareness.

[0007] In summary, existing technologies generally suffer from the following common shortcomings: First, they lack a single perceptual modality. Most existing methods only analyze low-level visual features of images, lacking joint modeling and understanding of semantic information in the interface text, resulting in a deficiency in the cognition of entity functions. Second, they lack high-level reasoning capabilities. Existing technologies typically only output the bounding box positions of elements and a simple category label, remaining at the level of "what is and where," unable to further analyze and reason about the spatial relationships, aggregation states, and behavioral trends between entities, making it difficult to support high-level situational understanding and decision inference. Third, they have poor adaptability to high-density, small-target scenarios. In the hotspot areas of large interactive systems, a large number of entity symbols are often densely deployed or even mutually occluded. Traditional detection models, due to receptive field limitations and feature confusion, are prone to serious false negatives in such scenarios, leading to a sharp drop in recall and creating a "blind spot" in situational awareness. Therefore, existing technologies have not yet achieved an end-to-end intelligent processing flow from raw, complex interactive system interface images to structured, semantic situational awareness. Summary of the Invention

[0008] This invention discloses a method and system for game scenario element detection and situational awareness based on VLM technology. It can adapt to different style interfaces with zero samples, robustly perceive dense small targets, and perform multi-level semantic reasoning, so as to break through the key bottleneck from pixel perception to semantic decision-making and meet the urgent needs of modern large-scale interactive systems for real-time, accurate and intelligent situational awareness.

[0009] To achieve the above objectives, the technical solution of the present invention includes the following:

[0010] A method for game scenario element detection and situational awareness based on VLM technology, the method comprising: By combining the map background image, identify the dense and non-dense areas of elements in the current game interaction interface; Recursively divide the densely populated areas of features until the division result is a non-densely populated area of ​​features or no larger than the minimum size setting; Identify elements in non-dense areas and dense areas of elements no larger than the minimum size setting, and construct a knowledge graph based on these elements; Reasoning based on knowledge graphs yields the current situational awareness results of the game interaction interface.

[0011] Furthermore, the step of combining the map background image to obtain the dense and non-dense areas of elements in the current game interaction interface includes: The RGB channel difference of the current game interaction interface and the map background image is calculated pixel by pixel, and the L2 norm is taken to obtain the grayscale residual image. The residual map is divided into a fixed number of grids, and the average value of the residual values ​​of all pixels within the grid is calculated. Based on this average value, the current game interaction interface is divided into a factor-intensive area and a factor-inefficient area.

[0012] Furthermore, the densely populated areas of features are recursively subdivided until the subdivision results in non-densely populated areas of features or areas not larger than the minimum size setting, including: Check the size of areas with dense features; For feature-dense areas that are larger than the minimum size setting, the feature-dense areas are divided into four-way blocks, and each block is determined to be a feature-dense area or a feature-non-dense area. If the block is a densely populated area, the size check of the densely populated area is performed again until the block is a non-densely populated area or the block is not larger than the minimum size setting.

[0013] Furthermore, identifying features in non-dense feature areas and dense feature areas not exceeding the minimum size setting includes: Based on the feature detection task, generate prompt text; The prompt text and input image are input into the visual language model to obtain the bounding box coordinates, semantic label, and confidence score of each feature in the input image; wherein, the input image includes non-dense feature areas or dense feature areas not larger than the minimum size setting.

[0014] Furthermore, a knowledge graph is constructed based on this element, including: Based on the bounding box coordinates, map the feature back to the global coordinate system; All elements from different regions are merged, and nodes in the knowledge graph are obtained based on the merged results; the attributes of the node include storage type, state, spatial location, confidence level, and detection time. The spatial relationship judgment function is called to add relationship edges between nodes; wherein, the relationship edges include spatial proximity edges, type clustering edges, geographic association edges, and threat association edges. The spatial proximity edges are used to describe the geographic distance between two entities that is less than a threshold. The type clustering edges are used to describe entities of the same type that are in the same grid. The geographic association edges are used to describe the relationship of entities located on geographic features. The threat association edges are used to describe the entity status that triggers threat rules.

[0015] A game scenario element detection and situational awareness system based on VLM technology, the system comprising: The image segmentation module is used to combine the map background image to obtain the dense and non-dense areas of elements in the current game interaction interface; the dense areas of elements are recursively segmented until the segmentation result is a non-dense area of ​​elements or not larger than the minimum size setting. The knowledge graph construction module is used to identify elements in non-dense areas and dense areas of elements no larger than the minimum size setting, and to construct a knowledge graph based on these elements. The situational awareness module is used to perform reasoning based on knowledge graphs to obtain the situational awareness results of the current game interaction interface.

[0016] A computer device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the game scenario element detection and situational awareness method based on VLM technology as described above.

[0017] A computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the game scenario element detection and situational awareness method based on VLM technology described above.

[0018] A computer program product, characterized in that, when the computer program product is run on a computer device, the computer device executes the game scenario element detection and situational awareness method based on VLM technology described above.

[0019] Compared with the prior art, the present invention has at least the following beneficial effects.

[0020] This invention demonstrates outstanding robustness and efficiency in complex interactive scenarios. In areas with a high density of entities, the system, through an adaptive block-segmentation mechanism, can continuously and accurately capture every key object, completely resolving the missed detection problem caused by the small size or feature confusion of traditional vision models, and providing users with a complete and blind-spot-free foundation for scene perception. The recursive block-segmentation strategy enables intelligent allocation of computing resources, avoiding the redundant overhead of global high-resolution processing, enabling the system response speed to reach millisecond levels, fully meeting the needs of real-time interaction.

[0021] The core breakthrough of this invention lies in achieving a leap from raw visual data to high-level semantic understanding. The system not only outputs an entity list but also analyzes spatial relationships between entities through a multi-level reasoning framework, including clustering patterns and path dependencies, and identifies dynamic behaviors such as traffic changes and resource distribution. Ultimately, the system can automatically generate interpretable situational conclusions. This significantly reduces the user's information processing burden, freeing them from tedious data filtering and allowing them to focus on optimizing strategy formulation and resource scheduling. The final output of structured data and natural language summaries intuitively matches user cognitive habits, greatly improving human-machine collaboration efficiency. As a general-purpose intelligent sensing technology, this invention provides a reliable technological foundation for decision support systems in fields such as large-scale virtual environment interaction and smart city management. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a method for game scenario element detection and situational awareness based on VLM technology.

[0023] Figure 2 A detailed flowchart of a game scenario element detection and situational awareness method based on VLM technology.

[0024] Figure 3 Block diagram of a sonar image denoising system based on neural network self-search.

[0025] Figure 4 A block diagram of computer equipment. Detailed Implementation

[0026] The present invention will be further described below with reference to possible accompanying drawings and specific embodiments, but this does not constitute any limitation on the present invention.

[0027] The game scenario element detection and situational awareness method based on VLM technology of the present invention, such as Figure 1 and Figure 2 As shown, it includes the following steps.

[0028] Step S1: Combine the map background image to obtain the dense and non-dense areas of elements in the current game interaction interface.

[0029] (1) System initialization and map background modeling.

[0030] Map background acquisition: Upon first launch of the system, the user is required to enter an "empty map" state in the system software, i.e., a clean interface with no entities, markers, or pop-ups on the map, and manually capture a standard map background image. The system archives this image as a benchmark for subsequent occlusion assessments.

[0031] Coordinate calibration: The system requires the user to input or obtain two key parameters simultaneously through interface parsing: the spatial coordinates of the upper left corner of the map. , ), and the map scale These parameters are used to convert pixel coordinates into geographic coordinate information.

[0032] (2) Real-time processing flow.

[0033] Image capture: By calling the system interface, the current image of the game interaction interface is acquired in real time at a certain frame rate. .

[0034] Coverage residual calculation: The system calculates the residual for the occlusion. and pre-stored Pixel-by-pixel RGB channel interpolation is calculated, and the L2 norm is taken to obtain a grayscale residual image. The smaller the residual value, the closer the pixel is to the original map, meaning it's less likely to be covered by entities; conversely, the larger the residual value, the more likely the pixel has been modified, meaning it's covered by entities. Additionally, in some cases, the average of local residuals can be used as an optional criterion.

[0035] Initial grid division and threshold determination: The entire image is divided into a fixed number of grids. In this invention, a uniform 16×16 grid is used for initial image division. For each grid, the average value of the residual values ​​of all pixels within it is calculated. An empirical threshold is then set. In the application of this invention, a typical empirical value of 30 was obtained through statistical learning with a small number of samples. If the average residual value of a certain grid is lower than... If so, then mark it as a candidate dense region.

[0036] In summary, this invention proposes using the degree of map background occlusion as a criterion for high-density areas. In the system interface, entities are typically overlaid on the map background as icons. When the density of entities on the map is too high, the map texture is severely obscured. Therefore, entity density can be indirectly assessed by calculating the visibility of the map background in a local area. Let the original map background be... The current interface is Then define the masking residual map. This indicates the visibility of the map texture after it is covered by the icon, specifically: In densely populated areas, The value is smaller because the entities cover a larger portion of the map background. Conversely, in areas with a smaller entity distribution, The value is relatively large. The system... After performing morphological opening operations, the computational region residual strength : in, This indicates the number of pixels in the region. If... ( If the threshold is set to a preset value, then the area is determined to be a dense area.

[0037] Step S2: Recursively divide the densely populated areas of features until the division result is a non-densely populated area of ​​features or reaches the minimum size setting.

[0038] Recursive processing function: The system starts a recursive function for each "candidate dense region" (let the function name be...). The logic of this function is as follows.

[0039] 1) Termination Condition Judgment: First, check the size of the current processing area. If its width or height is already smaller than the fixed geographic area (minimum size setting), then no further segmentation is performed, and the process jumps directly to step 3). The minimum fixed geographic area is a preset parameter that can be calculated using spatial scale and scale information. This serves as the minimum recommended size that the model can effectively process; areas that are too small are considered insufficient to contain dense adversarial unit areas.

[0040] 2) Quadrilateral Division: If the termination condition is not met, the current region is precisely divided into four equal-sized sub-regions along the horizontal and vertical midlines. For each of the four generated sub-regions, the system recalculates their respective average residual values. For any sub-region, if its average residual value is still below the threshold η, the process is recursively called on that sub-region. function.

[0041] Specifically, for areas identified as densely populated areas The system divides it into four equal sub-blocks along its length and width: , , , Execute respectively The execution result will be passed through The operation merges the execution results. When processing each sub-block, if its occlusion intensity is still below the threshold, recursively divide the block and amplify it; otherwise, stop splitting and call the detector on that sub-block. Perform the detection. This process can be represented as: This mechanism avoids unnecessary global amplification and significantly improves processing efficiency.

[0042] 3) Output image set. This image set contains non-dense areas of features and dense areas of features smaller than the minimum size setting.

[0043] Step S3: Identify elements in non-dense areas and dense areas of elements that are no larger than the minimum size setting, and construct a knowledge graph based on these elements.

[0044] First, for non-dense areas and feature-dense areas not exceeding a minimum size setting, this invention invokes a visual language model to perform detection. The model's input is an image and a fixed cue word string. Image patches of the current region (or sub-region) are input into the visual language model. In this invention's application, the cue word is set to: "Identify all entities in the image, including type, state, precise location, and visible geographic features such as rivers, roads, and mountains." In one embodiment, the present invention uses Qwen-VL as the base model to achieve zero-shot semantic detection. Qwen-VL possesses visual grounding capabilities, responding to natural language cues to directly locate and classify entities from images. Therefore, the visual language model serves as the detector. : The input image is The prompt text is T, and the model output is... For a set of triples: in For bounding box coordinates, For semantic labels, such as "entity of type A", The confidence level is used. This mechanism can generalize to unknown entity types without training. The visual language model returns detection results in JSON format, which is parsed to obtain information about all detected features.

[0045] Then, since each local image patch is cropped from the original large image, the bounding box coordinates returned by the visual language model are relative to that local patch. The system then maps all entity positions to the global coordinate system based on the patch's starting coordinates in the original image.

[0046] Next, the system merges all results from different regions. To avoid duplicate detection, the system iterates through all detection boxes. If detection boxes from different regions describe entities of the same category, and the total number of entities exceeds the initially agreed-upon number, they are sorted according to their confidence scores, and results with higher confidence scores are retained. This invention creates each unique entity as a knowledge graph node, storing attributes such as type, state, spatial location, confidence score, and detection time, and establishes a spatial index based on geographic coordinates to support fast regional queries.

[0047] Finally, based on the generated map nodes, the system calls the spatial relationship judgment function to add various relationship edges between the nodes: spatial proximity edges describe that the geographical distance between two entities is less than a threshold; type clustering edges describe that entities of the same type are in the same grid; geographical association edges describe the relationship of entities located on geographical features; and threat association edges describe the entity status that triggers threat rules.

[0048] Step S4: Based on the knowledge graph, reason to obtain the situational awareness results of the current game interaction interface.

[0049] After constructing the edges in the graph, the system performs high-level reasoning using a graph traversal algorithm to generate situational information such as clusters and threats. Finally, the system serializes the node and relationship states of the knowledge graph, integrates them into a situational report, and outputs it in two ways: first, it outputs a file in a pre-defined structured JSON format; second, it generates a natural language summary describing the situational awareness results.

[0050] It should be noted that this invention can be applied to multiple scenarios such as traffic scheduling and autonomous driving. For example, in the complex urban road scenario of autonomous driving, the system first acquires real-time environmental perception data generated by the fusion of sensors such as vehicle cameras and LiDAR, and compares it with pre-stored high-precision maps and prior static environment models. Then, a dynamic target residual map is quickly generated through a change detection algorithm.

[0051] By analyzing the map, the system can instantly identify densely interacting areas among traffic participants, such as busy intersections, merging ramps, pedestrian crossings, and ordinary road sections with sparse traffic. For high-density, high-risk interaction areas, the system initiates a spatial adaptive recursive segmentation mechanism, progressively decomposing complex scenarios into more manageable sub-regions until the number of targets and interaction complexity within each sub-unit are reduced to a level suitable for fine-grained analysis.

[0052] Subsequently, the system inputs this target area data into a multimodal perception and understanding model. This model can not only accurately identify various traffic participants such as cars, pedestrians, bicycles, and traffic police, and output their positions, bounding boxes, and classification labels, but also deeply integrate visual appearance (such as turn signals and gestures), trajectory, speed, and other information to infer their real-time state and intentions. For example, it can detect when a vehicle is activating its left turn signal, when a pedestrian is standing still and looking around, possibly preparing to cross the road, or when a bicycle is going against traffic.

[0053] The system maps all identified and inferred entities and their attributes to a unified spatiotemporal coordinate system, constructing a dynamic traffic scene knowledge graph. The nodes of the graph represent entities (such as vehicle V, pedestrian P, traffic light L) and include attributes (type, location, speed, intent); the edges of the graph represent rich interactions and logical relationships between entities, such as "vehicle V1 is following vehicle V2", "pedestrian P is on crosswalk Z", and "vehicle V3 poses a potential right-turn conflict threat to pedestrian P".

[0054] Ultimately, based on this map, the system performs high-level reasoning, generating accurate and interpretable situational awareness reports that directly serve the planning and control modules. Examples include threat warnings (a vehicle is cutting into the lane from the left ahead at close range; it is recommended to slow down and give way), opportunity recognition (traffic is stagnant ahead in the current lane, traffic is faster in the left lane, and there is sufficient safe distance behind, providing an opportunity to change lanes and overtake), and compliance detection (the green light ahead is about to end; the current speed is insufficient to safely cross the stop line; it is recommended to initiate smooth deceleration and stop). Therefore, this invention achieves end-to-end conversion from raw sensor signals to high-level semantic understanding, enabling autonomous driving systems not only to see the world but also to understand and reason about the ever-changing traffic dynamics.

[0055] Based on the same concept, this invention also discloses a game scenario element detection and situational awareness system based on VLM technology, such as... Figure 3 As shown, the system includes: The image segmentation module is used to combine the map background image to obtain the dense and non-dense areas of elements in the current game interaction interface; the dense areas of elements are recursively segmented until the segmentation result is a non-dense area of ​​elements or not larger than the minimum size setting. The knowledge graph construction module is used to identify elements in non-dense areas and dense areas of elements no larger than the minimum size setting, and to construct a knowledge graph based on these elements. The situational awareness module is used to perform reasoning based on knowledge graphs to obtain the situational awareness results of the current game interaction interface.

[0056] Based on the same concept, this invention also discloses a computer device, which may be a terminal, a laptop computer, a desktop computer, a server, a computer cluster, or other types of computer devices. For example... Figure 4 As shown, the computer device may include at least one processor and memory. The processor can execute instructions stored in the memory. The processor is communicatively connected to the memory via a data bus. In addition to the memory, the processor can also be communicatively connected to input devices, output devices, and communication devices via the data bus.

[0057] The processor can be any conventional processor. Processors may include central processing units (CPUs), graphics processing units (GPUs), field-programmable gate arrays (FPGAs), systems on chips (SoCs), application-specific integrated circuits (ASICs), or combinations thereof.

[0058] Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0059] In this embodiment of the invention, an executable instruction is stored in a memory. The processor can read the executable instruction from the memory and execute the instruction to implement all or part of the steps of the method of the invention.

[0060] Based on the same concept, the present invention also discloses a computer-readable storage medium including a computer program product or storing the computer program product. The computer product includes computer program instructions that can be executed by a processor to perform all or part of the steps described in the exemplary embodiments above.

[0061] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. These programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages, and scripting languages ​​(e.g., Python). The program code can be executed entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0062] Computer-readable storage media can take the form of any combination of one or more readable media. A readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media include: static random access memory (SRAM) having one or more electrically connected wires; electrically erasable programmable read-only memory (EEPROM); erasable programmable read-only memory (EPROM); programmable read-only memory (PROM); read-only memory (ROM); magnetic storage; flash memory; magnetic disk or optical disk; or any suitable combination thereof.

[0063] The above embodiments are merely illustrative of the technical solutions of the present invention and are not intended to limit it. Those skilled in the art can modify or make equivalent substitutions to the above technical solutions based on the concept of the present invention, and such modifications or equivalent substitutions should all be covered within the protection scope of the present invention. The protection scope of the present invention is defined by the claims.

Claims

1. A method for game scene element detection and situation awareness based on VLM technology, characterized in that, The method includes: By combining the map background image, identify the dense and non-dense areas of elements in the current game interaction interface; Recursively divide the densely populated areas of features until the division result is a non-densely populated area of ​​features or no larger than the minimum size setting; Identify elements in non-dense areas and dense areas of elements no larger than the minimum size setting, and construct a knowledge graph based on these elements; Reasoning based on knowledge graphs yields the current situational awareness results of the game interaction interface.

2. The method of claim 1, wherein, The step of combining the map background image to obtain the dense and non-dense areas of elements in the current game interaction interface includes: The RGB channel difference of the current game interaction interface and the map background image is calculated pixel by pixel, and the L2 norm is taken to obtain the grayscale residual image. The residual map is divided into a fixed number of grids, and the average value of the residual values ​​of all pixels within the grid is calculated. Based on this average value, the current game interaction interface is divided into a factor-intensive area and a factor-inefficient area.

3. The method of claim 1, wherein, Recursively segment densely populated areas of features until the segmentation result is a non-densely populated area of ​​features or no larger than the minimum size setting, including: Check the size of areas with dense features; For feature-dense areas that are larger than the minimum size setting, the feature-dense areas are divided into four-way blocks, and each block is determined to be a feature-dense area or a feature-non-dense area. If the block is a densely populated area, the size check of the densely populated area is performed again until the block is a non-densely populated area or the block is not larger than the minimum size setting.

4. The method of claim 1, wherein, Features in non-dense areas and dense areas not exceeding the minimum size setting are identified, including: Based on the feature detection task, generate prompt text; The prompt text and input image are input into the visual language model to obtain the bounding box coordinates, semantic label, and confidence score of each feature in the input image; wherein, the input image includes non-dense feature areas or dense feature areas not larger than the minimum size setting.

5. The method of claim 4, wherein, A knowledge graph is constructed based on this element, including: Based on the bounding box coordinates, map the feature back to the global coordinate system; All elements from different regions are merged, and nodes in the knowledge graph are obtained based on the merged results; the attributes of the node include storage type, state, spatial location, confidence level, and detection time. The spatial relationship judgment function is called to add relationship edges between nodes; wherein, the relationship edges include spatial proximity edges, type clustering edges, geographic association edges, and threat association edges. The spatial proximity edges are used to describe the geographic distance between two entities that is less than a threshold. The type clustering edges are used to describe entities of the same type that are in the same grid. The geographic association edges are used to describe the relationship of entities located on geographic features. The threat association edges are used to describe the entity status that triggers threat rules. 6.A system for game scene element detection and situation awareness based on VLM technology, characterized in that, The system includes: The image segmentation module is used to combine the map background image to obtain the dense and non-dense areas of elements in the current game interaction interface; the dense areas of elements are recursively segmented until the segmentation result is a non-dense area of ​​elements or not larger than the minimum size setting. The knowledge graph construction module is used to identify elements in non-dense areas and dense areas of elements no larger than the minimum size setting, and to construct a knowledge graph based on these elements. The situational awareness module is used to perform reasoning based on knowledge graphs to obtain the situational awareness results of the current game interaction interface.

7. A computer device, comprising: The computer device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the game scenario element detection and situational awareness method based on VLM technology as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the game scenario element detection and situational awareness method based on VLM technology as described in any one of claims 1-5.

9. A computer program product, characterised in that, When the computer program product is run on a computer device, the computer device executes the game scenario element detection and situational awareness method based on VLM technology as described in any one of claims 1-5.

Images