WO2019144346A1 - Procédé de traitement d'objet dans une scène virtuelle, dispositif et support d'informations - Google Patents

Procédé de traitement d'objet dans une scène virtuelle, dispositif et support d'informations Download PDF

Info

Publication number
WO2019144346A1
WO2019144346A1 PCT/CN2018/074156 CN2018074156W WO2019144346A1 WO 2019144346 A1 WO2019144346 A1 WO 2019144346A1 CN 2018074156 W CN2018074156 W CN 2018074156W WO 2019144346 A1 WO2019144346 A1 WO 2019144346A1
Authority
WO
WIPO (PCT)
Prior art keywords
real
data
neural network
training sample
time
Prior art date
Application number
PCT/CN2018/074156
Other languages
English (en)
Chinese (zh)
Inventor
李德元
李源纯
姜润知
黄柳优
王鹏
魏学峰
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Priority to PCT/CN2018/074156 priority Critical patent/WO2019144346A1/fr
Priority to CN201880003364.1A priority patent/CN110325965B/zh
Publication of WO2019144346A1 publication Critical patent/WO2019144346A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/44—Arrangements for executing specific programs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/44—Arrangements for executing specific programs
    • G06F9/451—Execution arrangements for user interfaces
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T19/00—Manipulating 3D models or images for computer graphics

Definitions

  • the present invention relates to an electrical digital data processing technology, and in particular, to an object processing method, device, and storage medium in a virtual scenario.
  • the display technology based on graphics processing hardware expands the sensing environment and the channel for obtaining information, especially the display technology of the virtual scene, and can realize the intelligent interaction of various virtual objects in the human, human and virtual scenes according to actual application requirements. .
  • the virtual scene can also achieve a real-world visual perception effect by means of stereoscopic display technology, typically using stereoscopic display technologies such as stereoscopic projection, virtual reality and augmented reality technology to output various virtual scenes.
  • stereoscopic display technologies such as stereoscopic projection, virtual reality and augmented reality technology to output various virtual scenes.
  • the behavior of the simulated object is adapted to the behavior of the object controlled by the user, and the simulated object is implemented according to the behavior of the user controlled object.
  • the behaviors that the user-controlled object implements are adapted to form an interaction process in the virtual environment.
  • the game is a typical application of the virtual scene display technology.
  • the user can run the game through the device.
  • the user-controlled game object cooperates with other game objects on the line to fight or play against each other.
  • the embodiment of the invention provides an object processing method, a device and a storage medium in a virtual scene, which can improve the intelligence degree of the simulated object by using an artificial neural network model.
  • An embodiment of the present invention provides a method for processing an object in a virtual scenario, including:
  • the artificial neural network model is trained by taking a scene data sample included in the pre-processed training sample set as an input, and outputting the operation data sample included in the pre-processed training sample set as an output.
  • An embodiment of the present invention provides an object processing method in a virtual scenario, including:
  • Real-time operation data corresponding to the fourth object is executed in the real-time virtual scene.
  • An embodiment of the present invention provides an apparatus for processing an object in a virtual scenario, including:
  • a memory for storing executable instructions
  • the processor when used to execute the executable instructions stored in the memory, implements a method for object processing in any virtual scene provided by the embodiment of the present invention.
  • the embodiment of the invention provides a storage medium, which stores executable instructions for causing a processor to execute an object processing method in any virtual scene provided by the embodiment of the present invention.
  • the problem of learning the skill of the user's operation object by the artificial neural network model is converted into the process of the artificial neural network model training according to the scene data sample and the operation data sample; since the transformation of the data by the artificial neural network model is essentially an iterative update parameter The process does not need to formulate the specific logic of the object to perform the operation data.
  • the operational data with rich expression form can be realized, which is close to the real operation of the user; and the artificial neural network algorithm has better anti-noise ability.
  • the application speed is fast, so the artificial neural network model rarely has operational errors like the user's real operation, and the decision speed is much faster than the user's response time.
  • the operation skill is higher than the learned user, which significantly improves the intelligence level. .
  • FIG. 1 is a schematic diagram of an optional application mode of an object processing method in a virtual scene according to an embodiment of the present invention
  • FIG. 2 is a schematic diagram of an optional application mode of an object processing method in a virtual scene according to an embodiment of the present invention
  • FIG. 3 is a schematic flowchart of an object processing method in a virtual scenario according to an embodiment of the present disclosure
  • FIG. 4 is a schematic flowchart of an object processing method in a virtual scenario according to an embodiment of the present invention.
  • FIG. 5 is a schematic structural diagram of an apparatus for processing an object in a virtual scene according to an embodiment of the present disclosure
  • FIG. 6 is an optional schematic diagram of training and application of an artificial neural network model according to an embodiment of the present invention.
  • FIG. 7 is an optional schematic flowchart of converting a joystick moving angle data into an action mode according to an embodiment of the present invention
  • FIG. 8A is a schematic structural diagram of an artificial neural network model according to an embodiment of the present invention.
  • 8B is an optional schematic diagram of an artificial neural network according to an embodiment of the present invention for predicting operational data according to real-time scene data;
  • FIG. 9A is a schematic structural diagram of an artificial neural network model according to an embodiment of the present invention.
  • FIG. 9B is an optional schematic diagram of an artificial neural network according to an embodiment of the present invention for predicting operational data according to real-time scene data.
  • the terms "including”, “comprising” or any other variations thereof are intended to cover a non-exclusive inclusion, such that a method or apparatus comprising a plurality of elements includes not only the Elements, but also other elements not explicitly listed, or elements that are inherent to the implementation of the method or device.
  • an element defined by the phrase “comprising a " does not exclude the presence of additional related elements in the method or device including the element (eg, operation in the method or unit in the device)
  • the unit here may be part of a circuit, part of a processor, part of a program or software, etc.).
  • the embodiments of the present invention provide a series of operations, but the method provided by the embodiments of the present invention is not limited to the operations provided.
  • the device provided by the embodiments of the present invention includes a series of units, but the present invention is implemented.
  • the apparatus provided by the example is not limited to including the unit explicitly provided, and may also include a unit that is required to be set for acquiring related information or processing based on the information.
  • AI Artificial Intelligence
  • Artificial neural network is a mathematical model that mimics the structure and function of biological neural networks.
  • the exemplary structures of artificial neural networks in this paper include BP (Back Propagation) neural network and Recurrent Neural Networks (RNN).
  • the former is trained by an error backpropagation algorithm, and the artificial neural network is used for function estimation or approximation, including an input layer, an intermediate layer, and an output layer.
  • Each layer is connected by a plurality of processing units, and each node uses an excitation function.
  • the input data is processed and output to other nodes, and exemplary types of excitation functions include threshold type, linear type, and Sigmoid type.
  • Each bit represents a state, with one bit being one and the rest being zero.
  • the virtual scene using the output of the device different from the real world scene, can form a visual perception of the virtual scene through the aid of the naked eye or the device, for example, through the two-dimensional image outputted by the display screen, through stereoscopic projection, virtual reality and augmented reality.
  • Three-dimensional images output by stereoscopic display technologies such as technology; in addition, various kinds of simulation real-world perceptions such as auditory perception, tactile perception, olfactory perception, and motion perception can be formed by various possible hardware.
  • Scene data which represents various features that the objects in the virtual scene are represented during the interaction, for example, may include the position of the object in the virtual scene.
  • the scene data may include various times when the various functions configured in the virtual scene need to wait (depending on the same time can be used in a specific time)
  • the number of functions can also represent attribute values of various states of the game character, including, for example, health (also known as red amount) and mana (also called blue amount); and, for example, in a virtual reality implementation store
  • health also known as red amount
  • mana also called blue amount
  • the customer's various preferences for the product can be expressed.
  • Operational data indicating that the objects in the virtual scene are subjected to various operations performed by the user/artificial neural network model control, such as operations related to controllers such as touch screens, voice switches, mice, keyboards, and joysticks.
  • Data whether the function of the object is used (ie, which functions are used by the object), various actions implemented (such as whether to jump, whether to rotate and whether to squat, etc.) and the functions used by the object;
  • the data may be acquired from the hardware layer of the device during the period in which the device outputs the virtual scene, and various functions used in the virtual scene may be read from the operation interface of the output virtual scene.
  • the device (such as a terminal or server) is controlled by the role of the artificial neural network model control.
  • Embodiments of the present invention provide an object processing method in a virtual scene, a device for implementing an object processing method in a virtual scene, and a storage medium storing executable instructions for executing the object processing method, in order to facilitate understanding of the implementation of the present invention.
  • An example of the object processing method in the virtual scenario provided by the example is first described in the exemplary implementation scenario of the object processing method in the virtual scenario provided by the embodiment of the present invention.
  • the virtual scenario may be completely based on the output of the terminal device or output based on the cooperation between the terminal device and the server. .
  • FIG. 1 is a schematic diagram of an optional application mode of an object processing method in a virtual scenario 100 according to an embodiment of the present invention, which is applicable to some computing systems that rely on the computing power of the terminal device 200 to complete the virtual mode.
  • the application mode of the related data calculation of the scene such as the game of the stand-alone/offline mode, completes the output of the virtual scene through the terminal device 200 such as a smartphone, a tablet, and a virtual reality/augmented reality device.
  • the terminal device 200 calculates the data required for display by the graphics computing hardware, and completes loading, parsing, and rendering of the display data, and outputs a video frame capable of visually percepturing the virtual scene in the graphic output hardware.
  • a two-dimensional video frame is presented on a display screen of a smart phone, or a video frame that realizes a three-dimensional display effect is projected on a lens of the augmented reality/virtual reality glasses; in addition, in order to enrich the sensing effect, the device can also utilize different hardware.
  • auditory perception, tactile perception, motion perception, and taste perception are examples of auditory perception.
  • the terminal device 200 runs a stand-alone version of the game application, and outputs a virtual scene including action role playing during the running of the game application.
  • the virtual scene is an environment for the game character to interact, for example, for the game character to fight.
  • the virtual scene includes a first object 110 and a second object 120,
  • the first object 110 may be a game character controlled by the user, that is, the first object 110 is controlled by the real player, Moving in a virtual scene in response to real player action on the controller (including touch screen, voice switch, keyboard, mouse, joystick, etc.), such as when the real player moves the joystick to the left, the first object will be virtual Moving to the left in the scene, it is also possible to remain in place, jump, and use various functions (such as skills and props);
  • the second object 120 may be an object that interacts with the first object 110 in the virtual scene, the second object 120 may be a game character implemented by a robot model in a game application
  • a shopping guide application is installed in the terminal, and a three-dimensional virtual scene of the store is output during the running of the shopping guide application, and the first object 110 and the second object 120 are included in the virtual scene, and the first object 120 may be the user/ The user's own three-dimensional image, the first object 120 can be freely moved in the store, and the naked eye/virtual reality device user can perceive the three-dimensional image of the store and various commodities, and the second object 120 can be a shopping guide character output by using the stereoscopic display technology.
  • the shopping guide role can appropriately answer the user's consultation and recommend appropriate products to the user.
  • FIG. 2 is a schematic diagram of an optional application mode of the object processing method in the virtual scenario 100 according to the embodiment of the present invention, which is applied to the terminal device 200/300 and the server 400.
  • the application mode suitable for the virtual scene calculation by the computing power of the dependent server 400 and outputting the virtual scene at the terminal device 200/300.
  • the server 400 performs calculation of the virtual scene related display data and transmits it to the terminal device 200/300, and the terminal device 200/300 relies on the graphics computing hardware to complete the loading, parsing, and rendering of the calculated display data.
  • a two-dimensional video frame may be presented on a display screen of the smart phone, or a video frame realizing a three-dimensional display effect may be projected on the lens of the augmented reality/virtual reality glasses;
  • the corresponding hardware output of the terminal device can be utilized, for example, using the microphone output to form an auditory perception, using the vibrator output to form a tactile perception, and the like.
  • the terminal device 200/300 runs a web version of the game application, and interacts with other users through a game server, and the terminal device 200/300 outputs a virtual scene of the game application, including the first object 110 and the second object 120.
  • the first object 110 may be a game character controlled by a user (also referred to as a real player, to be distinguished from the robot model), and the first object 110 is controlled by the real player and will respond to the real player to the controller (including the keyboard and the mouse).
  • the second object 120 is a game character capable of interacting with the first object 110.
  • the number of the second objects 120 may be one or more.
  • the second object 120 may be controlled by a robot model.
  • One or more game characters in the online version of the game, the second object 120 may be one or more game characters controlled by the robot model, or may be one or more game characters controlled by other users on the line, and may also These are the two types of game characters.
  • a robot model implemented using a finite state machine divides behavior into multiple states in a finite state machine, and when a event trigger condition is satisfied, a state change is caused, and the game character performs a corresponding behavior.
  • a robotic model implemented using a hierarchical state machine in a hierarchical state machine, classifies behaviors and groups the same type of behavior into a state machine in order to facilitate maintenance of a large number of game character behaviors, and then Each state machine is then combined into a higher-level state machine to implement query and execution of different types of behavior.
  • each node of the tree structure of the behavior tree is a game behavior
  • the decision process of the game behavior is to search the behavior tree according to a certain rule from the root node, and find The qualified nodes are executed and the game behavior of the corresponding nodes is executed and executed.
  • the specific logic of the robot model is determined by the developer, the technical cost is high, the level of intelligence depends on the developer's technical level, the expression form is limited, and the performance behavior of the game character controlled by the real player is diverse and random. The characteristics are very different.
  • an embodiment of the present invention further provides a robot model implemented by an artificial neural network, which is also referred to as an artificial neural network model, and is collected from a sample virtual scene to represent an interactive object including an interactive object (hereinafter, the first object and a second object) a sample of the scene data of the feature, and collecting samples of the operational data representing the behavior of the object in the virtual scene, forming a set of training samples to train the artificial neural network model, so that the artificial neural network model learns the operational skills in the interaction process (
  • the response is an artificial neural network model parameter); for the trained artificial neural network model, real-time scene data of a user-controlled object (hereinafter, assumed to be a third object) in a real-time virtual scene is acquired, according to the learned operation skill To predict the operation data executed by the object of the user control object (hereinafter, assumed to be the fourth object), the operation data is executed by the control object to implement the corresponding behavior, and the intelligent interaction between the objects is realized.
  • an artificial neural network model which is also referred to as
  • FIG. 3 is a schematic flowchart diagram of an object processing method in a virtual scene according to an embodiment of the present invention. The process of forming a training sample training artificial neural network model by collecting sample virtual scenes will be described below with reference to FIG.
  • an interaction process between the first object and the second object in the sample virtual scene is collected; as an example, the terminal is a terminal having graphics computing capability and graphics output capability, including a smartphone, a tablet, and a virtual reality/augmented reality glasses.
  • the virtual scene output by one or more terminals may be collected as a sample virtual scene for collecting related data, for example, for a virtual scene of the game, for a plurality of online users joining the same map, the game
  • the server collects the interaction process of the game characters controlled by the plurality of online users when calculating and synchronizing the virtual scene related data to the online user.
  • the virtual scene that the terminal outputs to itself may also be used. Collect as a virtual scene sample.
  • the first object is an object controlled by the first terminal user in the virtual scene
  • the second object is an object that interacts with the first object in the virtual scene
  • the interaction process between the first object and the second object Including a process in which either party uses a function or a relative position changes, for example, a process in which a first object chases or flees a second object in a game, and the first object and the second object perform cooperative combat using their respective functions (including skills and props).
  • the process of the battle may divide the output process of the virtual environment according to a specific duration (for example, 10 minutes), or may be divided according to a specific state, for example, a game in which the outcome of the game is determined in the game application. , the process of completing the purchase of a commodity in the shopping guide application, and the like.
  • basic information of the virtual scene is formed according to the characteristics of the collected virtual scene; in some embodiments of the present invention, the basic information includes features describing the virtual scene, such as the size of the virtual scene, and the interaction process occurring in the virtual scene. The results and other information.
  • the basic information may include the size of the game character's battle environment, and the win/negative/flat result of each game.
  • the basic information may include the size of the store environment, whether the customer purchased the product, and the like.
  • the collection result is combined to form a scene data sample of the corresponding sampling point according to the position of the first object and the second object acquired in the interaction process, and the waiting time of the first object and the second object using function.
  • an uninterrupted acquisition manner is adopted for the interaction process, that is, the location and waiting time of the object in the virtual environment are continuously collected during the entire interaction process, for example, when the terminal outputs each image frame of the virtual scene.
  • Collecting the position of the object in the virtual environment and the waiting time of the function of the corresponding object the combination of at least two image frames constitutes a sampling point, the positions of the first object and the second object collected at each sampling point, and the first The waiting time of the object and the second object are combined to form a scene data sample corresponding to one sampling point, and the manner of performing acquisition once for each image frame can form a complete scene data sample of the interaction process.
  • the interaction process includes a series of image frames, and each of the two image frames constitutes one sampling point, and an optional recording form of the scene data sample collected at the sampling points (image frames 1, 2) is:
  • the waiting time corresponding to the function of the first object in the image frame 2 and the waiting time corresponding to the function of the second object in the image frame 2].
  • a sampling point is set at a certain number of image frames (or time) according to the acquisition precision, and is collected in the window time of each sampling point of the interaction process, and the sampling point window generally includes a plurality of image frames, collecting the positions of the first object and the second object at the sampling point, and the waiting time of the first object and the second object using the function, and combining the collection results at one sampling point to form a scene data sample of the corresponding sampling point
  • the sampling point collection method the data of the virtual scene in the interaction process can be relatively completely collected on the one hand, and the repetition degree of the scene data sample is significantly reduced on the other hand, and the diversity of the training sample based on the scene data sample is ensured.
  • the acquisition accuracy is positively correlated with the number of sampling points, that is, the higher the acquisition precision, the more the number of sampling points are set, and vice versa; generally, the sampling points can be uniformly set during the interaction process. For example, when the sampling period is 1 second, the acquisition is started every interval of 1 second during the interaction, and the position and waiting time of the object in the virtual scene are sampled in the window time of the sampling point.
  • the setting of the sampling point can be adaptively set according to the amount of information of different stages in the interaction process, by learning the variation law of the information amount at different stages in the virtual scene of the game, according to the information amount and sampling of different stages.
  • the number of objects in the virtual scene, the number of functions used, and the number of objects being moved; the greater the amount of information in one phase of the interaction, the more the number of sampling points is set, that is, the two are positively correlated.
  • the information volume at different stages of the interaction process can be adaptively collected, so that the constructed training samples can train the artificial neural network, and the artificial neural network can perform sufficient skill learning according to the details of the user operation.
  • the amount of information may also be represented by referring to a change of a signal related to the output process of the virtual scene.
  • the code rate is higher than the average bit rate of the interactive process, usually an action.
  • the amount of information is relatively large due to the relatively frequent transformation, so the number of sampling points is higher than other stages of the interaction process, thereby realizing the effect of comprehensively collecting the scene data on the basis of reducing the number of sampling operations.
  • the actual application scenario selection may be selected.
  • the function in the virtual scenario of the game application, the function may be a skill or an item used by the game character; in the virtual scenario of the shopping guide application, the function may be a customer-to-product The behavior of attention, as well as the introduction behavior of the merchandiser.
  • the first object and the second object when the interaction process of collecting the virtual scene forms the scene data sample, the first object and the second object may be acquired in addition to the location of the first object and the second object collected in the virtual process and the waiting time.
  • Various attribute values for collecting various attribute values of the first object and the second object for each image frame/sampling point of the virtual scene, together with the attribute values and locations and usage functions collected at the corresponding image frame/sampling point The waiting time is combined to form a scene data sample of the corresponding image frame/sampling point.
  • the attribute values of each sample point can also be collected: including the red amount and the blue amount, wherein the red quantity refers to the game character life value/health value (HP, He alth Point).
  • the blue value refers to the mana point/magic point of the game character.
  • the value of the attribute in the game is not limited thereto, and may include, for example, a wealth value.
  • a scene data sample including an attribute value may be recorded in the following form: [the position of the first object, the position of the second object, the waiting time corresponding to the function of the first object, the attribute value of the first object, the second object The waiting time corresponding to the function, the attribute value of the second object].
  • operation 104 based on the controller operation data implemented when controlling the first object acquired during the interaction process, whether the function possessed by the first object is released, the operation data samples of the corresponding sampling points are merged.
  • a continuous acquisition manner may be adopted for the interaction process, that is, various operation data executed when the user controls the first object without interruption during the entire interaction process, for example, at least one of the following operation data: the controller
  • the operational data, the pose mode of the first object (including whether it is stationary, whether it is rotated, whether it jumps or not, etc.), the function used by the first object, and the way the first object uses the function; will be in each image frame
  • the collected various types of operational data are combined to form a scene data sample corresponding to one image frame.
  • the interaction process includes a series of image frames, each of which constitutes one sample point, and an optional record form of the operation data samples collected at the image frames 1, 2 is:
  • the first object operates in the image frame 1 controller, the first object uses the function in the image frame 1,
  • the first object operates on the data of the image frame 2, and the first object uses the function in the image frame 2,
  • the second object operates on the data of the controller of the image frame 2, and the second object operates the data in the image frame 2 controller.
  • various operation data executed by the user when controlling the first object are intermittently collected at the sampling point of the interaction process, for example, at least one of the following operation data: controller operation data, and a pose mode of the first object ( Including whether it is stationary, whether it is rotated, whether it jumps or not, etc., and the way the first object uses the function, the various types of operation data collected at each sampling point are combined into a form record of the data set to form a corresponding one.
  • a sample of scene data of a sample point for example, a vector of each type of operation data as a dimension, and various types of vectors are combined into a vector of a higher dimension.
  • the interaction process includes a series of image frames, and the scene data samples acquired by each image frame may be recorded with reference to the following form of the operation data samples collected in the image frame 1: [control operation data of the first object, first The way the object uses the function, the way the first object uses the function; the controller of the second object manipulates the data, the second object uses the way of function 1, and the second object uses the way of the function].
  • the virtual scene is selected according to the actual application of the virtual scene.
  • the controller operation data may be the operation data of the rocking joystick, and the function mode may be whether to release. Skills, whether to initiate a normal attack, etc.; in the virtual scenario of the shopping guide application, the controller operation data can be used to control the data of the object forward, backward, and leave, and the function mode can be to recommend different commodities.
  • the training samples of the corresponding sampling points are constructed according to the scene data samples and the operation data samples collected at different sampling points, and the training sample sets are formed according to the training samples constructed at different sampling points of the interaction process.
  • the scene data sample and the operation data sample corresponding to each video frame are continuously collected during the interaction, the scene data sample and the operation data sample collected in one video frame and the collected interaction process are collected for any one video frame.
  • the basic information is combined to form a training sample corresponding to one video frame, and the training samples corresponding to the multiple video frames form a training sample set.
  • the scene data sample and the operation data sample are collected at each sampling point of the interaction process, for any one of the sampling points, the scene data sample and the operation data sample collected at one sampling point, and the collected interaction process
  • the basic information is combined to form a training sample corresponding to one video frame, and the training samples corresponding to the multiple video frames form a training sample set.
  • the training samples in the formed training sample set are recorded in the following form: [training sample 1 of image frame 1, training sample of image frame 2 n...], for each training sample, the following can be used to record [the basic information of the virtual scene to which the image frame 1 belongs, the scene data sample of the image frame 1, and the scene data sample of the image frame 2].
  • a pre-processing for adapting the artificial neural network model is performed on the training sample set.
  • the exemplary embodiments provided below describe a scheme for different forms of pre-processing of a set of training samples, it being understood that some or all types of pre-processing may be performed for a set of training samples.
  • the operation data included in the training sample is collected from an interaction process in which the user operates insufficiently, and if the artificial neural network model is trained using such a sample, it is difficult to quickly learn the user.
  • the skill of operating the first object affects the learning efficiency of the artificial neural network model.
  • Such an interaction process is essentially an interaction process that is ineffective for training the artificial neural network model. Therefore, it is determined from the interaction process of the training sample from the training sample set.
  • Interaction process Delete training samples that are not part of a valid interaction process in the training sample set.
  • the following conditions can be used for filtering: the sum of the number of cumulative usage functions of the first object and the second object in the interaction process exceeds the threshold of the total function usage times of an interaction process.
  • the interaction between objects in a game is insufficient to avoid collecting samples from such invalid interactions; when the overall skill usage threshold of a game is exceeded, the interaction process is valid, that is, in this game.
  • the interaction of the objects is sufficient, and a comprehensive range of scene data samples and operational data samples can be collected from the interaction process. Therefore, the artificial neural network model trained by the training samples can fully learn the game control skills of real players. .
  • target function of the statistical use number it may be a specific function of the first object and the second object, and the skill/props control skill of the game character that needs to be learned according to the artificial neural network model, and the skill to be learned is collected/
  • the props form operational data samples, which can effectively reduce the complexity of constructing the training samples based on the operational data samples, and improve the learning efficiency of the artificial neural network model; of course, the target functions of the above statistical usage times can also be the first object and the second object. All the features used by the object.
  • the operation result included in each training sample in the training sample set may be obtained.
  • the operational results of the process eg, a game
  • selecting training samples with valid attributes to train the artificial neural network can accelerate the artificial neural network model to learn the user from the training samples Controls the speed of the first object's operational skills and saves training time.
  • the number of times of the cumulative use function of the first object in the interaction process may be continuously determined, whether the threshold value of the function usage of the first object is exceeded, and the invalid attribute or the valid attribute of the training sample is marked according to the judgment result, and the following two cases are involved:
  • the training corresponding to the interaction process that the user does not reach the target by the normal operation skill is filtered out by the number of times the cumulative use function of the first object in the interaction process is performed.
  • the sample ensures that the training samples used to train the artificial neural network model include the user's real operational skills.
  • normalized preprocessing may be performed on the scene data samples included in each training sample in the training sample set, for artificial
  • the training process of the neural network model is a process of continuously iterating and optimizing the parameters of the artificial neural network model for the input part and the output part of the training sample. In this process, if the scene data samples constituting the training sample are normalized The value will be compressed into a unified value space, which will speed up the optimal solution of the parameters of the artificial neural network model, and improve the accuracy of the artificial neural network.
  • determining different types of fields included in the scene data samples may include distance components and usage functions according to differences in scenes in different application modes. Waiting time and attribute values, etc., an example is as follows:
  • the different types of fields are separately normalized, and the normalized field data is connected as a component, for example, by the following form:
  • the scene data samples in the training samples are replaced with vectors formed by normalized component connections.
  • the distance component is compared with the component of the virtual scene in the direction of the corresponding coordinate axis, and the normalized result of the distance component in the corresponding coordinate axis direction is obtained according to the ratio operation.
  • the scene data sample obtained by the acquisition interaction process may include the positions of the first object and the second object, and the distance component is obtained by replacing the position of the scene data sample including the first object with the position of the second object.
  • the position p1 of the first object and the position p2 of the second object included in the scene data sample are respectively mapped into the same reference coordinate system as the virtual scene, and the first object and the second object have a certain value in the reference coordinate system.
  • the distance d is calculated according to the projection principle.
  • the distance components formed based on different coordinate axes, for example, based on the x-axis and the y-axis, are dx and dy, and the positions of the first object and the second object included in the training sample (x, y) are Replaced with the distance component (dx, dy).
  • the waiting time is compared with the sum of the waiting times of the corresponding functions in the training sample set, and the normalized result of the corresponding waiting time is obtained according to the ratio operation.
  • the waiting time of the first skill 1 of the first object recorded in all the training samples is counted and summed , that is, T1
  • the ratio of the waiting time of the skill 1 to the summation t1/T is used as the normalized result of t1.
  • the attribute value is compared with the sum of the same type of attribute values in the training sample set, and the normalized result of the attribute value is obtained according to the ratio operation.
  • the type of the field in the training sample as the life value of the first object (indicated as Life1) as an example
  • the waiting time t1 of the skill 1 as an example
  • the first recorded in all the training samples is counted.
  • the waiting time of the life value of the object is added and recorded as LIFE, and the ratio Life1/LIFE of the first object's life value Life1 and the added LIFE is used as a normalization result.
  • the action mode of the first object when the training sample set is subjected to preprocessing for adapting the artificial neural network model, the action mode of the first object may be determined according to the controller operation data included in the training sample, and the training sample may be The controller operation data of an object is replaced with a corresponding action mode scheme, and the action mode of the first object relative to the second object is replaced with the controller operation data of the corresponding plurality of image frames in the training sample, thereby effectively simplifying the data of the training sample. Complexity, without loss of information, and thus significantly reduces the complexity of training samples.
  • the device forms a visual perception of the virtual scene by outputting an image frame corresponding to the virtual scene. Therefore, when collecting at each sampling point of the interaction process of the virtual scene, due to the randomness of the user operating the controller, it may appear in this
  • the controller operation data is not collected at the sampling point (because the user does not operate the controller during the window time of this sampling point), and the operation data collected to the controller for some or all of the image frames included in the sampling point may also occur;
  • the number of times the controller operates data is continuously acquired during the output of the image frame, also referred to herein as the number of image frames corresponding to the controller operation data, which will reflect whether the user operates the controller and controls the first object through the controller. The way the action is implemented.
  • the action mode is determined according to the number of image frames corresponding to the controller operation data in the operation data sample, and the following two cases are involved:
  • the normalization processing for the scene data samples in the training samples has been described above, and in some embodiments of the present invention, the operational data in the training samples When the sample is preprocessed, the normalization process can be performed so that the values of the data in the finally formed training samples are in a small value space, which reduces the complexity of the artificial neural network model training and improves the training efficiency.
  • the set coding sequence determines a bit of a different type of field in the corresponding operation data sample in the coding sequence; sets the determined bit, and obtains the coding result of the operation data sample according to the set; replaces the operation data sample with the operation The result of encoding the data sample.
  • each operation data sample includes whether to use function 1, whether to use function 2, whether to advance, or whether to escape data of 4 fields
  • initialize a 4-bit bit sequence [0, 0, 0, 0] to 1 means yes, 0 means no, the field corresponding to the different operation data samples is set, otherwise the state is not set, and the corresponding coding sequence is formed.
  • a scenario for preprocessing a scene data sample and an operation data sample included in a training sample when the artificial neural network model is trained, deleting the flag from the training sample set is invalid.
  • the training sample of the attribute, the training sample is selected from the remaining training samples marked as effective attributes to train the artificial neural network model, and the selected training samples may be unevenly distributed in the number of occurrences of the operation, which may lead to artificial nerves.
  • the network model has been over-learned for the operational skills of certain functions, and the operation skills of some functions are not well learned. Therefore, the training samples are balanced in the number of occurrences of different types of operations, thus making the artificial neural network model. It can fully learn the user's control skills for different types of functions, and will be described in combination with different equalization methods.
  • Equilibrium mode 1 For training samples whose statistics are marked as valid attributes, the number of occurrences of different types of operations in the statistical operation data samples is copied, and the training samples corresponding to the operations whose occurrence times are less than the order of magnitude are copied until the number of occurrences is lower than the order of magnitude.
  • the number of occurrences of the operation reaches an order of magnitude, where the order of magnitude can be a predetermined order of magnitude or the order of magnitude corresponding to the highest number of operations, and the balanced training samples can be used for training of the artificial neural network model.
  • the operation type involved in the training sample includes: using the operation data of the skill 1, using the operation data of the skill 2, using the operation data of the item 1, and using the operation data of the item 2; the corresponding operation times are 100, 1000, 1200, respectively. , 1100, then the training samples corresponding to the operation data of the skill 1 are randomly copied 900, that is, the total number reaches 1000, so as to be consistent with the training samples of other types of operations.
  • Equilibrium mode 2 For training samples whose statistics are marked as valid attributes, count the number of occurrences of different types of operations in the training samples, determine the types of operations whose operation times are higher than the order of magnitude, and randomly select from the training samples including the corresponding types of operations.
  • An order of magnitude training sample, along with an operation type whose number of operations is not higher than the order of magnitude, together with a training sample that includes an operation type that is not higher than the order of magnitude, can be used to train an artificial neural network model.
  • the operation type involved in the training sample includes: using the operation data of the skill 1, using the operation data of the skill 2, using the operation data of the item 1, using the operation data of the item 2; the corresponding operation times are 100, 150, 120 respectively 1000, then the training samples corresponding to the operation data of the props 2 are randomly selected 100 to be consistent with the training samples of other types of operations for the subsequent training operations of the artificial neural network model.
  • the scene data samples included in the pre-processed training sample set are input, and the artificial neural network model is trained using the operational data samples included in the pre-processed training sample set as outputs.
  • the artificial neural network model may adopt a type such as BP, RNN, etc., of course, not limited thereto.
  • the training samples may be subjected to adaptation processing, for example, in the training sample when using RNN.
  • the scene data sample needs to correspond to multiple image frames.
  • the basic structure of the initial artificial neural network includes an input layer, an intermediate layer and an output layer, initially including input of corresponding scene data samples, output of corresponding operational data samples, and loss function of artificial neural network model parameters; each time in the artificial neural network model In the iterative training process, the training function is selected according to the training sample selected from the training sample set, and the artificial neural network model parameters corresponding to the minimum value of the loss function are obtained, and the artificial neural network model parameters are updated according to the solved artificial neural network model parameters.
  • FIG. 1 is a schematic diagram of an optional application mode of the object processing method in the virtual scenario 100 according to the embodiment of the present invention, which is applicable to some devices that are completely dependent on the terminal device 200 .
  • the computing capability can complete the application mode of the related data calculation of the virtual scene, and the terminal device 200 outputs the application of the virtual scene, such as the game application and the shopping guide application of the stand-alone/offline mode described above, through the smartphone, the tablet and the virtual
  • the terminal device 200 such as a reality/augmented reality device completes the output of the virtual scene.
  • the object controlled by the user in the virtual scene hereinafter referred to as the third object, it can be understood that the third object is only convenient for the following description.
  • the third object and the first object may be the same object controlled by the user, or may be Different objects, for example, the user can choose to join different battle teams in the APRG so that the controlled game characters are different; in the process of the user controlling the interaction between the third object and the fourth object, in order to improve the intelligence degree of the fourth object
  • the artificial neural network model is embedded in the application, and the artificial neural network model is used to predict the real-time scene data of the fourth object. According to the operation data predicted to be executable by the fourth object, the operation skill of the user to control the first object is inherited. From the perspective of the user's perception of controlling the third object, it is found that the operational data of the fourth object is rich in expression and close to the real operation of the user.
  • FIG. 2 is a schematic diagram of an optional application mode of the object processing method in the virtual scenario 100 according to the embodiment of the present invention, which is applied to the terminal device 200/300 and
  • the server 400 generally, is adapted to perform virtual scene calculations depending on the computing power of the server 400, and output an application mode of the virtual scene at the terminal device 200/300.
  • the server 400 performs calculation of the virtual scene related display data and sends it to the terminal device 200/300.
  • the terminal device 200/300 relies on the graphics computing hardware to complete the loading, parsing and rendering of the calculated display data, and outputs the virtual scene according to the graphic output hardware to form
  • the terminal device 200 collects the real-time scene data of the user and sends it to the server 400, and the server 400 runs the artificial neural network.
  • the model uses an artificial neural network model to include real-time scene data of the third object for prediction. According to the operation data predicted to be executable by the fourth object, the operation skill of the user controlling the first object is inherited, and is perceived by the user controlling the third object. From the perspective, it will be found that the operational data of the fourth object is rich in expression and close to the real operation of the user.
  • FIG. 4 is an optional schematic flowchart of an object processing method in a virtual scene according to an embodiment of the present invention.
  • An exemplary implementation of the object processing method in the virtual scenario in the foregoing application scenario will be described below with reference to FIG.
  • corresponding real-time scene data is merged according to the collected position and the waiting time of the third object and the fourth object use function.
  • the operation 202 can be easily implemented, for example, the manner of non-stop acquisition for the interaction process, that is, the location of the object in the virtual environment is continuously collected during the entire interaction process. Waiting time, the method of performing acquisition for each image frame can form complete real-time scene data of the interaction process, or set the sampling point according to the acquisition precision for the interaction process, and collect in the window time of each sampling point of the interaction process,
  • the window of the sampling point generally includes a plurality of image frames, the positions of the third object and the fourth object are collected at the sampling point, and the waiting time of the third object and the fourth object using the function, and the collected results at one sampling point are combined to form a corresponding Real-time scene data of the sampling point.
  • the real-time scene data is subjected to pre-processing for adapting the artificial neural network model.
  • the pre-processing for the real-time scene data may be easily implemented, including the location of the third object included in the real-time scene data.
  • the positions of the four objects are mapped into the reference coordinate system; the distance components of the third object and the fourth object based on different coordinate axes are calculated in the reference coordinate system; the positions of the third object and the fourth object included in the real-time scene data are replaced by the distance Component.
  • the following preprocessing is performed on the real-time scene data: determining different types of fields included in the real-time scene data; normalizing different types of fields separately; normalizing the processed Each field data is connected as a component; the real-time scene data is replaced with a vector formed according to the connection.
  • the probability of controlling different action modes implemented by the fourth object and the probability of different function usage modes are predicted based on the pre-processed real-time scene data.
  • the middle layer of the artificial neural network model is composed of a series of nodes.
  • the real-time scene data is input to the middle layer through the input layer, and the input data is transformed by the excitation function in the nodes of the middle layer, and finally the output is transformed into different action modes.
  • different function usage methods and corresponding probabilities for example, in game applications, the probability of outputting forward, backward, jump, squat, etc., as well as ordinary attacks, release skills, The probability of using a function such as a prop.
  • an action mode that satisfies the probability condition and a function use mode are determined, and the real-time operation data of the fourth object is merged.
  • the combination Forming operational data to be executed for the fourth object; for the former may be used to form a response to a one-time operation of the third object, such as performing an action of the game character and implementing a function, for which the latter may be used
  • a response to a continuous operation of the third object is formed, such as a series of consecutive actions and functions of the game character.
  • real-time operational data corresponding to the fourth object is executed in the real-time virtual scene in response to the operation performed by the user controlling the third object during the interaction between the third object and the fourth object.
  • the terminal device 200 illustrated in FIG. 1 outputs a virtual scenario of an application, and controls an operation of the third object for the user in the virtual scenario, and invokes the operation interface to control the fourth object to execute real-time operation data, such as an action in the game, and the shopping guide application.
  • real-time operation data such as an action in the game, and the shopping guide application.
  • FIG. 5 is an optional structural diagram of an apparatus for processing an object in a virtual scene according to an embodiment of the present invention.
  • An optional structure of an apparatus according to an embodiment of the present invention will be described with reference to FIG. 5.
  • the apparatus 500 shown in FIG. 5 may include at least one processor 510, at least one communication bus 540, a user interface 530, at least one network interface 520, and a memory 550.
  • the various components in device 500 are coupled together via communication bus 540.
  • communication bus 540 is used to implement connection communication between these components.
  • the communication bus 540 includes a power bus, a control bus, and a status signal bus in addition to the data bus.
  • various buses are labeled as communication bus 540 in FIG.
  • the user interface 530 may include a display, a keyboard, a mouse, a trackball, a click wheel, a button, a button, a touch panel, or a touch screen.
  • Network interface 520 can include a standard wired interface, a wireless interface, and a typical wireless interface is a WiFi interface.
  • the memory 550 can be a high speed RAM memory or a non-Volatile memory, such as at least one disk memory. Memory 550 can also be at least one storage system remote from processor 510.
  • the memory 550 in the embodiment of the present invention is used to store various types of data to support the operation of the device 500. Examples of such data include: any computer program for operating on the device 500, such as an operating system, a network communication module, a user interface module, and a virtual scene function.
  • the program for implementing the object processing method in the virtual scene of the embodiment of the present invention may be Included in the virtual scenes application.
  • the object processing method in the virtual scene disclosed in the embodiment of the present invention may be applied to the processor 510 or implemented by the processor 510.
  • Processor 510 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each operation of the object processing method in the virtual scene may be completed by an integrated logic circuit of hardware in the processor 510 or an instruction in a form of software.
  • the processor 510 described above can be a general purpose processor, a DSP or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, or the like.
  • the processor 510 can implement or perform various methods, operations, and logic blocks provided in the embodiments of the present invention.
  • a general purpose processor can be a microprocessor or any conventional processor or the like.
  • the object processing method in the virtual scene provided by the embodiment of the present invention may be directly implemented as a hardware decoding processor, or may be executed by a combination of hardware and software modules in the decoding processor.
  • the software module may be located in a storage medium, and the storage medium is located in the memory 550.
  • the processor 510 reads the information in the memory 550 and combines the hardware to complete the object processing method in the virtual scene provided by the embodiment of the present invention.
  • the object processing method in the virtual scene provided by the embodiment of the present invention may be implemented by using the following virtual scene function module stored in the memory 550: an acquisition unit 551, a sample unit 552, and a pre-processing Unit 553 and model unit 554.
  • the collecting unit 551 is configured to collect an interaction process between the first object and the second object in the sample virtual scene.
  • the sample unit 552 is configured to construct a training sample according to the collected scene data samples and operation data samples, and combine the training samples constructed at different sampling points of the interaction process to form a training sample set;
  • the pre-processing unit 553 is configured to configure the training sample set to be pre-processed to adapt to the artificial neural network model
  • the model unit 554 is configured to input the scene data sample included in the pre-processed training sample set, and output the operation data sample included in the pre-processed training sample set as an output, and train the artificial neural network model to predict the corresponding according to the scene data. The performance of the operational data.
  • the collecting unit 551 is further configured to collect, at a sampling point of the interaction process, a location of the first object and the second object, and the first object and the first The waiting time of the second object using the function; combining the collected result of the sampling point to form a scene data sample of the corresponding sampling point.
  • the sample unit 552 is further configured to collect, at a sampling point of the interaction process, an attribute value indicating a state of the first object, and an attribute value indicating a state of the second object.
  • the attribute values collected at the sampling point, and the positions and waiting times collected at the corresponding sampling points are combined to form a scene data sample of the corresponding sampling point.
  • the sample unit 552 is further configured to acquire at least one of the following operational data of the first object at a sampling point of the interaction process: the executed controller operation data, the The pose mode of the first object and the manner in which the first object uses the function; combining the acquisition results of the sample points to form an operation data sample of the corresponding sample point.
  • the sample unit 552 is further configured to collect basic information of the interaction process, where the basic information includes an operation result of the first object in the interaction process, and the virtual scenario The size of the interaction process, and the scene data samples and operation data samples collected at different sampling points of the interaction process are combined to form a training sample of the corresponding sampling point.
  • the pre-processing unit 553 is further configured to determine, from an interaction process in which the training samples are derived from the training sample set, an effective interaction process that satisfies the following conditions: The sum of the first object and the second object cumulative usage function times exceeds the overall function usage count threshold; and the training samples not belonging to the valid interaction process are deleted from the training sample set.
  • the pre-processing unit 553 is further configured to acquire an operation result included in each training sample in the training sample set; when the operation result indicates that the first object is in the interaction process When the operation result reaches the target, the effective attribute of the training sample is marked; when the operation result indicates that the operation result of the first object in the interaction process does not reach the target, the invalid attribute of the training sample is marked.
  • the pre-processing unit 553 is further configured to: when the operation result indicates that the operation result of the first object in the interaction process reaches a target, and the The number of times the function of the cumulative use of an object exceeds the threshold of the number of times the function uses the first object, and the effective attribute of the training sample is marked.
  • the pre-processing unit 553 is further configured to: when the operation result indicates that the operation result of the first object in the interaction process reaches a target, and the The number of times the function of the cumulative use of an object does not exceed the threshold of the number of times the function uses the first object, and the invalid attribute of the training sample is marked.
  • the pre-processing unit 553 is further configured to perform the following pre-processing on the scene data samples included in each training sample in the training sample set:
  • the position of the first object and the second object included in the training sample is replaced with the distance component.
  • the pre-processing unit 553 is further configured to perform the following pre-processing on the scene data samples included in each training sample in the training sample set:
  • the normalized field data is connected as a component
  • the scene data samples in the training samples are replaced with vectors formed by the component connections.
  • the pre-processing unit 553 is further configured to compare the distance component with a component of the virtual scene in a direction of a corresponding coordinate axis when the type of the field is a distance component. Calculating, according to the ratio operation, obtaining a normalized result of the distance component in the direction of the corresponding coordinate axis,
  • the waiting time is compared with the sum of the waiting times of the corresponding functions in the training sample set, and the corresponding waiting time is obtained according to the ratio operation.
  • the attribute value is compared with the sum of the same type of attribute values in the training sample set, and the normalized result of the attribute value is obtained according to the ratio operation.
  • the pre-processing unit 553 is further configured to perform the following pre-processing on the operation data samples included in each training sample in the training sample set:
  • the determined action mode is substituted for controller operation data in the training sample.
  • the pre-processing unit 553 is further configured to: remove training samples marked as invalid attributes from the training sample set; and statistically mark different types of operations in the operational data samples in the training samples of the valid attributes The number of occurrences, and performing at least one of the following pre-processing: copying the training samples corresponding to the operations whose number of occurrences is less than the order of magnitude until the number of occurrences of the number of occurrences less than the order of magnitude reaches the order of magnitude; The samples corresponding to the operations of the order of magnitude are randomly selected, and the number of selected training samples conforms to the order of magnitude.
  • the pre-processing unit 553 is further configured to perform the following pre-processing on the operation data samples included in each training sample in the training sample set:
  • the model unit 554 is further configured to initialize an input layer, an intermediate layer, and an output layer of the artificial neural network; initially including the input, the output, and an artificial neural network model parameter.
  • a loss function in the iterative training process of the artificial neural network model, the training function is selected according to the training sample selected from the training sample set, and the artificial neural network model parameter corresponding to the minimum value of the loss function is obtained;
  • the artificial neural network model is updated according to the solved artificial neural network model parameters.
  • the collecting unit 551 is further configured to collect an interaction process between the third object and the fourth object in the real-time virtual scene;
  • the pre-processing unit 553 is further configured to perform pre-processing configured to adapt the artificial neural network model according to the collected real-time scenario data;
  • the pre-processing unit 553 is further configured to predict, in the artificial neural network model, a probability of controlling different action modes implemented by the fourth object according to the pre-processed real-time scene data, and different function usage manners. Probability; determining an action mode that satisfies a probability condition and a function usage manner, and combining to form real-time operational data of the fourth object;
  • the model unit 554 is further configured to execute real-time operation data corresponding to the fourth object in the real-time virtual scene in response to an interaction process between the third object and the fourth object.
  • the collecting unit 551 is further configured to: according to the collected positions of the third object and the fourth object, and the third object and the fourth object using functions Waiting time, combining to form corresponding real-time scene data; collecting the positions of the third object and the fourth object at the sampling point of the interaction process, and waiting for the third object and the fourth object to use the function Time; combining the acquisition results of the sampling points to form real-time scene data of corresponding sampling points.
  • the pre-processing unit 553 is further configured to perform the following pre-processing on the real-time scene data: the location of the third object included in the real-time scene data and the fourth Mapping a position of the object into a reference coordinate system; calculating a distance component of the third object and the fourth object based on different coordinate axes in the reference coordinate system; and the third object included in the real-time scene data The position with the fourth object is replaced with the distance component.
  • the pre-processing unit 553 is further configured to perform the following pre-processing on the real-time scene data: determining different types of fields included in the real-time scene data; and using the different types of fields Perform normalization processing separately; connect each normalized field data as a component; replace the real-time scene data with a vector formed according to the connection.
  • the model unit 554 is further configured to transmit the real-time scene data to an intermediate layer of the artificial neural network model at an input layer of the artificial neural network model;
  • the middle layer in the neural network model transforms the input real-time scene data according to the excitation function of the middle layer node to form different types of action modes and corresponding probabilities, and forms different types of function usage modes and corresponding probabilities;
  • the output layer of the artificial neural network outputs an action mode that satisfies the probability condition and a function use mode.
  • the device 400 in the embodiment of the present invention may be implemented as a terminal device, such as the terminal 200 shown in FIG. 1, where the terminal device 200 runs a virtual scene application, where the user controls the first object and the virtual scene.
  • the operation data sample in the virtual scene output by the user and the virtual scene data sample are used to construct the training sample, and the artificial neural network model is trained after the training sample is preprocessed;
  • the object (which may be the same object as the first object or the object different from the first object) and the fourth object (for example, the robot controlled by the virtual scene application may be the same object as the second object, also
  • the operation data of the fourth object is predicted according to the real-time scene data of the virtual scene, and the interface of the virtual scene application is invoked to control the fourth object to implement the operation data to form a user control.
  • the response of the operation performed by the third object is performed by the third object.
  • the device 400 provided by the embodiment of the present invention may be implemented as a server, such as the server 400 shown in FIG. 2, where the terminal 200/300 runs a virtual scenario application (as a client), and collects a user to control the first object.
  • the process of interacting with the second object in the virtual scene includes collecting operation data samples of the user in the output virtual scene, and constructing training samples of the virtual scene data samples, transmitting to the server 400, and the server 400 performs pre-processing training on the training samples.
  • the artificial neural network model, the trained machine learning model is synchronized to the virtual scene application of the terminal 200/300; when the user operates the object third object (which may be the same object as the first object, may also be the first object)
  • the object third object which may be the same object as the first object, may also be the first object
  • the fourth object the robot controlled by the server 400, which may be the same object as the second object or the object different from the second object
  • the real-time scene data of the virtual scene is predicted.
  • the operation data of the fourth object is implemented, and the interface of the virtual scene application is invoked to control the fourth object to implement real-time For data formed in response to user control actions of the third object of the embodiment.
  • the artificial neural network model is realized by learning the user's operation, without paying attention to the specific implementation logic of the behavior, the expression form is rich, the user operation is close, the turnover rate is low, the decision speed is fast, and the battle is fast. The level can exceed the user being studied.
  • the embodiment of the present invention further provides a storage medium, where an executable program is stored, and when the executable program is executed by the processor, the object processing method in the virtual environment provided by the embodiment of the present invention is implemented, for example, FIG. 3 or FIG.
  • the storage medium provided in the embodiment of the present invention may be a storage medium such as an optical disk, a flash memory, or a magnetic disk, and may be a non-transitory storage medium.
  • FIG. 6 is an optional schematic diagram of training and application of an artificial neural network model according to an embodiment of the present invention, and relates to training and application of an artificial neural network model.
  • the following parts are involved: (a) collecting training samples; (b) pre-processing the training samples; (c) using the pre-processed sample-pair models according to the artificial neural network algorithm Training; will be explained separately.
  • the real player controls the game character to fight, records basic game information, real-time information of the scene, and operational data samples of the real player during the battle, and the recorded data combination is obtained.
  • the data set is used as a training sample.
  • the real player performs the ARPG game combat operation on the mobile device or computer, collects the basic information of the game during the battle, and performs multiple sampling in each game, each sample point has a certain window time, and each sample point collects the scene data. A sample of the sample and the actual player's operational data.
  • the basic information of the game includes: whether the game character controlled by the real player wins, the size of the scene (the scene, that is, the space for the object to move in the game, such as the platform in the battle game), the total cooling time of each skill; the cooling time refers to continuous
  • the time required to use the same skill (or item), referred to as the CD the sum of the CDs of each skill of the player-controlled game character, which is the total time of the corresponding skill.
  • the scene data includes: the position of our (ie, the position of the game character controlled by the real player), the position of the enemy (ie, the game character that is fighting the game character controlled by the real player), the current CD of each skill of our skill, and the enemy.
  • the current CD of each skill there is a corresponding set of the above data in each sample.
  • the operation data includes: the skill usage mode (whether each skill is released, whether a normal attack is performed), the joystick movement angle, whether the game character controlled by the real player jumps; it should be noted that the joystick movement angle is only when the joystick is operated. The corresponding data will be collected. If the real player does not operate the joystick within the window time of one sampling point, the data of the joystick movement angle will not be collected at this time. Since the operation data is incomplete, the data collected by the sampling point will be Will be discarded to ensure that the resulting training samples include rocker angle data.
  • the collected training samples should be as full as possible. Take the 1v1 battle scene of an ARPG mobile game as an example. It is necessary to collect more than 30 rounds (game rounds) of scene data samples and The data sample is manipulated, taking 20 samples per game as an example, and after screening, about 500 effective training samples.
  • Pre-processing the training samples. Pre-processing operations such as screening, data conversion, and class balancing are performed on the collected samples.
  • the effective data screening involves: selecting the training samples collected in the final winning game round, and the number of skill releases in the game round (the total number of skill releases by our side and the enemy) is greater than the threshold of the total number of times.
  • the threshold of an ARPG mobile game is 20 times.
  • an artificial neural network model based on an artificial neural network is mainly obtained by learning a training sample, and thus a game round that is selected by a real player to collect a training sample, and the total number of times the game rounds the actual number of skill releases by us and the enemy is obtained.
  • the threshold of the total number of times is greater than the threshold of the artificial neural network model.
  • the position conversion involves: converting the position of our training and the position of the enemy into the distance between us and the enemy, and the conversion formula is:
  • the location information is mainly used to describe the positional relationship between us and the enemy: whether it is within the scope of attack and attack. Therefore, using the distance between us and the enemy is more direct, and can reduce the information dimension and reduce the complexity of the model.
  • normalization involves normalizing the scene data to [0, 1], respectively, as follows:
  • enemy skill 1CD enemy skill 1CD / skill 1 cooling total time
  • enemy skill 2CD enemy skill 2CD / skill 2 cooling total time
  • the rocker angle data conversion involves: converting the rocker movement angle into two operations of chasing the enemy and escaping the enemy.
  • FIG. 7 is a moving angle of the joystick provided by the embodiment of the present invention.
  • a multi-frame game screen is outputted during the sampling window time, so that the formed training samples include rocker angle data corresponding to the multi-frame picture, and the joystick included in each training sample is included.
  • the angle data judges whether the rocker angle data of more than a frame picture is included; if the rocker angle data of more than a frame picture is included, the position of our and the enemy in the training sample is determined:
  • the distance between our and the enemy in the a frame is reduced, it is determined whether the distance between our enemy and the enemy in the a frame exceeds the distance threshold. If the threshold is exceeded, we use the action mode of the pursuit place. If not exceeded, discard the joystick movement angle data.
  • the role of the threshold a is to filter out the situation that the real player is accidentally touched or the operation is not obvious.
  • the functions of the thresholds b and c are to filter out the joystick operation in which the player's intention is not obvious.
  • encoding the operation data with One-Hot involves: serializing the operation data as [whether skill 1 is released, skill 2 is released, ..., whether a normal attack is performed, whether the enemy is pursued, whether or not the enemy is fleeing from the enemy , whether to jump], the bit corresponding to the player's operation is set to 1, the rest is 0. If the player releases skill 2, it is encoded as [0,1,...,0,0,0,0].
  • the total number of times of each operation is counted according to the operation data of all the training samples, the training samples corresponding to the operations with a larger total number of times are randomly sampled, or the training samples of the operations with less total number of times are randomly copied, so that each operation is performed.
  • the total number of times is the same order of magnitude and the highest level is equal, which is beneficial to reduce the complexity of model training and reduce or avoid the tuning of the output of artificial neural network model.
  • the artificial neural network model is trained using the pre-processed training samples, wherein the operational data is used as an output, and the corresponding scene data before the operation is taken as an input, as follows:
  • FIG. 8A is an optional structural schematic diagram of an artificial neural network model according to an embodiment of the present invention, and an artificial neural network.
  • the training of the model is performed on the device according to the BP neural network algorithm; of course, for the artificial neural network algorithm to be used, in addition to the BP neural network, other artificial neural network algorithms, such as the RNN algorithm, in which the input of the input RNN model is required, may be used. Adjust to scene data samples for consecutive multi-frames in the game.
  • FIG. 8B is an optional schematic diagram of an artificial neural network according to an embodiment of the present invention for predicting operational data according to real-time scene data, and relates to an application phase of an artificial neural network model.
  • the following parts (a) real-time acquisition of scene data during combat; (b) pre-processing of scene data; (c) inputting pre-processed scene data into an artificial neural network model, and calculating the output of the model Operation data; (d) Invoking the corresponding game interface according to the operation of the model output, so that the game character completes the corresponding game behavior;
  • the game program obtains the scene data in the battle process in real time, which is consistent with the collection of the training samples.
  • the scene data mainly includes: our position, enemy position, our skill CD, and enemy skill CD.
  • the scene data is preprocessed in the game program, which is consistent with the preprocessing of the training samples, including the position of ours, the enemy position converted to the distance between us and the enemy, and the normalization of the scene data.
  • the pre-processed scene data is taken as input, and the output is calculated by the neural network algorithm, that is, the decision of the artificial neural network model.
  • the BP neural network algorithm in the artificial neural network is used, and the output is calculated from the input information in the game program according to the algorithm operation.
  • the output of the artificial neural network model is a set of numbers, corresponding to [whether skill 1 is released, skill 2 is released, ..., whether it is a normal attack, whether it is chasing the enemy, whether it flees the enemy, whether it jumps or not], according to the output
  • the game interface is called to execute the game operation corresponding to the maximum value in the output.
  • the execution strategy of chasing the enemy and fleeing the enemy may be further optimized according to the characteristics of the game.
  • the execution strategy may adopt a forward/reverse corresponding route according to different games.
  • the pursuit of the enemy operation adopts a strategy of preferentially moving toward the enemy's horizontal axis: Take our position as the coordinate origin, and the moving direction is (the enemy's horizontal axis coordinate/2, the enemy's longitudinal axis coordinate).
  • the training data is used to transform the operational data into the skill usage mode of the decision execution, the pursuit of the enemy, the action mode of fleeing the enemy, and the like, and the artificial neural network. Models are more likely to achieve higher levels of combat.
  • the input and output of artificial neural network models should be adjusted accordingly for different ARPG battle scenarios: thus, it has strong versatility and low technical cost. .
  • the number of model skills should be the same as the actual number of skills in the game; if the skill consumption is blue, the amount of blue between us and the enemy should be added to the input; if the game character cannot jump, remove it from the operation data. Jump operation; according to the specific game situation, if the decision needs to refer to the blood volume of our side and place, then the sample can be collected into the blood volume of our side and the enemy, so that the artificial neural network model can refer to the blood for the operation data decision; In addition, the type of the field that remains in place can be added to the operation data, and the type of the field of the enemy skill CD in the input can also be removed from the operation data. Of course, the artificial neural network model trained in this way is implemented. The level of combat will be reduced to some extent.
  • FIG. 9A is an optional structure of the artificial neural network model provided by the embodiment of the present invention.
  • FIG. 9A is an optional structure of the artificial neural network model provided by the embodiment of the present invention.
  • 9B is an optional schematic diagram of an artificial neural network according to an embodiment of the present invention for predicting operational data according to real-time scene data, a three-dimensional modeling of a store in a virtual scene output by a shopping guide application, and further including a first object and a second
  • the object the first object is a customer role controlled by the user through the virtual reality device, is a rendering of the user image in the virtual environment by using the three-dimensional modeling technology, and the second object is a shopping guide realized in the virtual environment based on the three-dimensional modeling technology.
  • the guide must have a wealth of product knowledge and recommended product skills.
  • the acquisition of the employee with rich sales experience controls the shopping guide in the virtual environment to interact with the first object controlled by the customer.
  • each service process of the purchaser for the customer including collecting the customer purchase or the customer leaving as the basic result of the interaction process, collecting the time required by the shopping guide to introduce different commodities, the time when the customer pays attention to different commodities, the purchaser and the customer
  • the distance forms a sample of the scene data, collects the actual purchase situation of the customer, the action mode of the purchaser in the interaction process, and the situation of the recommended product to construct an operation data sample, and then trains the artificial neural network model through preprocessing, so that the artificial neural network model learns the salesperson's Product knowledge and sales skills.
  • the artificial neural network model is used to control the shopping guide in the virtual scene. After the customer character enters the store, the location of the shopping guide and the customer role is collected in real time, the time of the different products after the customer enters the store, and the time required by the shopping guide to introduce the product.
  • the artificial neural network model is used to calculate the probability of recommending different commodities and the probability of the action mode adopted.
  • the guide character will move according to the customer's role, and continue to collect the customer's role to pay attention to different products; if the recommended product 2 has the highest probability, the shopping guide controls the shopping guide to introduce the customer to the product 2 through the control interface of the application. .
  • the artificial neural network model is realized by learning the operational data of real users, without specific logic, rich representation, close to the operation of real users, and artificial neural network algorithm has better The anti-noise ability is fast, and the calculation speed is fast when applied. Therefore, the artificial neural network model rarely has operational errors like real users, and the decision speed is much faster than the real user's response, and its behavior level is higher than that of the learned user.

Landscapes

  • Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • General Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Marketing (AREA)
  • Computer Graphics (AREA)
  • Computer Hardware Design (AREA)
  • Economics (AREA)
  • Development Economics (AREA)
  • Human Computer Interaction (AREA)
  • Processing Or Creating Images (AREA)

Abstract

La présente invention porte sur un procédé de traitement d'objet dans une scène virtuelle, sur un dispositif ainsi que sur un support d'informations, ledit procédé consistant : à acquérir un processus d'interaction entre un premier objet et un second objet dans une scène virtuelle d'échantillon; à construire des échantillons d'apprentissage en fonction des échantillons de données de scène acquis et des échantillons de données d'opération et à combiner les échantillons d'apprentissage construits à différents points d'échantillonnage dans le processus d'interaction de sorte à former un ensemble d'échantillons d'apprentissage; à prétraiter l'ensemble d'échantillons d'apprentissage; et à former un modèle de réseau neuronal artificiel en prenant les échantillons de données de scène inclus dans l'ensemble d'échantillons d'apprentissage prétraité en tant qu'entrée et en prenant les échantillons de données d'opération inclus dans l'ensemble d'échantillons d'apprentissage prétraité en tant que sortie.
PCT/CN2018/074156 2018-01-25 2018-01-25 Procédé de traitement d'objet dans une scène virtuelle, dispositif et support d'informations WO2019144346A1 (fr)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2018/074156 WO2019144346A1 (fr) 2018-01-25 2018-01-25 Procédé de traitement d'objet dans une scène virtuelle, dispositif et support d'informations
CN201880003364.1A CN110325965B (zh) 2018-01-25 2018-01-25 虚拟场景中的对象处理方法、设备及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/074156 WO2019144346A1 (fr) 2018-01-25 2018-01-25 Procédé de traitement d'objet dans une scène virtuelle, dispositif et support d'informations

Publications (1)

Publication Number Publication Date
WO2019144346A1 true WO2019144346A1 (fr) 2019-08-01

Family

ID=67394493

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/074156 WO2019144346A1 (fr) 2018-01-25 2018-01-25 Procédé de traitement d'objet dans une scène virtuelle, dispositif et support d'informations

Country Status (2)

Country Link
CN (1) CN110325965B (fr)
WO (1) WO2019144346A1 (fr)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112245921B (zh) * 2020-11-17 2022-04-15 腾讯科技（深圳）有限公司 虚拟对象控制方法、装置、设备及存储介质
CN112947809A (zh) * 2021-01-29 2021-06-11 北京三快在线科技有限公司 知识学习方法、装置、电子设备
CN114442888A (zh) * 2022-02-08 2022-05-06 联想(北京)有限公司 对象确定方法、装置及电子设备
CN115953706B (zh) * 2023-01-09 2024-04-26 支付宝(杭州)信息技术有限公司 虚拟形象处理方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105867932A (zh) * 2016-04-07 2016-08-17 浙江工业大学 基于元数据反射链的游戏云设计方法
CN106502890A (zh) * 2016-10-18 2017-03-15 乐视控股（北京）有限公司 测试用例生成方法及系统
CN107102727A (zh) * 2017-03-17 2017-08-29 武汉理工大学 基于elm神经网络的动态手势学习与识别方法
CN107291232A (zh) * 2017-06-20 2017-10-24 深圳市泽科科技有限公司 一种基于深度学习与大数据的体感游戏交互方法及系统

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104866860A (zh) * 2015-03-20 2015-08-26 武汉工程大学 一种室内人体行为识别方法
CN106503787B (zh) * 2016-10-26 2019-02-05 腾讯科技（深圳）有限公司 一种获取游戏数据的方法及电子设备
CN106897738B (zh) * 2017-01-22 2019-07-16 华南理工大学 一种基于半监督学习的行人检测方法
CN106909891B (zh) * 2017-01-24 2019-11-15 华南农业大学 一种基于自反馈基因表达式编程的人体行为识别方法
CN107168530A (zh) * 2017-04-26 2017-09-15 腾讯科技（深圳）有限公司 虚拟场景中的对象处理方法和装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105867932A (zh) * 2016-04-07 2016-08-17 浙江工业大学 基于元数据反射链的游戏云设计方法
CN106502890A (zh) * 2016-10-18 2017-03-15 乐视控股（北京）有限公司 测试用例生成方法及系统
CN107102727A (zh) * 2017-03-17 2017-08-29 武汉理工大学 基于elm神经网络的动态手势学习与识别方法
CN107291232A (zh) * 2017-06-20 2017-10-24 深圳市泽科科技有限公司 一种基于深度学习与大数据的体感游戏交互方法及系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LUO, LINBO ET AL.: "Towards a Data-Driven Approach to Scenario Generation for Serious Games", COMPUTER ANIMATION AND VIRTUAL WORLD, vol. 25, 19 May 2014 (2014-05-19), pages 393 - 402, XP055627614 *

Also Published As

Publication number Publication date
CN110325965B (zh) 2021-01-01
CN110325965A (zh) 2019-10-11

Similar Documents

Publication Publication Date Title
US11291917B2 (en) Artificial intelligence (AI) model training using cloud gaming network
KR102127670B1 (ko) 멀티플레이어 비디오 게임 매치메이킹 시스템 및 방법들
US20210374538A1 (en) Reinforcement learning using target neural networks
WO2019144346A1 (fr) Procédé de traitement d'objet dans une scène virtuelle, dispositif et support d'informations
US11213746B2 (en) Providing automated user input to an application during a disruption
JP7447299B2 (ja) 仮想シーンのアダプティブ表示方法及び装置、電子機器、並びにコンピュータプログラム
CN113633983A (zh) 虚拟角色表情控制的方法、装置、电子设备及介质
CN114307160A (zh) 训练智能体的方法
US10279267B2 (en) Monitoring game activity to detect a surrogate computer program
CN114125529A (zh) 一种生成和演示视频的方法、设备及存储介质
CN112138394A (zh) 图像处理方法、装置、电子设备及计算机可读存储介质
CN113509726B (zh) 交互模型训练方法、装置、计算机设备和存储介质
CN109445573A (zh) 一种用于虚拟化身形象互动的方法与装置
KR20210013529A (ko) 사용자 체험을 관리하기 위한 시스템 및 그 방법
WO2024012007A1 (fr) Procédé et appareil de génération de données d'animation, et produit associé
CN116943220A (zh) 一种游戏人工智能控制方法、装置、设备及存储介质
CN116943204A (zh) 虚拟对象的控制方法、装置和存储介质及电子设备
CN113360343B (zh) 分析内存占用状况的方法、装置、存储介质及计算机设备
JP2024505320A (ja) 人工知能を対象とする効率的なゲームプレイ訓練
EP4385592A1 (fr) Procédé mis en uvre par ordinateur pour commander un avatar virtuel
JP7008970B2 (ja) ゲーム装置、ゲーム遂行方法、およびプログラム
WO2022229639A2 (fr) Procédé mis en œuvre par ordinateur de commande d'avatar virtuel
CN118036694A (zh) 一种训练智能体的方法、装置和设备及计算机存储介质
CN117482528A (zh) 虚拟场景的汇总信息处理方法、装置、设备及存储介质
CN111359212A (zh) 游戏对象控制、模型训练方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18902160

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18902160

Country of ref document: EP

Kind code of ref document: A1