Methods, apparatus, devices, and computer programs for extracting features from 3D scenes

Abstract

The present application discloses a method, apparatus, device, and storage medium for extracting features from a 3D scene, and related embodiments are applicable to various scenes such as artificial intelligence, reducing the cost of extracting 3D scene features. The method includes: projecting a set of visual cone rays from a target character object; when the visual cone rays hit a hit object, returning attribute information of the hit object; performing vector transformation on the object attribute information; performing feature dimension reduction on the basic ray feature vector obtained by the transformation to obtain a ray feature vector; collecting altitude value matrices corresponding to multiple different granularities using the location of the target character object as the collection center; performing feature dimension reduction on the altitude value matrices corresponding to each finer granularity to obtain an altitude map feature vector; and integrating the ray feature vector and the altitude map feature vector into a 3D scene feature.

Description

Technical Field

[0001] This application claims the priority of a Chinese patent application filed with the China National Intellectual Property Administration on December 1, 2022, with an application number of 202211525532.8 and an invention title of "Method, Apparatus, Device, and Storage Medium for Extracting Features of 3D Game Scenes", and all its contents are incorporated herein by reference.

[0002] Embodiments of this application relate to the technical field of image processing, and in particular, to the technical field of feature extraction of 3D scenes.

Background Art

[0003] In actual applications, 3D scenes (such as 3D game scenes, etc.) generally have characteristics such as being unstructured, having a complex variety of objects, complex object shapes, and being difficult to model as data information. Correspondingly, how to efficiently and accurately extract the scene features of 3D scenes has currently become a difficult problem to solve. The current methods for extracting features from 3D scenes mainly include extracting visual image features, depth map features, etc.

[0004] However, there are deficiencies in using the extraction of visual image features to describe 3D scenes. The reason is that the extracted visual image features are essentially projections of 3D scenes on a 2D plane, only containing 2D information of the 3D scene, with a high data dimension of the visual image features. When performing related machine learning tasks based on such visual image features, a high learning cost is consumed.

[0005] Although depth map features contain 3D information of the scene, the data dimension of depth map features is still high, with certain duplication and redundancy in the information contained in pixel points, and depth map features only contain distance information. To perform inductive learning on depth map features, related machine learning tasks still consume a high cost.

[0006] Combining visual image features and depth map features can extract richer scene features, but the problem remains that the extracted feature data is high-dimensional, resulting in a high training cost for machine learning tasks. [Overview of the project] [Problems that the invention aims to solve]

[0007] Embodiments of this application provide a method, apparatus, device, and storage medium for extracting features from a 3D scene, ensuring that the extracted 3D scene features are concise, efficient, semantically rich, and adequately describe the 3D scene. By performing a machine learning task based on these 3D scene features, the ability of the machine learning model to perceive the 3D scene is enhanced, accelerating the development and manufacturing of the machine learning model and reducing the training cost of the machine learning task. [Means for solving the problem]

[0008] One embodiment of the present application provides a method for extracting features from a 3D scene performed by a computer device. The process involves projecting a set of cone-shaped projections from a target character object onto a 3D scene screen, If the aforementioned cone ray hits the hit object, the step of returning the object attribute information of the hit object, The process involves performing a vector transformation on the object attribute information of each received hit object to obtain the basic trajectory feature vector, The steps include: performing a feature dimensionality reduction process on the aforementioned basic ray feature vector to obtain a ray feature vector; The steps include: collecting elevation value matrices corresponding to each granularity, with the location of the target character object as the collection center, based on different granularities; A feature dimensionality reduction process is performed on the elevation value matrix corresponding to each of the aforementioned granularities to obtain the elevation map feature vector; The method includes the step of integrating the ray feature vector and the altitude map feature vector into a 3D scene feature corresponding to the 3D scene screen.

[0009] Another aspect of this application provides a feature extraction device for a three-dimensional scene. A processing unit that projects a set of cone-shaped projections from a target character object onto a 3D scene screen, The system includes an acquisition unit that, when the aforementioned cone ray hits an object, returns object attribute information of the hit object, The processing unit further performs a vector transformation on the object attribute information of each received hit object to obtain a basic trajectory feature vector. The processing unit further performs a feature dimensionality reduction process on the basic ray feature vector to obtain a ray feature vector, The processing unit further collects elevation value matrices corresponding to each granularity, with the location of the target character object as the collection center, based on different granularities. The processing unit further performs a feature dimensionality reduction process on the height value matrix corresponding to each granularity to obtain a height map feature vector. The system further includes a decision unit that integrates the aforementioned ray feature vector and the aforementioned altitude map feature vector into a 3D scene feature corresponding to the 3D scene screen.

[0010] Other aspects of this application provide computer equipment, Memory for storing programs, When executing a program in memory, a processor implements each of the above methods, It includes a bus system that connects memory and a processor for communication.

[0011] Another aspect of this application provides a computer-readable storage medium in which commands are stored, and when the commands are executed by a computer, the computer causes the computer to execute the methods of each of the above aspects. [Effects of the Invention]

[0012] As can be seen from the above technical proposals, the embodiments of this application have the following beneficial effects: A set of cone rays is emitted from the target character object. When a cone ray hits a target object, the object attribute information of the target object is returned. A vector transformation is performed on the received object attribute information of each target object to obtain the corresponding basic ray feature vector. Feature dimensionality reduction is performed on the basic ray feature vector to obtain a ray feature vector. Additionally, elevation value matrices corresponding to each granularity are collected, with the location of the target character object as the collection center, based on different granularities. Feature dimensionality reduction is performed on each elevation value matrix corresponding to each granularity to obtain an elevation map feature vector. Finally, the ray feature vector and elevation map feature vector are integrated into a 3D scene feature corresponding to the 3D scene screen.

[0013] According to the above method, object attribute information of the hit object can be quickly obtained from the cone rays that hit the object. Furthermore, the corresponding base ray feature vector is obtained by transforming based on the object attribute information of the hit object. As a result, the base ray feature vector contains the attribute information of the object in the 3D scene, and since the number of cone rays emitted is far lower than the number of pixels such as visual image features and depth map features, the obtained base ray feature vector is concise and non-redundant. By performing dimensionality reduction on the base ray feature vector, the data dimensionality can be further effectively reduced. In addition, by collecting elevation value matrices corresponding to different granularities, the target character object can be... By displaying the elevation of the surrounding terrain and performing dimensionality reduction on the elevation value matrix to obtain elevation map feature vectors, the dimensionality of the data can be effectively reduced. From the elevation map feature vectors, terrain and topographic information of the 3D scene can be extracted, resulting in a more complete description of the 3D scene. Furthermore, by integrating the ray feature vectors and elevation map feature vectors into the 3D scene features, the 3D scene features become concise, efficient, semantically rich, and include terrain and topographic information, adequately describing the 3D scene. By training a machine learning model using the integrated 3D scene features as sample features, the machine learning model's ability to perceive 3D scenes is enhanced, accelerating the development of the machine learning model and reducing the training cost of the machine learning task. [Brief explanation of the drawing]

[0014] [Figure 1] This is a schematic diagram of the architecture of the image data control system in the embodiment of this application. [Figure 2] This is a flowchart of one embodiment of the feature extraction method for a 3D scene in the embodiments of this application. [Figure 3] This is a flowchart of another embodiment of the 3D scene feature extraction method in the embodiments of this application. [Figure 4]It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 5] It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 6] It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 7] It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 8] It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 9] It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 10] It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [[ID=1)) [Figure 11] It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 12] It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 13] It is a flowchart of another embodiment of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 14] It is a schematic flowchart of the principle of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 15] It is a schematic diagram of the principle of feature processing of the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 16] It is a schematic diagram of the effect of a set of frustum rays emitted by the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 17] It is a schematic diagram of the envelope surface of a set of frustum rays emitted by the method for extracting features of a 3D scene in the embodiments of the present application. [Figure 18]This is a schematic diagram of the cross-sectional distribution where the horizontal distance of the emission of a set of cone rays ejected by the feature extraction method for a 3D scene according to an embodiment of this application reaches a distance threshold. [Figure 19] This is a schematic diagram of an embodiment of a 3D scene feature extraction device according to the present invention. [Figure 20] This is a schematic diagram of an embodiment of a computer device according to the present application. [Modes for carrying out the invention]

[0015] The terms “first,” “second,” “third,” “fourth,” etc. (if any) in the specification, claims, and drawings of this application are not used to indicate a specific order or sequence, but to distinguish similar objects. Where appropriate, data used in this manner may be interchangeable, and embodiments of this application described herein may be carried out, for example, in an order other than those illustrated or described herein. Furthermore, the terms “including,” “corresponding,” and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus including a series of steps or units is not limited to these steps or units explicitly mentioned, but may include others not explicitly mentioned, or other steps or units inherent to these processes, methods, products, or apparatus.

[0016] To facilitate understanding, we will first define some of the terms or concepts to which the embodiments of this application relate. 1. Machine Learning: Machine learning is a method for automatically analyzing data features to acquire rules and using those rules to make predictions and inferences about unknown data. It is a means of realizing artificial intelligence. Machine learning includes supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. 2. Game AI: Game AI is a program or character that combines artificial intelligence-related technologies and is introduced into a game to enrich the way the game is played and improve the player's gaming experience. 3. 3D Game Scene: A 3D game scene is the scene in a 3D game where a character is located, and it is a collection of all spatial information, including the surrounding environment and various objects. 4. Feature Extraction: Feature extraction is the process of constructing information-rich and non-redundant derived values ​​(called feature values) based on an initial set of source information. Feature extraction is a process of simplification and dimensionality reduction of the original information, which makes induction and learning easier for machine learning models. 5. Visual Image Features: Visual image features are rendered images of the field of view in which the character is currently located in the 3D scene. Rendered images generally contain three channels, RGB, and their visual effects essentially match the user-visible 3D scene screen. After acquiring visual image features, machine learning models (e.g., game AI) typically encode the image features using convolution or attention mechanisms, and the encoded parameters are generally learnable. 6. Depth Map Features: Depth map features are characterized by the depth values ​​of pixel points in the field of view where the character is currently located in the 3D scene. It is necessary to calculate the distance of each object pixel point in the field of view where the character is currently located. Since the depth value matrix can be considered as an image, the processing method of depth map features by machine learning models (e.g., game AI) is similar to that of visual image features and is generally encoded by convolution or attention mechanisms.

[0017] Herein, with regard to relevant data such as object attribute information in the specific embodiments of this application, when the above embodiments of this application are applied to a specific product or technology, it is necessary to obtain the user's permission or consent, and the collection, use, and processing of relevant data must comply with the relevant laws, regulations, and standards of the relevant country or region.

[0018] The 3D scene feature extraction method provided in this application is applicable to various scenes, including but not limited to artificial intelligence, cloud technology, maps, and smart transportation. By integrating ray feature vectors and altitude map feature vectors into 3D scene features, it can be applied to training machine learning models and is applicable to scenes with 3D scenes, such as first-person shooter games, competitive games, and smart transportation games.

[0019] The 3D scene feature extraction method submitted in this application is applicable to the image data control system shown in Figure 1. Referring to Figure 1, which is a schematic diagram of the architecture of the image data control system in an embodiment of this application, as shown in Figure 1, the server acquires a 3D scene image provided by a terminal device, emits a set of cone rays from a target character object, and if there is a cone ray that hits a target object in the set of cone rays, it returns the object attribute information of the target object, performs a vector transformation on the received object attribute information of each target object to obtain a basic ray feature vector corresponding to the cone ray, and performs a feature dimensionality reduction process on the basic ray feature vector to obtain a ray feature vector. Furthermore, based on different granularities, using the location where the target character object is located as the collection center, elevation value matrices corresponding to each granularity are collected, and feature dimensionality reduction is performed on the elevation value matrices corresponding to each granularity to obtain an elevation map feature vector, and then integrates the ray feature vector and the elevation map feature vector into a 3D scene feature corresponding to the 3D scene screen.

[0020] According to the above method, object attribute information of the hit object can be quickly obtained from the cone rays of the hit object, and furthermore, the corresponding base ray feature vector is transformed and obtained based on the object attribute information of the hit object. As a result, the base ray feature vector contains the attribute information of the object in the 3D scene, and since the number of cone rays emitted is far lower than the number of pixels such as visual image features and depth map features, the obtained base ray feature vector is concise and non-redundant. Furthermore, by performing dimensionality reduction processing on the base ray feature vector, the dimensionality of the data can be reduced even more effectively. Furthermore, by collecting elevation value matrices corresponding to different granularities, the elevation of the terrain around the target character object is displayed. By performing dimensionality reduction on the elevation value matrices and obtaining elevation map feature vectors, the dimensionality of the data is effectively reduced, and terrain and topographic information of the 3D scene can be extracted from the elevation map feature vectors. This results in a more complete description of the 3D scene. Moreover, by integrating the ray feature vectors and elevation map feature vectors into the 3D scene features, the 3D scene features become concise and efficient, semantically rich, and include terrain and topographic information, fully describing the 3D scene. By training a machine learning model using the integrated 3D scene features as sample features, the machine learning model's ability to perceive 3D scenes is enhanced, accelerating the development of the machine learning model and reducing the training cost of the machine learning task.

[0021] Here, Figure 1 shows only one terminal device, but in a real-world scenario, more types of terminal devices may participate in the data processing process. Terminal devices include, but are not limited to, mobile phones, computers, smart voice interaction devices, smart home appliances, and in-car terminals. The specific number and types will be determined according to the real-world scenario, and the details are not limited here. Similarly, Figure 1 shows one server, but in a real-world scenario, multiple servers may participate. In particular, in a multi-model training and interaction scenario, the number of servers will be determined according to the real-world scenario, and the details are not limited here.

[0022] In this embodiment, the server may be an independent physical server, a server cluster consisting of multiple physical servers, or a distributed system. Furthermore, it may be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), big data, and artificial intelligence platforms. Terminal devices and servers may be connected directly or indirectly by wired or wireless communication, and terminal devices and servers may be connected to form a blockchain network; however, this application is not limited to these.

[0023] Combining the above introduction, the 3D scene feature extraction method of this application will now be introduced, with reference to Figure 2. One embodiment of the 3D scene feature extraction method in this application includes the following steps: In step S101, a pair of cone-shaped projections are emitted from the top of the target character object onto the 3D scene screen.

[0024] Here, the 3D scene feature extraction method provided in the embodiment of this application extracts 3D scene features corresponding to a 3D game scene screen, and in actual application, the 3D scene features corresponding to a 3D game scene screen may be used to train a game AI.

[0025] Here, game AI can not only provide a better game experience but also support game development. For example, it can predict game win rates, design more rational challenges and numerical values, and even design more interesting game characters. Therefore, designing smarter, more comprehensive game AI is a crucial part of game development. In 3D scene games, the design of game AI requires that the game AI effectively perceive the game scene information in its current state. Because the information dimension of the original 3D scene is too high and the content is complex, it is difficult to use it directly as input for machine learning models. Therefore, it is necessary to extract features from the original 3D scene information. However, 3D scenes in games generally have characteristics such as being unstructured, having a wide variety of objects, complex object shapes, and being difficult to model as data information.

[0026] Therefore, in order to better support game AI using machine learning based on the extracted 3D scene features, it is necessary to maintain the accuracy and completeness of the described 3D scene information, as well as to perform dimensionality reduction and simplification on the 3D scene information.

[0027] To better extract features from a 3D scene, as shown in Figure 14, in this embodiment, tools such as a cone ray tool and an altitude map tool are first constructed. The cone ray tool (cone ray tool in Figure 14) can be developed based on the ray function built into the game engine. After each ray is emitted, it returns the hit point and the type and material information of the hit object. The altitude map tool (altitude map tool in Figure 14) first derives an altitude description file for the entire game environment (3D FPS game in Figure 16) (for example, altitude description files for various terrains such as rainforest, snow, or island, and terrains such as grass or sand), and can provide a corresponding altitude value search interface. Here, the tool construction process only needs to be run once after the game environment is determined, and tools can be duplicated within the same game environment.

[0028] Furthermore, after constructing the cone ray tool and advanced map tool, the process proceeds to the feature collection stage. For example, during a trial run of the game (the 3D FPS game in Figure 16), the cone ray tool (the cone ray tool in Figure 14) is used to emit a set of cone rays from the top of the game character object (i.e., the target character object) (the head of the game character in the 3D FPS game in Figure 16) according to the distribution of cones, for each frame of the 3D game scene screen. This allows for real-time acquisition of ray results. Here, since it is necessary to dynamically calculate the hit information of the rays, feature collection of the cone rays must also be performed in real time as the game progresses. The number and distribution of rays in a set of cone rays can be set according to the actual needs and are not specifically limited here.

[0029] In practical application, the method provided by the embodiments of this application is applicable to feature extraction of 3D game scene screens, and may also be applied to other scenes containing 3D scene screens to extract features of 3D scene screens in those other scenes, and is not limited to the application scenes of the method provided by the embodiments of this application.

[0030] In step S102, if the visual cone rays hit an object, the object attribute information of that object is returned.

[0031] In the embodiments of this application, after a set of cone rays is emitted from a target character object, if there is a cone ray in the set of cone rays that hits an object in a 3D scene (in the embodiments of this application, an object hit by a cone ray is called a hit object), the cone ray of the hit object returns object attribute information corresponding to the point of impact, such as the position information of the hit object, the type information of the hit object, and the material information of the hit object.

[0032] Specifically, after entering the feature collection stage, for example, according to the distribution of visual cones, a pair of visual cone rays are emitted from the top of the game character object (the head of the game character in the 3D FPS game in Figure 16) (in the 3D FPS game in Figure 16, a pair of cone-shaped visual cone rays are emitted from the head of the game character), and the ray results are acquired in real time. If there is a visual cone ray in the pair of visual cone rays that hit an object in the 3D game scene, the visual cone ray of the hit object returns object attribute information corresponding to the point of hit to the server. For example, in the 3D FPS game in Figure 16, if a pair of cone-shaped visual cone rays are emitted from the head of the game character and one of the visual cone rays hits an obstacle (the container in Figure 16), the visual cone ray that hit the obstacle returns the position information of the hit object corresponding to the point of hit (for example, the point of hit where it hit the container's casing, in the 3D game in Figure 16). The system returns object attribute information to the server, such as the current on-screen coordinates in an FPS game, the type of object hit (e.g., container casing), and material information (e.g., high-strength steel).

[0033] In step S103, a vector transformation is performed on the object attribute information of each received hit object to obtain the basic trajectory feature vector.

[0034] In the embodiment of this application, after the cone rays return object attribute information of the hit object, feature processing is performed on the received object attribute information, and a vector transformation is performed on the object attribute information of all received hit objects to obtain a basic ray feature vector, thereby enabling better acquisition of 3D scene features based on the basic ray feature vector thereafter.

[0035] Specifically, as shown in Figure 14, after returning object attribute information of the hit object, a feature processing step is performed. In the feature processing step, encoding, i.e., vector transformation, is performed on the data acquired in the feature collection step, i.e., all received object attribute information of the hit objects, to obtain a basic ray feature vector. This basic ray feature vector is then input into the subsequent machine learning model, and the machine learning game AI performs inductive learning on the extracted 3D scene features. Here, generally, the parameters of the encoding process can also be learned and optimized.

[0036] In step S104, a feature dimensionality reduction process is performed on the basic ray feature vector to obtain a new ray feature vector.

[0037] In the embodiments of this application, after obtaining the basic ray feature vector, the machine learning model inputs the feature vector to the machine learning model. In order to better perform inductive learning on the 3D scene features extracted by the machine learning model, the machine learning model performs a feature dimensionality reduction process on the basic ray feature vector to obtain a ray feature vector with lower dimensionality, thereby achieving dimensionality reduction and simplification of the feature vector. This allows for better dimensionality reduction and simplification of the 3D scene features extracted based on the ray feature vector thereafter.

[0038] Specifically, as shown in Figure 14, after obtaining the basic ray feature vector, a feature dimensionality reduction process is performed on the basic ray feature vector via a neural network (as shown in Figure 14, a high-density network is used for the ray vector) to obtain a ray feature vector, specifically a one-dimensional ray feature vector.

[0039] In step S105, elevation value matrices corresponding to each granularity are collected, with the location of the target character object as the collection center, based on different granularities.

[0040] In the embodiments of this application, after constructing the cone radiation tool and the elevation map tool, elevation value matrices corresponding to each granularity are collected in all directions around the current character coordinates of the target character object, based on different granularities. This allows for better acquisition of terrain and topographic information in a 3D environment based on the elevation value matrices, thereby better supporting machine learning models in sensing the 3D scene.

[0041] Specifically, after constructing the cone ray tool and the elevation map tool, the process proceeds to the feature collection stage. For example, during a trial run of a game (the 3D FPS game in Figure 16), the elevation map tool (the elevation map tool in Figure 14) is used for each frame of the 3D game scene screen to obtain the current character coordinates of the game character object in the 3D game scene screen (the current orientation and current screen coordinates of the game character in the 3D FPS game in Figure 16). An elevation value search interface is constructed in the elevation map tool to search for the terrain map corresponding to the current character coordinates. Furthermore, using the current character coordinates of the game character object as the collection center, grid diagrams corresponding to each granularity are collected in all directions around the collection center based on different granularities, for example, granularity a. Finally, based on the elevation value search interface, an elevation value matrix corresponding to each grid diagram is obtained.

[0042] Here, since terrain conditions are generally stable and do not change with gameplay, the acquisition of elevation matrix values ​​and extraction of elevation map features can be performed offline, further reducing computational overhead and, to some extent, lowering the development cost of the game AI.

[0043] In step S106, feature dimensionality reduction is performed on the elevation value matrix corresponding to each granularity to obtain elevation map feature vectors.

[0044] In the embodiments of this application, after obtaining the elevation matrix corresponding to each granularity, the machine learning model is input to the machine learning model. In order to enable the machine learning model to perform inductive learning on the extracted 3D scene features using the feature vectors, the machine learning model performs a feature dimensionality reduction process on the elevation matrix corresponding to each granularity to obtain an elevation map feature vector. By inputting this elevation map feature vector into the machine learning model, the machine learning model can then perform inductive learning on the extracted 3D scene features. Generally, the parameters of the encoding process can also be learned and optimized.

[0045] Specifically, as shown in Figure 14, after obtaining the elevation value matrix corresponding to each granularity, a feature dimensionality reduction process is performed on the elevation value matrix corresponding to all granularities via a neural network (a convolutional network is used as shown in Figure 14) to obtain an elevation map feature vector, specifically a one-dimensional elevation feature vector.

[0046] In step S107, the ray feature vector and the elevation map feature vector are integrated into the 3D scene feature corresponding to the 3D scene screen.

[0047] In the embodiments of this application, after obtaining the ray feature vector and the elevation map feature vector, the ray feature vector and the elevation map feature vector are integrated into the 3D scene features corresponding to the 3D scene screen, thereby enabling the machine learning model to perform inductive learning on the extracted 3D scene features.

[0048] Specifically, as shown in Figure 14, after obtaining the ray feature vector and the altitude map feature vector, the ray feature vector and the altitude map feature vector are stitched together to obtain 3D game scene features corresponding to the 3D game scene screen. This allows the machine learning-based game AI to perform inductive learning.

[0049] Embodiments of this application provide a feature extraction method for a 3D scene. According to the above method, object attribute information of the hit object can be quickly obtained from the cone rays of the hit object, and furthermore, a corresponding basic ray feature vector is obtained by transforming based on the object attribute information of the hit object. As a result, the basic ray feature vector contains the attribute information of the object in the 3D scene, and since the number of emitted cone rays is far lower than the number of pixel points such as visual image features and depth map features, the obtained basic ray feature vector is concise and non-redundant. Furthermore, by performing dimensionality reduction processing on the basic ray feature vector, the dimensionality of the data can be further effectively reduced, and by collecting elevation value matrices corresponding to different granularities, the current target... By displaying the elevation of the terrain surrounding a character object and performing dimensionality reduction on the elevation value matrix to obtain elevation map feature vectors, the dimensionality of the data is effectively reduced. From the elevation map feature vectors, terrain and topographic information of the 3D scene can be extracted, resulting in a more complete description of the 3D scene. Furthermore, by integrating the ray feature vectors and elevation map feature vectors into the 3D scene features, the 3D scene features become concise, efficient, semantically rich, and include terrain and topographic information, fully describing the 3D scene. By training a machine learning model using the integrated 3D scene features as sample features, the machine learning model's ability to perceive 3D scenes is enhanced, accelerating the development of the machine learning model and reducing the training cost of the machine learning task.

[0050] Optionally, in another preferred embodiment of the 3D scene feature extraction method provided by the embodiment of this application, based on the embodiment corresponding to Figure 2 above, the object attribute information of the hit object includes the position information, type information, and material information of the hit object, and as shown in Figure 3, step S103, which performs a vector transformation on the object attribute information of each received hit object to obtain a basic ray feature vector, includes the following steps: In step S301, vector transformation is performed on the object attribute information of each hit object, including the position information, type information, and material information of the hit object, to obtain an information vector corresponding to the cone ray that hit the hit object; In step S302, the information vectors corresponding to each of the visual cone rays are integrated into the basic ray feature vector.

[0051] In the embodiment of this application, after returning object attribute information of the hit object, feature processing is performed on the received object attribute information, that is, vector transformation is performed on the position information, type information, and material information of the hit object to obtain an information vector corresponding to the cone rays that hit the hit object, and the information vectors corresponding to the cone rays of all hit objects are integrated into a basic ray feature vector, thereby enabling better acquisition of 3D scene features based on the basic ray feature vector thereafter.

[0052] Specifically, as shown in Figure 14, after returning object attribute information of the hit object, a feature processing step is performed. In the feature processing step, the data acquired in the feature collection step, namely the position information of the hit object, the type information of the hit object, and the material information corresponding to the cone rays of the hit object, are encoded, i.e., vector transformed, to obtain an information vector corresponding to the cone rays. Then, by stitching together the information vectors corresponding to the cone rays of all hit objects, a basic ray feature vector is obtained. Subsequently, the basic ray feature vector is input into a machine learning model, and inductive learning is performed on the 3D scene features extracted by the machine learning game AI.

[0053] Thus, according to the above method, vector transformation is performed on the positional information, type information, and material information of the object hit by the cone rays to obtain an information vector corresponding to the cone rays. Furthermore, the information vectors corresponding to the cone rays of each hit object are integrated to obtain a basic ray feature vector. As a result, the obtained basic ray feature vector can more accurately describe the objects present in the 3D scene, that is, it achieves a sufficient description of the 3D scene.

[0054] Optionally, in another preferred embodiment of the 3D scene feature extraction method provided by the embodiment of this application, based on the embodiment corresponding to Figure 3 above, as shown in Figure 4, step S301 includes steps S401 to S403, and step S302 includes step S404, by performing a vector transformation on the position information of the hit object, the type information of the hit object, and the material information to obtain an information vector corresponding to the cone rays that hit the hit object; In step S401, the position information of the hit object is subjected to vector normalization to obtain a position vector; In step S402, feature encoding is performed on the type information and material information of the hit object to obtain an encoded vector; In step S403, the position vector and the encoded vector are stitched together to obtain an information vector corresponding to the cone ray that hit the object in question; In step S404, the information vectors corresponding to each of the visual cone rays are sequentially stitched together to obtain the basic ray feature vector.

[0055] In the embodiment of this application, after returning object attribute information of the hit object, feature processing is performed on the received object attribute information, that is, vector normalization processing is performed on the position information of the hit object to obtain a position vector, feature encoding is performed on the type information and material information of the hit object to obtain an encoded vector, further, the position vector and the encoded vector are stitched together to obtain an information vector corresponding to the cone rays that hit the hit object, and then the information vectors corresponding to the cone rays of all hit objects are stitched together sequentially to obtain a ray feature vector.

[0056] Specifically, as shown in Figure 14, after returning object attribute information of the hit object, a feature processing step is performed, in which the data acquired in the feature collection step, i.e., the position information of the hit object, is subjected to vector normalization processing. Specifically, first, a series of operations such as horizontal translation, zooming, and rotation of 3D coordinates are performed on the coordinate information of the hit point of the cone ray to obtain a position vector corresponding to the coordinate information, or normalization processing is performed on the coordinate information, for example, by performing a vector transformation on the coordinate information based on a linear function (matrix multiplication, for example, matrix composed of base vectors * coordinate value = vector) to obtain a position vector corresponding to the coordinate information. Other vector normalization processing methods may also be adopted, and are not specifically limited here.

[0057] Furthermore, feature encoding may be performed on the type information and material information of the hit object, specifically employing one-hot encoding for the type information and material information respectively, and then embedding processing may be performed on the encoded data to obtain an embedding vector corresponding to the type information (object type embedding in Figure 15) and an embedding vector corresponding to the material information (object material embedding in Figure 15). Then, the embedding vector corresponding to the type information (object type embedding in Figure 15) and the embedding vector corresponding to the material information (object material embedding in Figure 15) may be stitched into the encoded vector, or other feature encoding methods may be employed for the type information and material information, and there are no specific limitations.

[0058] Furthermore, as shown in Figure 15, after obtaining a position vector corresponding to the position information of the hit object (the ray hit point coordinate vector in Figure 15), and encoded vectors corresponding to the type information and material information of the hit object (the object type embedding and object material embedding in Figure 15), the position vector and encoded vector are stitched together (for example, stitched together in the order of the ray hit point coordinate vector, object type embedding, and object material embedding in Figure 15) to obtain an information vector (the single ray information in Figure 15) for a single cone ray (i.e., the cone ray that hit the hit object). Then, by stitching together the information vectors corresponding to the cone rays of all hit objects according to a predetermined fixed order (for example, stitching together a set of cone rays in the right diagram of Figure 17 in a counterclockwise direction, from outside to inside), a basic ray feature vector is obtained. Subsequently, the basic ray feature vector is input into a machine learning model, which allows the machine learning model to perform inductive learning on the extracted 3D scene features.

[0059] Thus, according to the above method, a position vector is generated based on the position information of the object hit by the cone rays, an encoded vector is generated based on the type information and material information of the hit object, and then the position vector and the encoded vector are stitched together to obtain an information vector corresponding to the cone rays, and then the information vectors corresponding to the cone rays of each hit object are stitched together sequentially to obtain a basic ray feature vector, thereby ensuring that the obtained basic ray feature vector more accurately describes the objects present in the 3D scene, that is, it achieves a sufficient description of the 3D scene.

[0060] Optionally, in another preferred embodiment of the 3D scene feature extraction method provided by the embodiment of this application, based on the embodiment corresponding to Figure 2 above, step S105, which collects elevation value matrices corresponding to each granularity with the location of the target character object as the collection center, based on different granularities, as shown in Figure 5, includes the following steps:

[0061] In step S501, based on different granularities, the location of the target character object is used as the collection center, and N*N grids corresponding to each granularity are collected in all four directions around it, where N is an integer greater than or equal to 1; In step S502, an elevation value matrix corresponding to each granularity is generated based on an N*N grid corresponding to each granularity.

[0062] In the embodiments of this application, after constructing a cone radiation tool and an elevation map tool, N*N grids corresponding to each granularity are collected in all directions around the current character coordinates of the target character object, based on different granularities. Then, based on the N*N grids corresponding to each granularity, an elevation value matrix corresponding to each granularity is generated. This allows for better acquisition of terrain and topographic information in a 3D environment based on the elevation value matrix, and furthermore, better supports machine learning models in sensing the 3D scene.

[0063] Specifically, after constructing the cone ray tool and the elevation map tool, the process proceeds to the feature collection stage. For example, during a trial run of the game (3D FPS game in Figure 16), the elevation map tool (elevation map tool in Figure 14) is used on the 3D game scene screen for each frame to determine the current character coordinates of the target character object in the 3D game scene screen (3D in Figure 16). The system obtains the current orientation and on-screen coordinates of the game character in an FPS game, constructs an altitude value search interface in the altitude map tool to search for the terrain map (game terrain in Figure 18) corresponding to the current character coordinates, and then, using the current character coordinates of the target character object as the collection center, collects grid diagrams corresponding to each granularity (e.g., granularity a) on all four sides of the collection center (for example, collecting 4*4 grids with side length a as shown in Figure 18), and then obtains altitude values ​​corresponding to the N*N grids corresponding to each granularity (for example, collecting 4*4 grids with side length a as shown in Figure 18) based on the altitude value search interface to generate an altitude value matrix (for example, the 4*4 altitude value matrix in Figure 18).

[0064] Thus, according to the above method, the current character coordinates of the target character object are used as the collection center, and N*N grids corresponding to different granularities are collected in all four directions around it. Based on the N*N grids corresponding to each granularity, an elevation value matrix corresponding to each granularity is generated, and the elevation of the terrain around the current target character object is displayed using the elevation value matrix corresponding to different granularities, thereby extracting terrain and ground feature information of the 3D scene, and thus providing a more complete description of the 3D scene.

[0065] Optionally, in another preferred embodiment of the 3D scene feature extraction method provided by the embodiments of this application, based on the embodiment corresponding to Figure 5 above, as shown in Figure 6, step S501, which collects N*N grids corresponding to each granularity in all four directions around the location where the target character object is located, based on different granularities, includes steps S601 to S602, and step S502 includes step S603; In step S601, different granularities are defined as the unit length of the grid to be collected, which has different sizes; In step S602, N*N grids corresponding to the unit lengths of various grids to be collected are collected in all four directions around the location where the target character object is located, with the collection center being the collection center, according to the unit lengths of the grids to be collected, which have different sizes. In step S603, the elevation value corresponding to the center point of each grid in the N*N grid is obtained, and an elevation value matrix corresponding to the granularity of the N*N grid is generated based on the elevation value.

[0066] In the embodiments of this application, after constructing the cone radiation tool and the elevation map tool, first, N*N grids corresponding to the unit length of each target grid are collected in all directions around the current character coordinates of the target character object, with different granularities as the unit length of the target grid having different sizes, according to the unit length of the target grid having different sizes, and the elevation value corresponding to each target grid is obtained, and an elevation value matrix corresponding to each granularity is generated based on the elevation value, thereby enabling better acquisition of terrain and ground surface information in the 3D environment based on the elevation value matrix thereafter, and further providing better support to the machine learning model in sensing the 3D scene.

[0067] Specifically, after constructing the cone ray tool and the elevation map tool, the process proceeds to the feature collection stage. For example, during a trial run of a game (the 3D FPS game in Figure 16), the elevation map tool (the elevation map tool in Figure 14) is used for each frame of the 3D game scene screen to obtain the current character coordinates of the target character object in the 3D game scene screen (the current orientation of the game character in the 3D FPS game in Figure 16, and the current coordinates on the screen). An elevation value search interface is then constructed in the elevation map tool to search for the terrain map (the game terrain in Figure 18) corresponding to the current character coordinates.

[0068] Furthermore, different granularities (e.g., granularity a in Figure 18) are used as the unit length of the grids to be collected, which have different sizes (e.g., unit length a of the grids to be collected in Figure 18). Additionally, the current character coordinates of the target character object are used as the collection center, and grid diagrams corresponding to each granularity are collected on all four sides around the collection center according to the unit length of the grids to be collected, which have different sizes (e.g., unit length a of the grids to be collected in Figure 18). For example, a 4x4 grid with side length a is collected as shown in Figure 18.

[0069] Furthermore, based on the altitude value search interface, altitude values ​​corresponding to the center points of each grid in the N*N grid corresponding to each granularity are obtained (for example, collecting the center points of each grid in a 4*4 grid where the side length of the grid is a, as shown in Figure 18), and an altitude value matrix is ​​generated according to the correspondence between the center points of each grid in the N*N grid corresponding to each granularity and the position of each element in the N*N altitude value matrix (for example, the altitude values ​​corresponding to the center points of each grid in the N*N grid corresponding to each granularity are written to the corresponding 4*4 altitude value matrix to obtain the 4*4 altitude value matrix in Figure 18).

[0070] Thus, according to the above method, the unit length of the grid to be collected, which has different sizes for different granularities, is determined, and based on this, N*N grids are collected. Furthermore, the elevation value of the center point of each grid in the N*N grid is written to the corresponding N*N matrix to obtain an elevation value matrix corresponding to that granularity, thereby ensuring that the determined elevation value matrix more accurately reflects the elevation of the terrain around the current target character object.

[0071] Optionally, in another preferred embodiment of the 3D scene feature extraction method provided by the embodiment of this application, based on the embodiment corresponding to Figure 2 above, step S106, which involves performing a feature dimensionality reduction process on the elevation value matrix corresponding to each granularity to obtain an elevation map feature vector, as shown in Figure 7, includes the following steps:

[0072] In step S701, stitching and tensor transformation are performed on the elevation value matrix corresponding to each granularity to obtain the elevation map feature tensor; In step S702, a feature dimensionality reduction process is performed on the advanced map feature tensor to obtain the advanced map feature vector.

[0073] In the embodiments of this application, after obtaining the elevation matrix corresponding to each granularity, a feature dimensionality reduction process is performed on the elevation matrix corresponding to all granularities to obtain an elevation map feature vector (for example, a one-dimensional elevation map feature vector) in order to enable the machine learning model to perform inductive learning on the 3D scene features extracted by the machine learning model using the feature vectors input to the machine learning model. Here, generally, the parameters of the encoding process can also be learned and optimized.

[0074] Specifically, as shown in Figure 15, after obtaining the elevation matrix corresponding to each granularity (for example, the elevation matrix corresponding to three different granularities as shown in Figure 15), stitching and tensor transformation processing are performed on the elevation matrix corresponding to all granularities. For example, if there are elevation matrix corresponding to three different granularities a, b, and c, the three elevation matrix are 4*4 elevation matrix with the same number of matrices. Therefore, there is no need to zero-padding the matrices to obtain matrices with the same number of matrices. Instead, the matrix after stitching or merging can be obtained by stitching or merging the array horizontally or vertically using the cat function, or other stitching or merging methods may be adopted, and there are no specific limitations here. Then, a tensor transformation is performed on the stitched or merged matrix. For example, first, the stitched or merged matrix is ​​converted to numpy type data using a Numpy array, and then the converted numpy type data is substituted into the function expression torch.from_numpy(np_array) using the torch.from_numpy function to generate an advanced map feature tensor (for example, an advanced map feature tensor containing different granularities as shown in Figure 15). Furthermore, other tensor transformation methods, such as the torch.tensor function, may also be used, and are not specifically limited here.

[0075] Furthermore, after obtaining the high-level map feature tensor (for example, a high-level map feature tensor containing different granularities, as shown in Figure 15), a feature dimensionality reduction process is performed on the high-level map feature tensor (for example, a high-level map feature tensor containing different granularities, as shown in Figure 15) via a neural network (using a convolutional network CNN, as shown in Figure 14) to obtain the high-level map feature vector.

[0076] In this embodiment, to extend the feature detection range, elevation matrices corresponding to different receptive fields are obtained by selecting collection granularity of different sizes, and the topography and features of the 3D scene are more adequately described based on the elevation map tensor formed from the elevation matrices corresponding to different receptive fields.

[0077] Step S104, optionally, involves performing a feature dimensionality reduction process on the basic ray feature vector to obtain a ray feature vector based on the embodiment corresponding to Figure 2 above. Specifically, this step includes performing a feature dimensionality reduction process on the basic ray feature vector via a neural network to obtain a ray feature vector. Step S106, which involves performing a feature dimensionality reduction process on the elevation value matrix corresponding to each granularity to obtain an elevation map feature vector, specifically includes performing a feature dimensionality reduction process on the elevation value matrix corresponding to each granularity via a neural network to obtain an elevation map feature vector.

[0078] In this way, the feature dimensionality reduction process for the base ray feature vector and the feature dimensionality reduction process for the elevation value matrix corresponding to each granularity are completed via the corresponding neural network. This ensures that the ray feature vector obtained through the feature dimensionality reduction process better retains the useful information from the original base ray feature vector, and that the elevation map feature vector obtained through the feature dimensionality reduction process better retains the useful information from the elevation value matrix corresponding to each granularity. As a result, the ray feature vector and elevation map feature vector more accurately describe the corresponding 3D scene.

[0079] Optionally, in another preferred embodiment of the 3D scene feature extraction method provided by the embodiment of this application, based on the embodiment corresponding to Figure 2 above, the neural network that processes the basic ray feature vector and the elevation value matrix (specifically, the elevation map feature tensor) corresponding to each granularity, as shown in Figure 8, may be the same convolutional neural network, and step S104, which performs a feature dimensionality reduction process on the basic ray feature vector to obtain the ray feature vector, includes step S801, and step S106, which performs a feature dimensionality reduction process on the elevation value matrix corresponding to each granularity to obtain the elevation map feature vector, includes step S802; In step S801, a feature dimensionality reduction process is performed on the underlying ray feature vector via a convolutional neural network to obtain a ray feature vector; In step S802, a feature dimensionality reduction process is performed on the elevation value matrix corresponding to each granularity via a convolutional neural network to obtain elevation map feature vectors.

[0080] In the embodiments of this application, after obtaining the basic ray feature vector and the elevation map feature tensor (generated based on the elevation value matrix corresponding to each granularity), the neural network used when performing dimensionality reduction on the basic ray feature vector and the elevation map feature tensor, respectively, may be the same convolutional neural network. That is, feature dimensionality reduction is performed on the basic ray feature vector via a convolutional neural network to obtain a one-dimensional ray feature vector, and feature dimensionality reduction is performed on the elevation map feature tensor via a convolutional neural network to obtain a one-dimensional elevation map feature vector. This reduces the construction and use of different neural network frameworks, reduces computational complexity to some extent, and improves the efficiency of acquiring 3D game scene features.

[0081] Specifically, after obtaining the base ray feature vector and the altitude map feature tensor, dimensionality reduction processing is performed on the base ray feature vector and the altitude map feature tensor, respectively. Subsequently, based on the dimensionality-reduced ray feature vector and the altitude map feature vector, dimensionality-reduced and simplified 3D scene features can be better obtained. Therefore, the neural network used when performing dimensionality reduction processing on the base ray feature vector and the altitude map feature tensor may be the same convolutional neural network. That is, feature dimensionality reduction processing is performed on the base ray feature vector via the convolutional neural network. Since the base ray feature vector is a feature vector obtained by performing feature normalization and stitching according to a fixed order, a 1D ray feature vector can be obtained by performing convolution on the base ray feature vector using 1xN convolutional cores in the convolutional neural network based on the magnitude of the feature values. Feature dimensionality reduction processing on the altitude map feature tensor via the convolutional neural network can be performed using a normal convolution operation method, that is, by processing through convolutional layers, pooling layers, and fully connected layers in sequence to obtain a 1D altitude map feature vector.

[0082] Optionally, in another preferred embodiment of the 3D scene feature extraction method provided by the embodiments of this application, based on the embodiment corresponding to Figure 2 above, as shown in Figure 9, the neural network includes a first neural network and a second neural network, wherein the first and second neural networks are different neural networks, and step S104 includes step S901 to obtain a ray feature vector by performing a feature dimensionality reduction process on the basic ray feature vector, and step S106 includes step S902 to obtain an altitude map feature vector by performing a feature dimensionality reduction process on the altitude value matrix corresponding to each granularity; In step S901, a feature dimensionality reduction process is performed on the underlying ray feature vector via the first neural network to obtain the ray feature vector; In step S902, a feature dimensionality reduction process is performed on the elevation value matrix corresponding to each granularity via a second neural network to obtain elevation map feature vectors.

[0083] In the embodiments of this application, after obtaining the basic ray feature vector and the elevation map feature tensor (generated based on the elevation value matrix corresponding to each granularity), the neural networks used when performing dimensionality reduction on the basic ray feature vector and the elevation map feature tensor, respectively, may be different neural networks. That is, feature dimensionality reduction processing is performed on the basic ray feature vector via a first neural network to obtain the ray feature vector, and feature dimensionality reduction processing is performed on the elevation value matrix corresponding to all granularities via a second neural network to obtain the elevation map feature vector. By performing dimensionality reduction processing on different types of feature vectors via different neural networks, the corresponding dimensionality reduction features can be obtained better and with greater accuracy, thereby improving the accuracy of obtaining 3D scene features.

[0084] The first neural network and the second neural network are different neural networks. Specifically, the first neural network may be a high-density network, such as the DNN neural network shown in Figure 15, and the second neural network may be a convolutional neural network, such as the CNN neural network shown in Figure 15.

[0085] Specifically, as shown in Figure 15, after obtaining the basic ray feature vector and the altitude map feature tensor, dimensionality reduction processing is performed on the basic ray feature vector and the altitude map feature tensor, respectively. Subsequently, based on the dimensionality-reduced ray feature vector and the altitude map feature vector, dimensionality-reduced and simplified 3D scene features can be better obtained. Therefore, the neural networks used when performing dimensionality reduction processing on the basic ray feature vector and the altitude map feature tensor may be different neural networks. Since the obtained basic ray feature vector is non-image data, and convolutional neural networks can more easily learn information about adjacent positions in image data, a high-density network, i.e., a first neural network, for example the DNN neural network in Figure 15, may be used to perform feature dimensionality reduction processing on the basic ray feature vector, which does not necessarily contain adjacent position information. Data related to atomic positions in the basic ray feature vector may be analyzed using a multilayer perceptron MLP, and a feedforward algorithm may be adopted for the sensed data to obtain a 1D ray feature vector. Since the advanced map feature tensor belongs to the image data, a convolutional neural network, i.e., a second neural network, which can more easily learn information about adjacent positions in the image data, may be used to perform feature dimensionality reduction on the advanced map feature tensor. In other words, a normal convolutional operation method may be employed, that is, processing through convolutional layers, pooling layers, and fully connected layers in sequence, to obtain a one-dimensional advanced map feature vector.

[0086] Optionally, in another preferred embodiment of the 3D scene feature extraction method provided by the embodiments of this application, based on the embodiment corresponding to Figure 2 above, step S101 of emitting a set of visual cone rays from a target character object includes step S1001, and step S102 includes step S1002, as shown in Figure 10; In step S1001, a set of cone rays are projected from the top of the target character object, simulating the cone field of view; In step S1002, after each of the cone rays emitted by simulating the cone field of view reaches a length threshold, if any cone rays hit an object, the object attribute information of that object is returned.

[0087] In the embodiments of this application, when extracting features from original 3D scene information, 3D scenes generally have characteristics such as being unstructured, having a wide variety of object types, complex object shapes, and being difficult to model as data information. Therefore, in these embodiments, a pair of cone rays are emitted from the top of the target character object by simulating the cone field of view. After each of the cone rays in the pair of cone rays emitted by simulating the cone field of view reaches a length threshold, if any cone rays hit an object, the object attribute information of the hit object is returned. This allows for better sensing of objects placed in the 3D environment and the attribute information of these objects. Subsequently, 3D scene features can be better extracted based on the object attribute information, and machine learning can be better supported based on the extracted 3D scene features.

[0088] Specifically, to better sense objects placed in a 3D environment and the attribute information of these objects, as shown in Figure 14, in this embodiment, first, tools such as a cone ray tool and an elevation map tool are constructed, and after constructing the cone ray tool and elevation map tool, the process proceeds to the feature collection stage, for example, during a trial run of a game (a 3D FPS game in Figure 16), the cone ray tool (the cone ray tool in Figure 14) is used to simulate the cone field of view angle from the top of the target character object (the head of the game character in the 3D FPS game in Figure 16) to the 3D game scene screen of each frame, and a set of cone rays is emitted. That is, as shown in the left figure of Figure 17, a set of cone rays with a conical envelope surface is emitted according to the cone ray direction and cone ray distribution, and the ray results are obtained in real time.

[0089] Furthermore, after each of the cone rays in a set of cone rays emitted by simulating the cone field of view reaches a length threshold (10 m, as shown in the left diagram of Figure 17), if any cone rays hit an object, the object attribute information of the hit object is returned. For example, in the set of cone rays (as shown in the left diagram of Figure 17, a set of cone rays with a conical envelope surface are emitted from the head of a game character), if any cone rays hit an object, the cone rays of the hit object return object attribute information corresponding to the point of impact to the server, for example, in the 3D of Figure 16. In an FPS game, a pair of cone-shaped trajectories are emitted from the head of the game character (as shown in the left diagram of Figure 17, a pair of trajectories with a cone-shaped envelope are emitted from the head of the game character). When the cone field of view is simulated and each of the emitted trajectories reaches a length threshold (10 m, as shown in the left diagram of Figure 17), assuming that one trajectory hits an object at 10 m (the grass in Figure 16), the trajectory that hit the object sends back object attribute information to the server, such as the position information of the hit object corresponding to the point of impact (for example, the current coordinates on the screen in the 3D FPS game in Figure 16, where the point of impact hit the grass), the type information of the hit object (for example, the grass ground), and the material information (for example, soil or soft soil).

[0090] Optionally, in another preferred embodiment of the 3D game scene feature extraction method provided by the embodiments of this application, based on the embodiment corresponding to Figure 10 above, step S1001 includes step S1101, and step S1002 includes step S1102, where a set of cone rays are emitted from the top of a target character object, simulating the cone field of view; In step S1101, p uniformly distributed ray directions are obtained with the top of the target character object as the center, and M ray clusters are emitted for each ray direction. The envelope of the M ray clusters is conical, and each ray cluster contains p visual cone rays. The p visual cone rays of each ray cluster are uniformly distributed in M ​​concentric circles, where p is an integer greater than or equal to 2, and M is an integer greater than or equal to 1. In step S1102, after the p cone rays of each of the M ray clusters have reached a length threshold, if any cone rays have hit an object, the object attribute information of that object is returned.

[0091] In the embodiments of this application, when extracting features from original 3D scene information, 3D scenes generally have characteristics such as being unstructured, having a wide variety of object types, complex object shapes, and being difficult to model as data information. Therefore, in these embodiments, with the top of the target character object as the center, first, p uniformly distributed ray directions are obtained, and then M ray clusters are emitted for each ray direction. After the p visual cone rays of each ray cluster among the M ray clusters reach a length threshold, if there are visual cone rays that hit an object, the object attribute information of the hit object is returned. This allows for better sensing of objects placed in the environment and the attribute information of these objects, and thereafter, 3D scene features can be better extracted based on the object attribute information, and machine learning can be better supported based on the extracted 3D scene features.

[0092] A ray cluster is a group of rays that contains p optical cone rays. The p optical cone rays of one emitted ray cluster are uniformly distributed in a circle, meaning that M emitted ray clusters are uniformly distributed in M ​​concentric circles, and thus the envelope of the M ray clusters is conical.

[0093] Specifically, to better sense objects placed in a 3D environment and the attribute information of these objects, as shown in Figure 14, in this embodiment, first, tools such as the cone ray tool and the elevation map tool are constructed, and after constructing the cone ray tool and the elevation map tool, the process proceeds to the feature collection step, for example, during a trial run of a game (3D FPS game in Figure 16), the cone ray tool (cone ray tool in Figure 14) is used on the 3D game scene screen of each frame to capture the top part of the target character object (3D in Figure 16) Using the head of a game character in an FPS game as the center of the circle, first, obtain p uniformly distributed ray directions, and if both p and M are 5, the 25 projected cone rays (where, if p is 5 and M is 3, 15 projected cone rays are produced, or if p is 5 and M is 7, 35 projected cone rays are produced) are uniformly distributed in the circle, that is, one circle is divided into 0°, 72°, 144°, 216° and 288°, and Assuming the length threshold is 10m in the left diagram of Figure 17, the maximum radius of the concentric circles is 10m. Since the M emitted ray clusters are uniformly distributed across the M concentric circles, the radii of each concentric circle are 2m, 4m, 6m, 8m, and 10m, respectively. Thus, the angle of the emitted ray in each ray cluster can be calculated as 9°, 18°, 27°, 36°, and 45°. Therefore, the direction of each visual cone ray can be determined based on the degree on the circle and the angle of the emitted ray.

[0094] Furthermore, as shown in the right-hand diagram of Figure 17, M ray clusters are emitted for each ray direction. For example, as shown in the right-hand diagram of Figure 17, ray clusters a, b, c, d, and e are emitted for 0°, 72°, 144°, 216°, and 288°, respectively. This allows us to obtain five concentric circles and 25 cone rays uniformly distributed within those concentric circles. When the horizontal distance of the emission of the 25 cone rays reaches a length threshold (10m, as shown in the left-hand diagram of Figure 17), the cross-sectional distribution shown in the right-hand diagram of Figure 17 can be obtained.

[0095] Furthermore, if any cone rays hit an object, the object attribute information of the hit object is returned. For example, in the set of cone rays (as shown in the left diagram of Figure 17, a set of cone rays with a conical envelope surface are emitted from the head of the game character), if any cone rays hit an object, the cone rays that hit the object return object attribute information corresponding to the point of impact to the server. For example, in the 3D of Figure 16... In an FPS game, a pair of cone-shaped cone rays are emitted from the head of the game character (as shown in the left diagram of Figure 17, a pair of cone rays with a conical envelope are emitted from the head of the game character). Assuming that the cone field of view is simulated and each of the emitted cone rays reaches a length threshold (10 m as shown in the left diagram of Figure 17), and one cone ray hits an object at 10 m (the grass in Figure 16), then the cone ray that hit the object returns object attribute information to the server, such as the position information of the hit object corresponding to the point of impact (e.g., the point of impact on the grass, the current coordinates on the screen in the 3D FPS game in Figure 16), the type information of the hit object (e.g., the grass ground), and the material information (e.g., soil or soft soil).

[0096] Here, after obtaining an information vector for a single cone ray (single ray information in Figure 15) based on the acquired object attribute information, the information vectors corresponding to the cone rays of all hit objects are stitched together according to a predetermined fixed order. For example, stitching is done in a counterclockwise direction from outside to inside, or from inside to outside in a counterclockwise direction. Regarding stitching in a counterclockwise direction, from outside to inside, for a set of cone rays in the right-hand diagram of Figure 17, first, in a counterclockwise order, that is, for the rays in the cross-sectional distribution in the right-hand diagram of Figure 17, starting from 0°, ray cluster a is stitched together, that is, in the order a0, a1, a2, a3 to a4. Similarly, the other ray clusters b, c, d and e are stitched together in a counterclockwise direction, and then, in an outside-to-inside order, that is, from ray cluster a to ray cluster e in the cross-sectional distribution in the right-hand diagram of Figure 17, the basic ray feature vectors are obtained. Subsequently, these basic ray feature vectors can be input into a machine learning model, which can then perform inductive learning on the extracted 3D scene features.

[0097] For example, in an optional embodiment of the 3D scene feature extraction method provided by the embodiment of this application, based on the embodiment corresponding to Figure 2 above, step S107 of integrating the ray feature vector and the elevation map feature vector into a 3D scene feature corresponding to the 3D scene screen, as shown in Figure 12, includes the following steps: In step S1201, the ray feature vector and the elevation map feature vector are stitched together sequentially to obtain the 3D scene features corresponding to the 3D scene screen.

[0098] In the embodiments of this application, after obtaining ray feature vectors and altitude map feature vectors, the ray feature vectors and altitude map feature vectors are sequentially stitched together to obtain 3D scene features corresponding to the 3D scene screen, thereby obtaining 3D scene features with a fixed feature order. Subsequently, the machine learning model can perform inductive learning based on these 3D scene features with a fixed feature order.

[0099] Specifically, as shown in Figure 15, after obtaining the ray feature vector and the altitude map feature vector, the ray feature vector and the altitude map feature vector are stitched together sequentially, that is, stitched together in the order from the ray feature vector to the altitude map feature vector as shown in Figure 15, thereby obtaining the 3D scene features corresponding to the 3D scene screen. This allows the machine learning model to perform inductive learning.

[0100] Optionally, in another preferred embodiment of the 3D scene feature extraction method provided by the embodiment of this application, based on the embodiment corresponding to Figure 2 above, after step S107 of integrating the ray feature vector and the elevation map feature vector into a 3D scene feature corresponding to the 3D scene screen, as shown in Figure 13, the method further includes the following steps: In step S1301, the 3D scene features corresponding to the 3D scene screen are used as feature training samples; In step S1302, the feature training sample is input into the win rate prediction model, and the win rate prediction model estimates the probability that the target character object will win next time; In step S1303, the model parameters are updated by performing reinforcement learning on the win rate prediction model based on the probability of winning next time and the expected win.

[0101] In the embodiments of this application, after acquiring 3D scene features corresponding to a 3D scene screen, these 3D scene features are used as feature training samples in a machine learning model. For example, in this embodiment, when extracting 3D scene features from a 3D game scene screen, these 3D scene features are used as feature training samples in the inductive learning of a game AI by machine learning. That is, the feature training samples are input into a win rate prediction model, the win rate prediction model estimates the probability that the target character object will win next time, and then reinforcement learning is performed on the win rate prediction model based on the probability of winning next time and the expectation of victory to update the model parameters. This allows the game AI to subsequently plan its win rate better based on the trained win rate prediction model.

[0102] Specifically, the win rate prediction model may be represented by a deep neural network with parameter θ using the strategy πθ(at|st). The win rate prediction model takes the observations and actions st=o1:t,a1:t-1 received in the game, and the 3D scene features corresponding to the 3D scene screen of each frame as feature training samples, and selects the action at as the output. Internally, it senses the game environment based on the feature training samples, encodes the observation result ot via convolutional and fully connected layers, merges it as a vector, processes it using a depth sequence network, and finally maps it as the probability distribution of the action, i.e., the probability value of the target character object winning next time. Furthermore, it obtains the expectation or value, i.e., the win expectation, based on the state value function or action value function, and then performs reinforcement learning on the win rate prediction model based on the iterative algorithm, the next win probability value, and the win expectation until the model converges, updating the model parameters. This allows for better planning of the game AI's win rate based on the trained win rate prediction model thereafter.

[0103] Here, after obtaining 3D scene features corresponding to the 3D scene screen, these 3D scene features may be used as feature training samples for inductive learning of a machine learning model. Furthermore, they may be applied to reinforcement learning of other models, such as clearing obstacle prediction, high score prediction, and obstacle prediction, or to supervised learning of other models, without any specific limitations.

[0104] In this way, by training a machine learning model using 3D scene features that are concise, efficient, semantically rich, and include topographic and geological information as sample features, we can enhance the machine learning model's ability to perceive 3D scenes, accelerate the development of machine learning models, and reduce the training cost of machine learning tasks.

[0105] The 3D scene feature extraction device of this application will be described in detail below, with reference to Figure 19, which is a schematic diagram of an embodiment of the 3D scene feature extraction device in this application, and the 3D scene feature extraction device 20 is, A processing unit 201 that projects a set of cone-shaped ray projections from a target character object onto a 3D scene screen, The unit includes an acquisition unit 202 that returns object attribute information of the hit object when the aforementioned cone ray hits the hit object, The processing unit 201 further performs a vector transformation on the object attribute information of each hit object received to obtain the basic trajectory feature vector. The processing unit 201 further performs a feature dimensionality reduction process on the basic ray feature vector to obtain a ray feature vector, The processing unit 201 further collects elevation value matrices corresponding to each granularity, with the location of the target character object as the collection center, based on different granularities. The processing unit 201 further performs feature dimensionality reduction processing on the elevation value matrix corresponding to each of the aforementioned granularities to obtain an elevation map feature vector. The system further includes a decision unit 203 that integrates the aforementioned ray feature vector and the aforementioned altitude map feature vector into a 3D scene feature corresponding to the 3D scene screen.

[0106] Optionally, based on the embodiment corresponding to Figure 19 above, the object attribute information of the hit object includes the position information of the hit object, the type information of the hit object, and the material information. In other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, the processing unit 201 specifically: For each of the object attribute information of the hit object, vector transformation is performed on the position information, type information, and material information of the hit object contained therein to obtain an information vector corresponding to the cone ray that hit the hit object. The information vectors corresponding to each of the aforementioned cone rays are integrated into the basic ray feature vector.

[0107] Optionally, in other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, based on the embodiment corresponding to Figure 19 above, the processing unit 201 specifically: The position information of the hit object is normalized into a vector to obtain the position vector. Feature encoding is performed on the type information and material information of the hit object to obtain an encoded vector. The position vector and the encoded vector are stitched together to obtain an information vector corresponding to the cone ray that hit the object. Specifically, the processing unit 201 sequentially stitches together the information vectors corresponding to each of the visual cone rays to obtain the basic ray feature vector.

[0108] Optionally, in other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, based on the embodiment corresponding to Figure 19 above, the processing unit 201 specifically: Based on different granularities, the location of the target character object is used as the collection center, and N*N grids corresponding to each granularity are collected in all four directions around it, where N is an integer greater than or equal to 1. Based on the N*N grid corresponding to each granularity, an elevation value matrix corresponding to each granularity is generated.

[0109] Optionally, in other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, based on the embodiment corresponding to Figure 19 above, the processing unit 201 specifically: Different granularities are used as the unit length of the grid to be collected, which has different sizes. According to the unit length of the grids to be collected, which have different sizes, the location of the target character object is used as the collection center, and N*N grids corresponding to the unit lengths of the various grids to be collected are collected in all four directions around it. Specifically, the processing unit obtains elevation values ​​corresponding to the center point of each grid in the N*N grid, and generates an elevation value matrix corresponding to the granularity of the N*N grid based on the elevation values.

[0110] Optionally, in other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, based on the embodiment corresponding to Figure 19 above, the processing unit 201 specifically: Stitching and tensor transformation are performed on the altitude value matrix corresponding to each granularity to obtain the altitude map feature tensor. The advanced map feature tensor is subjected to a feature dimensionality reduction process to obtain the advanced map feature vector.

[0111] Optionally, in other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, based on the embodiment corresponding to Figure 19 above, the processing unit 201 specifically: The basic ray feature vector is subjected to a feature dimensionality reduction process via a neural network to obtain the ray feature vector, The elevation value matrix corresponding to each granularity is subjected to feature dimensionality reduction via a neural network to obtain the elevation map feature vector.

[0112] Optionally, based on the embodiment corresponding to Figure 19 above, the neural network that performs feature dimensionality reduction on the basic ray feature vector and the neural network that performs feature dimensionality reduction on the elevation value matrix corresponding to each granularity are the same convolutional neural network, and in other embodiments of the 3D scene feature extraction device provided by the embodiment of this application, the processing unit 201 specifically, By performing a feature dimensionality reduction process on the underlying ray feature vector via a convolutional neural network, we obtain a new ray feature vector. Specifically, the processing unit 201 performs feature dimensionality reduction on the elevation value matrix corresponding to each granularity via a convolutional neural network to obtain elevation map feature vectors.

[0113] Optionally, based on the embodiment corresponding to Figure 19 above, the neural network includes a first neural network and a second neural network, wherein the first neural network and the second neural network are different neural networks, and in other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, the processing unit 201 specifically, The first neural network performs a feature dimensionality reduction process on the underlying ray feature vector to obtain the ray feature vector, Specifically, the processing unit 201 performs feature dimensionality reduction processing on the elevation value matrix corresponding to each granularity via a second neural network to obtain elevation map feature vectors.

[0114] Optionally, in other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, based on the embodiment corresponding to Figure 19 above, the processing unit 201 specifically: From the top of the target character object, a pair of cone-shaped projectiles are emitted by simulating the cone field of view. Specifically, the acquisition unit 202 simulates the field of view of the cone and, after each of the cone rays emitted reaches a length threshold, if any of the cone rays hit an object, it returns the object attribute information of the hit object.

[0115] Optionally, in other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, based on the embodiment corresponding to Figure 19 above, the processing unit 201 specifically: With the top of the target character object as the center, p uniformly distributed ray directions are obtained, and M ray clusters are emitted in each ray direction, the envelope of the M ray clusters is conical, each ray cluster contains p optical cone rays, and the p optical cone rays of each ray cluster are uniformly distributed in M ​​concentric circles, where p is an integer of 2 or more, and M is an integer of 1 or more. Specifically, the acquisition unit 202 returns object attribute information of the hit object if, after the p cone rays of each of the M ray clusters have reached the length threshold, any of the cone rays have hit the hit object.

[0116] Optionally, in other embodiments of the 3D scene feature extraction device provided by the embodiments of this application, based on the embodiment corresponding to Figure 19 above, the decision unit 203 specifically: The ray feature vector and elevation map feature vector are stitched together sequentially to obtain 3D scene features corresponding to the 3D scene screen.

[0117] In possible designs, in other embodiments of the embodiments of this application, The decision unit 203 further uses 3D scene features corresponding to the 3D scene screen as feature training samples. The processing unit 201 further inputs the feature training samples into the win rate prediction model, and the win rate prediction model estimates the probability that the target character object will win next time. The processing unit 201 further updates the model parameters by performing reinforcement learning on the win rate prediction model based on the probability of winning next time and the expectation of winning.

[0118] Other aspects of this application provide schematic diagrams of other computer equipment, as shown in Figure 20, which is a schematic diagram of the structure of computer equipment provided by embodiments of this application, the computer equipment 300 which may vary greatly in arrangement or performance and may include one or more main processors (central processing units, CPUs) 310 (e.g., one or more processors) and memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 331 or data 332. The memory 320 and storage media 330 may be temporary or permanent storage. The programs stored in the storage media 330 include one or more modules (not shown), each module including a set of command operations in the computer equipment 300. Furthermore, the main processor 310 is arranged to communicate with the storage media 330 and to execute a set of command operations in the storage media 330 in the computer equipment 300.

[0119] The computer equipment 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 333, such as Windows Server®, Mac OS X®, Unix®, Linux®, FreeBSD®, etc.

[0120] The above-described computer device 300 further performs the steps in the embodiment corresponding to Figures 2 to 13.

[0121] Another aspect of this application provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps in the method described in the embodiments of Figures 2 to 13 are realized.

[0122] Another aspect of this application provides a computer program product which includes a computer program that, when executed by a processor, accomplishes the steps in the method described in the embodiments of Figures 2 to 13.

[0123] As those skilled in the art will understand, for the sake of simplicity and conciseness, the specific working processes of the systems, apparatus, and units described above can be described by referring to the corresponding processes in the above-described method embodiments, and no further explanation is provided here.

[0124] In some embodiments provided by this application, the disclosed systems, apparatus and methods may be implemented in other forms. For example, the apparatus embodiments described above are schematic, and for instance, the division of the units is merely a division of logical functions, and in actual implementation, other division methods may exist, for example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not implemented. Also, the combinations, direct combinations, or communication connections between them that are indicated or discussed may be indirect combinations or communication connections by some interfaces, apparatus or units, which may be electrical, mechanical or other forms.

[0125] The units described as individual components may or may not be physically separated, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Based on actual needs, some or all of these units can be selected to achieve the objectives of the solution of this embodiment.

[0126] Furthermore, each functional unit in each embodiment of this application may be integrated into a single processing unit, each unit may exist individually and physically, or two or more units may be integrated into a single unit. The above integrated unit may be implemented in hardware form or in the form of a software functional unit.

[0127] The integrated unit may be implemented in the form of a software function unit and, when sold or used as an independent product, may be stored on a computer-readable storage medium. Based on this understanding, the essence of the proposed technology of this application, or a portion that contributes to the prior art, or all or part of the proposed technology, may be embodied in the form of a software product, which is stored on a storage medium and contains several instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in each embodiment of this application. The storage medium may include various media capable of storing program code, such as U disks, portable hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. [Explanation of Symbols]

[0128] 20 3D Scene Feature Extraction Device 201 units acquired 202 Processing Units 203 Generation Units 300 Computer Equipment 310 Main Processor 320 memory 330 Storage medium 331 Application Programs 332 data 333 Operating Systems 340 power supply 350 Wired or wireless network interfaces 360 Input / Output Interface

Claims

1. A method for extracting features from a 3D scene performed by a computer device, The process involves projecting a set of cone-shaped projections from a target character object onto a 3D scene screen, If the aforementioned cone ray hits the hit object, the step of returning the object attribute information of the hit object, The process involves performing a vector transformation on the object attribute information of each received hit object to obtain the basic trajectory feature vector, The steps include: performing a feature dimensionality reduction process on the aforementioned basic ray feature vector to obtain a ray feature vector; The steps include: collecting elevation value matrices corresponding to each granularity, with the location of the target character object as the collection center, based on different granularities; The steps include: performing a feature dimensionality reduction process on the elevation value matrix corresponding to each of the aforementioned granularities to obtain an elevation map feature vector; A method comprising the step of integrating the ray feature vector and the elevation map feature vector into a three-dimensional scene feature corresponding to the three-dimensional scene screen.

2. The object attribute information of the hit object includes the position information of the hit object, the type information of the hit object, and the material information. The step of performing a vector transformation on the object attribute information of each received hit object to obtain a basic trajectory feature vector is as follows: The steps include: performing a vector transformation on the object attribute information of each of the hit objects, applying the position information, type information, and material information of the hit object included in the object attribute information to obtain an information vector corresponding to the cone ray that hit the hit object; The method according to claim 1, comprising the step of integrating information vectors corresponding to each of the aforementioned cone rays into the basic ray feature vector.

3. The step of performing a vector transformation on the position information of the hit object, the type information of the hit object, and the material information included in the object attribute information to obtain an information vector corresponding to the cone ray that hit the hit object is: The steps include: performing a vector normalization process on the position information of the hit object to obtain a position vector; The steps include: performing feature encoding on the type information and material information of the hit object to obtain an encoded vector; The step includes stitching the position vector and the encoded vector to obtain an information vector corresponding to the cone ray that hit the hit object, The step of integrating the information vectors corresponding to each of the aforementioned cone rays into the basic ray feature vector is: The method according to claim 2, further comprising the step of sequentially stitching together information vectors corresponding to each of the aforementioned cone rays to obtain the basic ray feature vector.

4. The step of collecting elevation value matrices corresponding to each granularity, with the location of the target character object as the collection center, based on different granularities, A step of collecting N*N grids corresponding to each granularity in all four directions around the location where the target character object is located, with the location of the target character object as the collection center, based on different granularities, wherein N is an integer of 1 or more. The method according to claim 1, comprising the step of generating an elevation value matrix corresponding to each granularity based on an N*N grid corresponding to each granularity.

5. The step of collecting N*N grids corresponding to each granularity in all directions around the location where the target character object is located, based on different granularities, is as follows: The steps include: setting the different granularities to be used as the unit length of the grid to be collected, which has different sizes; The process includes the step of collecting N*N grids corresponding to various unit lengths of the grids to be collected, with the location of the target character object as the collection center, according to the unit length of the grids to be collected, which have different sizes, in all four directions. The step of generating an elevation value matrix corresponding to each granularity based on an N*N grid corresponding to each granularity is: The method according to claim 4, further comprising the steps of obtaining an altitude value corresponding to the center point of each grid in the N*N grid, and generating an altitude value matrix corresponding to the granularity corresponding to the N*N grid based on the altitude values.

6. The step of performing a feature dimensionality reduction process on the elevation value matrix corresponding to each of the aforementioned granularities to obtain an elevation map feature vector is as follows: The steps include performing stitching and tensor transformation on the altitude value matrices corresponding to each of the aforementioned granularities to obtain an altitude map feature tensor, The method according to claim 1, comprising the step of performing a feature dimensionality reduction process on the advanced map feature tensor to obtain the advanced map feature vector.

7. The step of performing a feature dimensionality reduction process on the aforementioned basic ray feature vector to obtain a ray feature vector is: The process includes the step of performing a feature dimensionality reduction process on the basic ray feature vector via a neural network to obtain the ray feature vector, The step of performing a feature dimensionality reduction process on the elevation value matrix corresponding to each of the aforementioned granularities to obtain an elevation map feature vector is as follows: The method according to claim 1, further comprising the step of performing a feature dimensionality reduction process on the elevation value matrix corresponding to each granularity via a neural network to obtain the elevation map feature vector.

8. The neural network for performing feature dimensionality reduction on the aforementioned basic ray feature vector and the neural network for performing feature dimensionality reduction on the height value matrices corresponding to each granularity are the same convolutional neural network. The step of performing a feature dimensionality reduction process on the basic ray feature vector via a neural network to obtain the ray feature vector is as follows: The process includes the step of performing a feature dimensionality reduction process on the underlying ray feature vector via the convolutional neural network to obtain the ray feature vector, The step of performing a feature dimensionality reduction process on the elevation value matrix corresponding to each granularity via a neural network to obtain the elevation map feature vector is as follows: The method according to claim 7, further comprising the step of performing a feature dimensionality reduction process on the elevation value matrix corresponding to each granularity via the convolutional neural network to obtain the elevation map feature vector.

9. The neural network includes a first neural network and a second neural network, wherein the first neural network and the second neural network are different neural networks. The step of performing a feature dimensionality reduction process on the basic ray feature vector via a neural network to obtain the ray feature vector is as follows: The process includes the step of performing a feature dimensionality reduction process on the basic ray feature vector via the first neural network described above to obtain the ray feature vector, The step of performing a feature dimensionality reduction process on the elevation value matrix corresponding to each granularity via a neural network to obtain the elevation map feature vector is as follows: The method according to claim 7, further comprising the step of performing a feature dimensionality reduction process on the elevation value matrix corresponding to each granularity via the second neural network to obtain the elevation map feature vector.

10. The step of emitting a set of visual cone rays from the target character object is: The step includes simulating the cone field of view angle and emitting a set of cone rays from the top part of the target character object, If the aforementioned cone ray hits the hit object, the step of returning the object attribute information of the hit object is: The method according to claim 1, further comprising the step of returning object attribute information of the hit object if, after each of the cone rays emitted by simulating the cone field of view has reached a length threshold, any of the cone rays have hit a hit object.

11. The step of simulating the cone field of view angle and emitting a set of cone rays from the top of the target character object is: A step comprising obtaining p uniformly distributed ray directions with the top of the target character object as the center, and emitting M ray clusters in each ray direction, wherein the envelope of the M ray clusters is conical, each ray cluster contains p optical cone rays, the p optical cone rays of each ray cluster are uniformly distributed in M ​​concentric circles, p is an integer of 2 or more, and M is an integer of 1 or more, After each of the cone rays emitted by simulating the cone field of view has reached a length threshold, if any of the cone rays have hit an object, the step of returning the object attribute information of the hit object is as follows: The method according to claim 10, further comprising the step of returning object attribute information of the hit object if, after the p cone rays of each ray cluster among the M ray clusters have reached the length threshold, any of the cone rays have hit the hit object.

12. The step of integrating the aforementioned ray feature vector and the aforementioned altitude map feature vector into the 3D scene features corresponding to the 3D scene screen is: The method according to claim 1, further comprising the step of sequentially stitching together the ray feature vector and the elevation map feature vector to obtain three-dimensional scene features corresponding to the three-dimensional scene screen.

13. After the step of integrating the ray feature vector and the elevation map feature vector into the 3D scene features corresponding to the 3D scene screen, the method The steps include using the 3D scene features corresponding to the aforementioned 3D scene screen as feature training samples, The steps include inputting the aforementioned feature training sample into a win rate prediction model and using the win rate prediction model to estimate the probability that the target character object will win next time, The method according to claim 1, further comprising the step of performing reinforcement learning on the win rate prediction model to update the model parameters based on the probability value of winning next time and the expectation of winning.

14. A feature extraction device for 3D scenes, A processing unit that projects a set of cone-shaped projections from a target character object onto a 3D scene screen, The system includes an acquisition unit that, when the aforementioned cone ray hits an object, returns object attribute information of the hit object, The processing unit further performs a vector transformation on the object attribute information of each received hit object to obtain a basic trajectory feature vector. The processing unit further performs a feature dimensionality reduction process on the basic ray feature vector to obtain a ray feature vector, The processing unit further collects elevation value matrices corresponding to each granularity, with the location of the target character object as the collection center, based on different granularities. The processing unit further performs a feature dimensionality reduction process on the height value matrix corresponding to each granularity to obtain a height map feature vector. The apparatus further includes a decision unit that integrates the ray feature vector and the altitude map feature vector into a three-dimensional scene feature corresponding to the three-dimensional scene screen.

15. Computer equipment, Includes memory in which computer programs are stored, a processor and a bus system, When the processor executes the computer program, it performs the method according to any one of claims 1 to 13. The computer device is characterized in that the bus system connects the memory and the processor to communicate with each other.

16. A computer program characterized by causing a computer to perform the method described in any one of claims 1 to 13.

Images