CN109224442A

CN109224442A - Data processing method, device and the storage medium of virtual scene

Info

Publication number: CN109224442A
Application number: CN201811022991.8A
Authority: CN
Inventors: 杨夏; 王洁梅; 周大军; 张力柯; 荆彦青
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2018-09-03
Filing date: 2018-09-03
Publication date: 2019-01-18
Anticipated expiration: 2038-09-03
Also published as: CN109224442B

Abstract

The present invention provides data processing method, device and the storage medium of a kind of virtual scene, and method includes: to identify to the image of virtual scene, with the stage locating for the determination virtual scene；According to different phase locating for the virtual scene, by the way of calling machine learning model or condition triggering, the object run of different phase locating for the corresponding virtual scene is determined；Control executes the corresponding object run of moment locating for the virtual scene.

Description

Data processing method, device and the storage medium of virtual scene

Technical field

The present invention relates to data processing technique more particularly to a kind of data processing methods of virtual scene, device and storage Medium.

Background technique

Based on the display technology of graphics process hardware, extends perception environment and obtain the channel of information, it is especially empty The display technology of quasi- scene, can be according to various virtual objects in practical application request realization person to person, people and virtual scene Intelligentized interaction.

Game is the typical case of virtual scene display technology, and user can be exported by equipment running game in equipment Virtual scene in, the game object of user's control cooperateed with other game objects on line fight or battle.

In the related technology, for the processing in scene of game using deeply learning algorithm training artificial intelligence (AI, Artificial Intelligence) it realizes, however, due to scene of game complexity, some scene of game (such as cut scene) Deeply learning algorithm is not particularly suited for solve.

Summary of the invention

The embodiment of the present invention provides data processing method, device and the storage medium of a kind of virtual scene, can be for not Stage locating for same virtual scene uses different processing modes, improves treatment effeciency.

The technical solution of the embodiment of the present invention is achieved in that

In a first aspect, the embodiment of the present invention provides a kind of data processing method of virtual scene, comprising:

The image of virtual scene is identified, with the stage locating for the determination virtual scene；

According to different phase locating for the virtual scene, by the way of calling machine learning model or condition triggering, Determine the object run of different phase locating for the corresponding virtual scene；

Control executes the corresponding object run of moment locating for the virtual scene.

Second aspect, the embodiment of the present invention provide a kind of data processing equipment of virtual scene, comprising:

Recognition unit is identified for the image to virtual scene, with the stage locating for the determination virtual scene；

Determination unit is touched for the different phase according to locating for the virtual scene using machine learning model or condition The mode of hair determines the object run of different phase locating for the corresponding virtual scene；

Execution unit executes the corresponding object run of moment locating for the virtual scene for controlling.

The third aspect, the embodiment of the present invention provide a kind of data processing equipment of virtual scene, comprising:

Memory, for storing executable instruction；

Processor when for executing the executable instruction stored in the memory, is realized provided in an embodiment of the present invention The data processing method of the virtual scene.

Fourth aspect, the embodiment of the present invention provide a kind of storage medium, are stored with executable instruction, the executable instruction For causing memory to execute the data processing method of the virtual scene provided in an embodiment of the present invention.

The embodiment of the present invention has the advantages that

1) it is identified by the image to virtual scene, determines different phase locating for virtual scene, that is to say, that is empty Quasi- scene is divided into the different stages, makes it possible to targetedly be handled for the different stages；

2) different phase according to locating for virtual scene, using the side for calling deep neural network model or condition triggering Formula determines the object run of different phase locating for corresponding virtual scene；The corresponding different stage, for calculating complexity not Together, object run is determined by the way of calling deep neural network model or condition triggering according to different computation complexities, The efficiency of determining object run is improved, meanwhile, so that deep neural network model is more targeted in the training process, training Speed is fast.

Detailed description of the invention

Fig. 1 is a schematic diagram of DQN model provided in an embodiment of the present invention；

Fig. 2 is the application model schematic diagram one of the data processing method of virtual scene provided in an embodiment of the present invention；

Fig. 3 is the application model schematic diagram two of the data processing method of virtual scene provided in an embodiment of the present invention；

Fig. 4 is the configuration diagram of the data processing system of virtual scene provided in an embodiment of the present invention；

Fig. 5 is the composed structure schematic diagram one of the data processing equipment of virtual scene provided in an embodiment of the present invention；

Fig. 6 is the flow diagram one of the data processing method of virtual scene provided in an embodiment of the present invention；

Fig. 7 is the flow diagram two of the data processing method of virtual scene provided in an embodiment of the present invention；

Fig. 8 is a schematic diagram of DQN model provided in an embodiment of the present invention；

Fig. 9 is the flow diagram three of the data processing method of virtual scene provided in an embodiment of the present invention；

Figure 10 is the schematic diagram of the image recognition of various dimensions provided in an embodiment of the present invention；

Figure 11 is the schematic diagram provided in an embodiment of the present invention for carrying out vital values identification；

Figure 12 is the schematic diagram of game keys provided in an embodiment of the present invention；

Figure 13 is the schematic diagram of DQN deep neural network model provided in an embodiment of the present invention；

Figure 14 is the method flow schematic diagram of on-line training DQN model provided in an embodiment of the present invention；

Figure 15 is the composed structure schematic diagram two of the data processing equipment of virtual scene provided in an embodiment of the present invention.

Specific embodiment

To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing to the present invention make into It is described in detail to one step, described embodiments are some of the embodiments of the present invention, instead of all the embodiments.According to this hair Embodiment in bright, all other implementation obtained by those of ordinary skill in the art without making creative efforts Example, shall fall within the protection scope of the present invention.

It should be noted that in embodiments of the present invention, the terms "include", "comprise" or its any other variant are intended to Cover non-exclusive inclusion, so that including the method for a series of elements or device not only includes wanting of clearly providing Element, but also including other elements that are not explicitly listed, or further include for implementation method or device intrinsic want Element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that including the element Method or device in there is also other relevant factor (such as the step in method or the unit in device, here Unit can be partial circuit, segment processor, subprogram or software etc.).

A series of step is contained for example, provided in an embodiment of the present invention, but method provided in an embodiment of the present invention It is not limited to provided step, similarly, device provided in an embodiment of the present invention includes a series of units, but the present invention is real The device for applying example offer is not limited to include the unit clearly provided, can also include to obtain relevant information or being based on information The unit of required setting when being handled.

In addition, " first ", " second ", " third " documented by the embodiment of the present invention are only used for distinguishing different objects, The difference of sequence or priority is not represented, it will be understood that in the absence of conflict, " first ", " second ", " third " institute's generation The object of table can be interchanged.

Before the present invention will be described in further detail, noun involved in the embodiment of the present invention and term are said Bright, noun involved in the embodiment of the present invention and term are suitable for following explanation.

1) AI refers to the intelligence shown by made system, refers to pair of the virtual scenes such as game herein As and/or scenery control process artificial intelligence.

2) machine learning (Machine Learning) model, (referred to as from the image pattern of the virtual scene of training set Sample) model training is carried out, the model after making training has according to the sample characteristics of sample outside training set to corresponding virtual scene The performance predicted of object run；In actual implementation, deep neural network model can be used, this is a kind of mimic biology The mathematical model of neural network structure and function, deep neural network model herein can for depth Q network model (DQN, Deep Q-Network) or asynchronous advantage actor reviewer algorithm model (A3C, Asynchronous Advantage Actor-Critic), deep neural network model is used for Function Estimation or approximation, including input layer, middle layer and output layer, often A layer is interconnected by a large amount of processing units and is constituted, and each node is handled and exported using data of the excitation function to input Other nodes are given, the illustrative type of excitation function includes threshold-type, lienar for and S growth curve (Sigmoid) type etc..

3) virtual scene, the scene for being different from real world exported using equipment, passes through naked eye or the auxiliary energy of equipment Enough visual perceptions formed to virtual scene, such as by the bidimensional image of display screen output, by stereoprojection, virtual existing The stereo display techniques such as real and augmented reality are come the 3-dimensional image that exports；Further, it is also possible to pass through various possible hardware Form the perception of the various simulation real worlds such as Auditory Perception, tactilely-perceptible, detected by scent and motion perception；A void herein The pictured scene that the quasi- corresponding frame image of scene is presented, scene switching correspond to the switching of the frame image of virtual scene.

4) object, the image of the various people and object that can interact in virtual scene, such as " soul bucket sieve: coming back " game Virtual scene in various game roles, the shop-assistant etc. in shop that virtual reality technology is realized.

5) in response to the condition or state relied on for indicating performed operation, when the relied on condition of satisfaction Or when state, performed one or more operations be can be in real time, it is possible to have the delay of setting；Do not saying especially In the case where bright, there is no the limitations for executing sequencing for performed multiple operations.

6) image recognition refers to and is handled image, analyzed and understood using computer, to identify various different modes Target and technology to picture.

Before the data processing scheme to virtual scene provided in an embodiment of the present invention is illustrated, the present invention is implemented DQN model in example is illustrated.

DQN model is to combine deep learning with intensified learning to realize from perception (perception) to movement (action) a kind of algorithm of end-to-end (end-to-end) study, is convolutional neural networks (C NN, Convolutional Neural Network) and Q-Learning combination, the input of CNN is raw image data (as state state), output It is then the corresponding value assessment Value Function (Q value) of each movement action.

DQN with a deep neural network as the network of Q value, parameter ω,

Q(s,a,ω)≈Q_π(s,a)；

It is applicable in mean square deviation mean-square error in Q value and comes objective function, i.e. loss function loss Function,

Wherein, s and a is respectively next state and movement, has used the table of depth enhancing study David Silver here Show mode, it appears that than more visible.It can be seen that Q value that Q-Learning to be updated is used herein as target value.Have Target value, and have current value, then deviation can be calculated by mean square deviation.

Calculate the gradient about loss function:

Then, end-to-end optimization is realized using stochastic gradient descent algorithm SGD, there is gradient above, by from depth It is calculated in degree neural networkStochastic gradient descent can be used to carry out parameter update, to obtain optimal Q Value.

Fig. 1 is the schematic diagram of DQN model provided in an embodiment of the present invention, and referring to Fig. 1, input is by gray proces The image of nearest 4 frame 84 × 84 is followed by two full articulamentums by several convolutional layers, and output is the Q value that each may be acted. After netinit, then input 84 × 84 × 4 calculates network output, selects a random movement according to probability (very little), Or the Q value that each movement is calculated in current network is input to according to current state, select the maximum movement of Q value (optimal movement)；The reward after executing optimal movement and the input (4 frame image) of next network are obtained, 4 frame images are made For state this moment, corresponding target value (Q value is updated as target value by the reward after the optimal movement of execution) is calculated, so Afterwards by SGD undated parameter weight, and iteration updates target action (target action)-Q value (value function) Parameter.

Inventor has found in the course of the research, in one embodiment, the realization game of vocal imitation skill can be inscribed based on game AI obtains game data (state) by In-game interface, thus the input as AI, however this needs game externally to provide Data-interface and operation interface, needing cooperation or the game of game developer itself is that open source provides, this swims many business It is infeasible for play.

In one embodiment, game AI can be realized based on game scripts, held by the action script of several fixations Movement in row game, thus realize AI function, however due to being rule-based and script, it is only suitable for some easy game scenes Test, in complicated scene of game, the state space that can be explored is very big, and rule is impossible to exhaust completely, and encounters random sexual behavior Part can not adjust adaptation, so not being available substantially.

In one embodiment, game AI can be realized based on deeply study, in a manner of end to end, image frame As input, by on-line training deep neural network model, the movement needed to be implemented is then exported by the model, however, should Mode can obtain good effect under single scene of game.But in most of commercial game, is swum and is illustrated with soul bucket sieve hand, (1) in scene of game, it is mingled with a lot of other scenes, such as cut scene, dialog box etc., it is strong that these are not particularly suited for depth Change learning algorithm to solve.(2) horizontal version shooting game, other than shooting fight, it is also necessary to which control role is walked by route, this portion Divide and is also not suitable for deeply learning algorithm to solve.

The embodiment of the present invention provides a kind of data processing method of virtual scene, the data processing for implementing virtual scene The device of method, the storage medium for being stored with executable instruction for configuration for executing data processing, for convenient for being easier to understand The data processing method of virtual scene provided in an embodiment of the present invention illustrates virtual scene provided in an embodiment of the present invention first The exemplary implement scene of data processing method, virtual scene can export based entirely on terminal device, or be set based on terminal The collaboration of standby and server exports.

In an implement scene, referring to fig. 2, Fig. 2 is the data processing of virtual scene 100 provided in an embodiment of the present invention The optional application model schematic diagram of one of method, can be complete suitable for some computing capabilitys for being completely dependent on terminal device 200 At the application model of the correlation data calculation of virtual scene, such as the game of standalone version/off-line mode, by smart phone, put down The output of the terminal devices 200 such as plate computer and virtual reality/augmented reality equipment completion virtual scene.

When forming the visual perception of virtual scene, terminal device 200 is calculated required for display by graphics calculations hardware Data, and complete display data load, parsing and rendering, images outputting hardware output can to virtual scene formed regard The video frame of perception is felt, for example, two-dimensional video frame is presented in the display screen in smart phone, alternatively, in augmented reality/virtual The video frame of Three-dimensional Display effect is realized in projection on the eyeglass of Reality glasses；In addition, in order to enrich perceived effect, equipment can be with The one or more of Auditory Perception, tactilely-perceptible, motion perception and taste perception are formed by different hardware.

As an example, terminal device 200 runs the game application of standalone version, defeated in the operational process of game application Out include the virtual scene for acting role playing, virtual scene be for game role interaction or mobile environment, such as can be with It is the map sought for the environment or progress precious deposits of ring, progress gunbattle for game role fight battle；Virtual scene In include target object 110, can also include and target object interacts the interactive object 120 of (as fight), target object 110 can be the game role being controlled by user, i.e. target object 110 is controlled by true player, will be responsive to true player's needle The operation of controller (including touch screen, audio-switch, keyboard, mouse and rocking bar etc.) is moved in virtual scene, such as When true player's touch-control indicates the operation button to move right, target object will move right in virtual scene, can be with It keeps static original place, jump and uses various functions (such as technical ability and stage property)；In the virtual scene of game, interactive object 120 can be the enemy to fight with target object 110, such as dogface, b oss, and the quantity of interactive object 120 can have more It is a.

As another example, shopping guide's application is installed in terminal, shop is exported in the operational process of shopping guide's application Three-dimensional virtual scene, includes target object 110 and interactive object 120 in virtual scene, and target object 110 can be user's sheet People/user 3-dimensional image, target object 110 can move freely in shop, and be used by naked eye/virtual reality device Family can perceive the 3-dimensional image of shop and extensive stock, and interactive object 120, which can be, utilizes stereo display technique output Shopping Guide role, Shopping Guide role according to robot model can the consulting to user suitably answered, and pushed away to user Recommend suitable commodity.

It is the data processing of virtual scene provided in an embodiment of the present invention referring to Fig. 3, Fig. 3 in another implement scene The optional application model schematic diagram of one of method, is applied to terminal device and server (illustratively shows terminal in figure 200/300 and server 30), generally, suitable for reliance server 30 computing capability complete virtual scene calculate and at end The application model of end equipment output virtual scene.

By taking the visual perception for forming virtual scene as an example, server 30 carries out the calculating of virtual scene display related data simultaneously It is sent to terminal device, terminal device is completed to calculate load, parsing and the rendering of display data dependent on graphics calculations hardware, according to Rely and export virtual scene in images outputting hardware to form visual perception, such as two can be presented in the display screen of smart phone The video frame of dimension, alternatively, the video frame of Three-dimensional Display effect is realized in projection on augmented reality/virtual reality glasses eyeglass； For the perception of the form of virtual scene, it will be understood that can be exported, such as made by means of the corresponding hardware of terminal device It is exported to form Auditory Perception with microphone, exports to form tactilely-perceptible etc. using vibrator.

As an example, the game application of terminal device operational network version passes through connection game server and other use Family carries out game interaction, and the virtual scene of terminal device output game application can also include including target object 110 The interactive object 120 interacted with target object 110, quantity can be one or more, and target object 110, which can be, to be benefited from The game role of family (also referred to as true player, to be different from robot model) control will be responsive to true player for control The operation of device (including keyboard, mouse and rocking bar etc.) and moved in virtual scene, such as when true player click jump key When, target object 110 executes skip operation in virtual scene；The control for being also based on user uses various function (such as skills Energy and stage property).

Next the image processing system of virtual scene provided in an embodiment of the present invention is illustrated, referring to fig. 4, Fig. 4 For the data processing system configuration diagram of virtual scene provided in an embodiment of the present invention, one exemplary application of support is realized, User terminal 200 (illustrating user terminal 200-1 and user terminal 200-2) connects server 30 by network 20, Network 20 can be wide area network or local area network, or be combination, realize that data are transmitted using Radio Link.

User terminal 200, for (illustrating graphical interfaces 210-1 and graphical interfaces by graphical interfaces 210 It 210-2) shows the image of virtual scene, and the image (as obtained multiframe picture by screenshotss) of virtual scene is sent to clothes Business device 30.

Server 30, for being identified to image of the user terminal 200 to virtual scene, to determine locating for virtual scene Stage；The different phase according to locating for virtual scene, by the way of calling deep neural network model or condition triggering, really Surely the object run of different phase locating for virtual scene is corresponded to；And object run data are returned into user terminal 200, make its control System executes the corresponding object run of moment locating for virtual scene.

The data processing equipment of virtual scene provided in an embodiment of the present invention may be embodied as hardware, software or software and hardware In conjunction with mode, illustrate the various exemplary implementations of the data processing equipment of virtual scene provided in an embodiment of the present invention below.

It elaborates to the hardware configuration of the data processing equipment of the virtual scene of the embodiment of the present invention, Fig. 5 is this hair The composed structure schematic diagram of the data processing equipment for the virtual scene that bright embodiment provides, it will be understood that Fig. 5 illustrate only void Intend the exemplary structure of the data processing equipment of scene rather than entire infrastructure, the part knot shown in Fig. 5 can be implemented as needed Structure or entire infrastructure.The data processing equipment of virtual scene provided in an embodiment of the present invention include: at least one processor 310, Memory 340, at least one network interface 320 and user interface 330.Various components in the data processing equipment of virtual scene It is coupled by bus system 350.It is understood that bus system 350 is for realizing the connection communication between these components.Always Linear system system 350 further includes power bus, control bus and status signal bus in addition in addition to including data/address bus.But in order to clear For the sake of Chu's explanation, various buses are all designated as bus system 350 in Fig. 5.

User interface 330 may include display, keyboard, mouse, trace ball, click wheel, key, button, touch-sensitive plate or Person's touch screen etc..

Memory 340 can be volatile memory or nonvolatile memory, may also comprise volatile and non-volatile Both memories.Wherein, nonvolatile memory can be read-only memory (ROM, Read Only Memory), programmable Read memory (PROM, Programmable Read-Only Memory), Erasable Programmable Read Only Memory EPROM (EPROM, Erasable Programmable Read-Only Memory), flash memory (Flash Memory) etc..Volatile memory can be with It is random access memory (RAM, Random Access Memory), is used as External Cache.By exemplary but not It is restricted explanation, the RAM of many forms is available, such as static random access memory (SRAM, Static Random Access Memory), synchronous static random access memory (SSRAM, Synchronous Static Random Access Memory).The memory 340 of description of the embodiment of the present invention is intended to include the memory of these and any other suitable type.

Processor 310 can be a kind of IC chip, the processing capacity with signal, such as general processor, number Word signal processor (DSP, Digital Signal Processor) either other programmable logic device, discrete gate or Transistor logic, discrete hardware components etc., wherein general processor can be microprocessor or any conventional processing Device etc..

Memory 340 can store executable instruction 3401 to support the operation of the data processing equipment 300 of virtual scene, The example of these executable instructions includes: program, plug-in unit and the foot for operating on the data processing equipment 300 of virtual scene The various forms of software modules such as this, program for example may include operating system and application program, wherein operating system includes each Kind system program, such as ccf layer, core library layer, driving layer etc., are based on hardware for realizing various basic businesses and processing Task.

Below in conjunction with the exemplary application of the data processing equipment of the virtual scene above-mentioned for realizing the embodiment of the present invention And implementation, illustrate the data processing method for realizing the virtual scene of the embodiment of the present invention.

It is an optional process of the data processing method of virtual scene provided in an embodiment of the present invention referring to Fig. 6, Fig. 6 Schematic diagram the step of showing in conjunction with Fig. 6, is illustrated the data processing method of the virtual scene of the embodiment of the present invention.

Step 401: the image of virtual scene being identified, with the stage locating for the determination virtual scene.

In one embodiment, the stage locating for virtual scene can be divided into two stages according to computation complexity, respectively Object interaction stage and non-object interaction stage；Wherein, the object interaction stage refers in virtual scene, exists and target pair As the object interacted, corresponding computation complexity is higher with respect to for the non-object interaction stage, and correspondingly, non-object The interaction stage refers to that in virtual scene, there is only target objects, and there is no the object interacted with target object, meters It calculates lower for the complexity relative object interaction stage；For example, in the virtual scene of game, when the corresponding role's (mesh of user Mark object) with enemy fight when, corresponding virtual scene is in the object interaction stage, and the stage, there are the wars between role Bucket, complicated for operation, corresponding computation complexity is higher, and when battle object is not present, the only corresponding role of user is in movement When, easy to operate, computation complexity is relatively low, and corresponding virtual scene is in the non-object interaction stage.

In one embodiment, the stage locating for virtual scene can be determined as follows:

Object identifying is carried out to the image of virtual scene；When there is only target objects in recognition result characterization virtual scene When, determine that the virtual scene is in the non-object interaction stage；Exist when recognition result characterizes in the virtual scene except described When interactive object other than target object, determine that the virtual scene is in the object interaction stage.

Here, in actual implementation, algorithm of target detection can be used, Object identifying is carried out to the image of virtual scene, to sentence With the presence or absence of the interactive object in addition to the target object in the fixed virtual scene.In one embodiment, target detection is calculated Method can be YOLO (You Only Look Once) object detection system or SSD (Single Shot MultiBox Detector) algorithm of target detection.

It is illustrated for carrying out Object identifying using image of the YOLO object detection system to virtual scene.In reality It further include the process to YOLO target detection model training before the image to virtual scene carries out Object identifying when implementation, In one embodiment, it can train in the following way and obtain YOLO target detection model:

Acquire the image in virtual scene comprising object (target object and/or interactive object)；To right in the image of acquisition As locating region is marked, training sample is formed；Based on obtained training sample, using Y OLO algorithm of target detection into Row model training obtains the YOLO target detection model that Object identifying is carried out for the image to virtual scene.For example, for void Quasi- scene is to acquire in scene of game comprising role corresponding to user and/or comprising role for scene of game as enemy's Region locating for different role in the image of acquisition is marked in frame image, forms training sample；Based on obtained training sample This, carries out model training using YOLO algorithm of target detection, obtains carrying out different role identification for the image to scene of game Y OLO target detection model.

Step 402: the different phase according to locating for virtual scene, using the side for calling machine learning model or condition triggering Formula determines the object run of different phase locating for the corresponding virtual scene.

In actual implementation, for the different phase of virtual scene, it is desirable that computational complexity it is different, can be according to complexity Degree so can be improved treatment effeciency using the processing mode in corresponding stage, in one embodiment, corresponding objects interaction rank Section, the computation complexity as corresponding to the virtual scene is higher, and machine learning model, such as deep neural network model can be used It handles, for the non-object interaction stage, since the corresponding computation complexity of the virtual scene is lower, without neural network Model treatment, the mode that condition triggering can be used are handled, and next determine that the process of object run carries out to different phase Explanation.

It in one embodiment, can object interaction stage in the following way when virtual scene is in the object interaction stage Object run:

The image of virtual scene comprising target object is inputted into deep neural network model, obtains target object in object The object run in interaction stage.It specifically, can be that the multiframe described image being sequentially arranged is inputted into depth nerve net Network model obtains the evaluation of estimate of each candidate operations in multiple candidate operations in object interaction stage；It is highest to choose evaluation of estimate Candidate operations are object run of the target object in the object interaction stage.Here the quantity of candidate operations can be according to practical need It is set.

Here, in one embodiment, before the image of virtual scene is inputted deep neural network model, need to figure As being pre-processed as follows: carrying out size adjusting to image, to be adapted to the input of deep neural network, then carry out ash to image Degreeization processing.

It is illustrated so that deep neural network model is DQN model as an example, by the continuous four frames figure according to timing arrangement As zooming in and out to the size (such as 84*84*4) of adaptation DQN model, gray processing processing is then carried out, is arranged in chronological order The picture queue of column, is then inputted DQN model, obtains the Q value of multiple candidate operations, chooses the highest candidate operations of Q value For target object the object interaction stage object run.

It in one embodiment, can be to depth nerve net before determining object run according to deep neural network model Network model is trained, such as on-line training, specifically, the training of deep neural network model can be realized in the following way:

Obtain the first state information that virtual scene is in the object interaction stage, and the first of corresponding first state information It is one of multiple candidate operations that first object operation is stated in object run；It controls target object and executes first object operation, obtain the Two-state information；Excitation value is calculated based on first state information and the second status information；Based on excitation value and the second state Information, training deep neural network model predict the object run of target object according to the status information in object interaction stage Performance.

In actual implementation, obtain virtual scene be in the first state information in object interaction stage can be with are as follows: obtain and work as The set of the history image frame of the picture frame and virtual scene of preceding virtual scene；Specifically, history image frame is display order Picture frame before current image frame can be a frame or multiframe, for example, first state information can be to arrange in chronological order The image of four frame virtual scenes of column, wherein last frame is the picture frame of current virtual scene.When target object executes first After object run, current virtual scene changes, and corresponding second status information is the four frames void being sequentially arranged The image of quasi- scene, wherein last frame is the picture frame of updated virtual scene.

Here, status information includes at least one of: the attribute value of target object, the attribute value of interactive object, target The object run result of object；For example, attribute value can be the corresponding vital values (blood of role in the virtual scene of game Amount)；For another example, in the virtual scene in the shop that virtual reality technology is realized, attribute value can indicate customer for each of commodity Kind preference；In the virtual scene of game, object run result can be the score of target object, fight triumph/unsuccessfully etc..

In practical applications, the attribute value of target object in every frame image, interactive object can be determined by image recognition The object run result of attribute value, target object；By taking virtual scene is scene of game as an example, it can be known by way of image recognition The vital values of other game role, specifically, vital values are characterized using haemal strand, positioning haemal strand region in the picture, then Binary conversion treatment is carried out to haemal strand region, and counts the ratio in haemal strand region shared by white pixel, can be obtained hundred shared by blood volume Divide ratio；When object run result using score to characterize when, can by number knowledge obtain specific score otherwise, when target grasp Make result using the text of similar " fight is won " can realize by preset template matching algorithm when characterizing, such as calls Text region algorithm identifies the result of displaying, and " fight triumph " still " fight failure " is shown with judgement.

In one embodiment, excitation value can be calculated in the following way:

Based on first state information and the second status information, state change information is obtained；According to state change information, state Mapping relations between variation and excitation value, are calculated excitation value.For example, state change includes the change of target object vital values Change, the mapping relations between state change and excitation value are linear relationship, if vital values are reduced, corresponding excitation value is negative value.

In one embodiment, deep neural network model can be trained according to the state in object interaction stage in the following way Information predicts the performance of the object run of target object:

According to the excitation value, the parameter of percentage regulation neural network model；According to the depth nerve net after adjusting parameter Network model obtains third state information；Excitation value is calculated based on the second status information and third state information；It repeats above-mentioned Operation, until the deep neural network model restrains.

Wherein, using the second status information as input, the deep neural network model after calling adjusting parameter is predicted to obtain Corresponding object run；Control target object executes the object run that prediction obtains, and obtains third state information.

In practical applications, the corresponding condition of convergence of deep neural network model can be configured according to actual needs, For example, it may be frequency of training reaches preset number, such as 100,000 times, or the model parameter that current training obtains is with before The difference of the model parameter once updated is less than threshold value.Since above-mentioned deep neural network model is repeatedly changed by sample Generation training, according to being obtained after excitation feedback data adjustment model parameter, therefore can step up model parameter it is accurate Property.

It in one embodiment, can object interaction rank in the following way when virtual scene is in the non-object interaction stage The object run of section:

The image recognition that at least one dimension is carried out to the image of the virtual scene obtains corresponding recognition result；Root According to the trigger condition that the recognition result is met, the object run for corresponding to the non-object interaction stage is obtained.

Specifically, in one embodiment, the image recognition carried out to the image of virtual scene can be at least one of:

Mobile instruction identification is carried out to the image of virtual scene, is used to indicate target to determine to whether there is in virtual scene Object carries out mobile mobile sign；

Operation instruction identification is carried out to the image of virtual scene, is used to indicate control to determine to whether there is in virtual scene Target object and/or the operation instruction mark for carrying out scene switching.

Here, by taking virtual scene is scene of game as an example, carrying out mobile instruction identification to the image of virtual scene can be Image recognition is carried out to the frame image of scene of game, to determine to be moved in scene with the presence or absence of mobile sign, such as instruction The sign that this arrow or instruction jumps；Carrying out operation instruction identification to the image of virtual scene can be for game The frame image of scene carries out image recognition, to determine such as to control game role with the presence or absence of operation button in scene and discharge technical ability Key, control carry out scene switching key (such as skipping key).

In one embodiment, mobile instruction identification carried out to the image of virtual scene and operation instruction identification can be used it is default Template matching algorithm realize, such as by under type realization:

It positions mobile sign or operation instruction identifies the region in described image；Feature is carried out to the region of positioning It extracts, is then matched with the feature of the mark of storage；There is movement in the image for determining virtual scene when successful match Sign or operation instruction mark determine that there is no corresponding marks in the image of virtual scene when it fails to match.

In one embodiment, in the non-object interaction stage, the corresponding non-object interaction stage can be obtained in the following way Object run:

It is mobile instruction identification in response to image recognition, when recognition result is the presence of mobile sign in virtual scene When, determine that the object run is that corresponding route movement is carried out based on the mobile sign；

When recognition result is that the mobile sign is not present in virtual scene, determine that the object run is foundation It is mobile that preset direction carries out corresponding route.

It include mobile instruction identification and operation instruction identification in response to image recognition, when recognition result is the virtual scene In there is no the mobile sign and exist be used to indicate control scene switching the operation instruction mark when, determine The object run is to carry out virtual scene switching；When recognition result is that the mobile indicateing arm is not present in the virtual scene When knowing and identifying there is no the operation instruction for being used to indicate control scene switching, process is ended processing.

Step 403: control executes the corresponding object run of moment locating for the virtual scene.

For example, being identified by the image to virtual scene, determination is currently at the object interaction stage, deep by calling Degree neural network model determines that the object run of target object to discharge a certain certain skills, then controls target object and executes release The operation of the certain skills.

Next by taking virtual scene is scene of game as an example, to the data processing method of the virtual scene of the embodiment of the present invention It is illustrated.Fig. 7 is an optional flow diagram of the data processing method of virtual scene provided in an embodiment of the present invention, Referring to Fig. 7, the data processing method of the virtual scene of the embodiment of the present invention includes:

Step 501: acquiring the frame image of current running game.

Here, game application is installed in terminal, such as " soul bucket sieve: coming back ", after game, which starts, to be run, carries out game picture The acquisition in face can obtain continuous multiframe picture (the i.e. frame figure of current running game in practical applications by way of screenshotss Picture), target object corresponds to the game role of user's control in scene of game.

Step 502: Object identifying being carried out to the frame image of acquisition, is judged whether according to recognition result in object interaction rank Section executes step 503 if being in the object interaction stage；If executing step 506 in the non-object interaction stage.

Here, in actual implementation, Object identifying is carried out using image of the algorithm of target detection to scene of game, to determine With the presence or absence of the interactive object in addition to target object in scene of game, such as determined by YOLO object detection system with screen picture In whether there is enemy.

Step 503: the frame image of acquisition is pre-processed.

In practical applications, the pretreatment carried out to the frame image of acquisition includes following operation:

It adjusts the size of full frame image to the size for being adapted to DQN model, then the frame image after progress size adjusting is carried out Gray processing processing, is sequentially arranged to form continuous frame image queue.

Step 504: pretreated frame image input will be carried out and train obtained DQN model, obtain multiple candidate operations Q value.

Here, input DQN model is by pretreated continuous 4 frame image, and a schematic diagram of DQN model is as schemed Shown in 8, DQN model includes that multilayer convolutional layer and the full articulamentum of multilayer, the specific number of plies can be configured according to actual needs. Image characteristics extraction layer includes convolutional layer (conv in Fig. 8), and every layer is made of multiple two-dimensional surfaces, and each plane is by multiple Independent neuron composition.Each of convolutional neural networks convolutional layer all followed by one is used to that local average is asked to mention with secondary The pond layer (the p ooling in Fig. 8) taken, this distinctive structure of feature extraction twice make network in identification to input figure As there is higher distortion tolerance.After obtaining image content features, the full articulamentum of first layer (FC in Fig. 8) can be input to In, input of the output of the full articulamentum of first layer as the full articulamentum of the second layer.After the mapping of the full articulamentum of multilayer, finally Exported (output).Here, the training of DQN model is not repeated herein referring to foregoing description.

Step 505: target object executes the maximum candidate operations of Q value in control scene of game.

Step 506: judge to carry out mobile mobile sign with the presence or absence of instruction target object in the frame image of acquisition, If it does, executing step 507；If it does not, executing step 508.

Here, in actual implementation, can judge to refer to the presence or absence of movement in image by preset template matching algorithm Indicating is known, for example, positioning the region of image locating for mobile sign, the image in region obtained to positioning carries out feature and mentions It takes, and matched obtained feature is extracted with the feature of storage, when successful match determines there is mobile sign, matching When failure, it is determined that there is no mobile signs.Mobile sign can be moved to certain certain bits for instruction game role The arrow set, or the arrow that instruction game role jumps.

Step 507: it is mobile that corresponding route being carried out based on mobile sign control target object.

Step 508: judge to identify in the frame image of acquisition with the presence or absence of the operation instruction for being used to indicate progress scene switching, If it does, executing step 509；If it does not, executing step 510.

Here, identify that determining, which whether there is in image, is used to indicate by the operation instruction of the frame image progress to acquisition Carry out the operation instruction mark of scene switching.Here operation instruction mark can indicate for such as the text of " skipping ".

Step 509: being identified based on operation instruction, control carries out scene of game switching.

Step 510: terminating this process flow.

Next by taking virtual scene is horizontal version soul bucket sieve scene of game as an example, to the number of the virtual scene of the embodiment of the present invention It is illustrated according to processing method.Fig. 9 is that one of the data processing method of virtual scene provided in an embodiment of the present invention is optional Flow diagram, referring to Fig. 9, the data processing method of the virtual scene of the embodiment of the present invention includes:

Step 601: obtaining the frame image of the current running game of terminal.

Here, the picture of the game of a frame or multiframe operation can be obtained by way of screen shot.

It should be noted that being referred to herein as leading role by the role of user's control in game, interacted with leading role Object is referred to as enemy, such as boss.

Step 602: image recognition being carried out to the frame image of the game of acquisition, obtains recognition result.

In one embodiment, image recognition can be carried out to game picture from multiple dimensions, for example, to the game of acquisition Frame image carries out identified below: the vital values of leading role, the vital values of Boss, each operable key, the enemy that same screen occurs, Score value, game result etc..

In practical applications, algorithm used by the identification of the different dimensions carried out to image is as shown in Figure 10, Tu10Wei The schematic diagram of the image recognition of various dimensions provided in an embodiment of the present invention is based on Figure 10, the figure next carried out to each dimension As identification is illustrated respectively.

1), vital values identify

In embodiments of the present invention, vital values identification includes that the identification of the vital values of game leading role and the vital values of boss are known Not；Figure 11 is an optional schematic diagram provided in an embodiment of the present invention for carrying out vital values identification, and referring to Figure 11, vital values are adopted It is characterized with haemal strand, binaryzation is carried out to haemal strand region, the pixel for then counting white accounts for the ratio of entire haemal strand, this ratio That is vital values percentage.

2), recognition by pressing keys

It in embodiments of the present invention, may include the identification of technical ability key, skip Button identification etc. to the identification of key, Figure 12 is this The schematic diagram for the game keys that inventive embodiments provide, referring to Figure 12, wherein it is technical ability key shown in number 11,12, number 13 be the skip Button for carrying out scene of game switching；The template matching algorithm that preset fixed position can be used obtains, for example, Text region can be carried out for the region to corresponding skip Button to skip Button identification, and judge whether that identification obtains the text of " skipping " Word.

3), enemy identifies

The dogface occurred in game, the assailable enemy such as small monster are detected with algorithm of target detection YOLO or SSD.? It when actual implementation, needs first to carry out the training of YOLO neural network model, is then based on the YOLO neural network model that training obtains Carry out enemy's identification.

Here the training of YOLO neural network model is illustrated, the training of YOLO neural network model specifically includes that

Acquisition includes the game picture (frame image) of enemy, and then artificial mark includes the region of enemy, generates training sample This, carries out model training using YOLO Target Recognition Algorithms, obtains the YO LO neural network model of enemy for identification.

4), score identifies

Due to the fixation band of position for obtaining branch and being shown in image fought, in the position area that score is shown Digital identification is carried out in domain, obtains current score, if score 50 is divided, obtains being scored at 50 known to 50 by number identification.

5), game result identifies

Whether game over picture shows the detection of triumph or failure, can make of the template matching algorithm of fixed position It arrives.When due to game over, game result can be shown in the specific region of image, such as image middle position, by the spy Determine region and carry out Text region, the game result of " fight triumph " or " fight failure " can be obtained.

Step 603: according to image recognition result, judging whether scene of game is scene of fighting, if it is scene of fighting, is held Row step 604；If not scene of fighting, step 606 is executed.

Here, in actual implementation, if the judgment basis for being scene of fighting is in frame image with the presence or absence of same with leading role The enemy of screen, if it is present determining that current scene of game is scene of fighting, if it does not exist, then determining current game The non-scene of fighting of scene.

Step 604: the DQN model that the frame image input training of continuous four frames game is obtained obtains 7 candidate operations Q value.

Here, the quantity of candidate operations and corresponding operation content can be according to actual needs into settings, such as candidate behaviour Work may include: no operation, be moved to the left, and move right, and squat down, jump, 2 technical ability of leading role.

Figure 13 is a schematic diagram of DQN deep neural network model provided in an embodiment of the present invention, referring to Figure 13, to obtaining Then the frame image scaling of the game taken carries out gray processing processing, from continuous arrangement in chronological order to the size of 176*108*4 Frame image queue in choose continuous four frames image, be input to the obtained DQN deep neural network model of training, export 7 The Q value of candidate operations.

The training of DQN deep neural network model is illustrated.Figure 14 is on-line training provided in an embodiment of the present invention The method flow schematic diagram of DQN model, referring to Figure 14, the method for on-line training DQN model includes:

Step 701: obtaining game picture in real time.

Step 702: image recognition is carried out to the game picture of acquisition.

Step 703: judging whether current scene of game is in combat phase according to recognition result, if so, executing step Rapid 704；If it is not, executing processing corresponding to non-combat phase.

Step 704: construction input state sample.

Here, input state sample is to pass through pretreated game frame image.

Step 705: calculating excitation value.

Here, excitation function is designed for the purpose of it can win fight.Specifically it can be according in image recognition result Some state changes are arranged excitation value, and there are mapping relations with excitation value for state change, and referring to table 1, table 1 shows sports ground Mapping relations between the state change and excitation value of scape.

State change	Meaning	Excitation value
			The blood volume of leading role is lost	Leading role's oligemia, then punish	-0.01/unit
The loss of Boss blood volume	Boss oligemia, then reward	0.01/unit
			Score increases	Score increases, then rewards	0.1/unit
It is injured	It comes to harm, then punishes	-0.01/unit

Table 1

Step 706: on-line training DQN model.

During carrying out DQN model training, map to obtain corresponding excitation value by obtaining state change, and pass through To excitation value the parameter of DQN deep neural network model is adjusted, it is further pre- by parameter DQN model adjusted Leading role's action which be to be is surveyed, movement is executed by controlling leading role, so that scene of game generating state changes.

Step 707: control output combat action, and execute step 701.

During model training, combat action of every output completes the change of a next state, is realized based on excitation value The adjustment of model parameter, by continuous repetitive exercise, until DQN model is restrained, training terminates.

Step 605: control game leading role executes the maximum candidate operations of Q value.

Step 606: judging whether to need to walk, if it is not needed, executing step 607；If desired, executing step 609.

Here, in actual implementation, if the judgment basis for needing to walk is that there are arrows in the frame image of game Or judged without enemy by carrying out image recognition to image in same screen.

Step 607: judging whether to need to carry out scene of game switching, if it is desired, execute step 608；If it is not needed, End processing process.

Here, if need scene switching judgment basis be can operation button disappear or occur skipping key, in reality In, player often occurs some cut scene scenes or session operational scenarios when Boss occurs, is in this in game Under scene of game, game can not be played, therefore when being obtained by identification in the presence of key is skipped, it is known that needed to control and be swum Play scene switching.

Step 608: control carries out scene of game switching.

Here, in practical applications, key realization scene switching can be skipped by controlling to click.

Step 609: judging with the presence or absence of arrow in scene of game, if it does, executing step 610；If do not deposited Executing step 611.

Further judged in image by image recognition with the presence or absence of the arrow for being used to indicate leading role's movement.

Step 610: control leading role is based on arrow and moves.

In horizontal version shooting game, route walking is divided into horizontal path walking and vertical path jump.Horizontal path is walked It (walks to the right) by default route walking.

Correspondingly, arrow is divided into the arrow of the arrow that instruction moves horizontally and instruction jump, jump if it is instruction Arrow, need to identify corresponding arrow locations by image recognition technology, then move at the arrow locations and jumped Jump, could carry out the jump of vertical path in correct position.

611: control leading role walks to the right.

Using the above embodiment of the present invention, have it is following the utility model has the advantages that

1) different phase according to locating for scene of game (combat phase, non-combat phase) uses different processing modes, For combat phase, using the operation of DQN model output fight, since DQN model only solves the problems, such as under scene of fighting, so With clearly defined objective, scene is single, can improve training speed；Non- combat phase is passed through without complicated neural network model Simple rule or condition triggering can be realized, reduce the requirement to the computing capability of equipment, improve treatment effeciency.

2) it realizes horizontal version shooting game AI, tester can be substituted enter in game and be played automatically, be played The purpose of auxiliary game test.

3) batch deployment AI carries out the test that game is played automatically, and a large amount of human resources that can save traditional test needs are opened Pin.

4) battle model of DQN is by means of deep neural network, so having good generalization.

The data processing equipment for continuing to explain virtual scene provided in an embodiment of the present invention is provided as the embodiment of the present invention Virtual scene the data processing equipment example that uses software and hardware combining to implement, virtual scene provided by the embodiment of the present invention Data processing equipment can be embodied directly in the various forms of software modules executed by processor 310, software module can be with In storage medium, storage medium is located at memory 340, and what software module included in the reading memory 340 of processor 310 can It executes instruction, completes this in conjunction with necessary hardware (e.g., including processor 310 and the other assemblies for being connected to bus 350) The data processing method for the virtual scene that inventive embodiments provide.Figure 15 is the data of virtual scene provided in an embodiment of the present invention The composed structure schematic diagram of processing unit, referring to Figure 15, device includes:

Recognition unit 341 is identified for the image to virtual scene, with rank locating for the determination virtual scene Section；

Determination unit 342, for the different phase according to locating for the virtual scene, using calling deep neural network mould Type or the mode of condition triggering, determine the object run of different phase locating for the corresponding virtual scene；

Execution unit 343 executes the corresponding object run of moment locating for the virtual scene for controlling.

In one embodiment, the determination unit is also used to be in the object interaction stage in response to the virtual scene, will The image of virtual scene comprising target object inputs deep neural network model, obtains the target object and hands in the object The object run in mutual stage.

In one embodiment, the determination unit, the multiframe described image input institute for being also used to be sequentially arranged Deep neural network model is stated, the evaluation of estimate of each candidate operations in multiple candidate operations in the object interaction stage is obtained；

Choosing the highest candidate operations of evaluation of estimate is object run of the target object in the object interaction stage.

In one embodiment, described device further include: acquiring unit and training unit；Wherein,

The acquiring unit, the first state information for being in the object interaction stage for obtaining the virtual scene, and The first object operation of the corresponding first state information, the first object operation is one of the multiple candidate operations；

The execution unit is also used to control the target object and executes the first object operation, obtains the second state Information；

The training unit, for excitation to be calculated based on the first state information and second status information Value；

And it is based on the excitation value and second status information, the training deep neural network model is according to right As the status information in interaction stage, the performance of the object run of the target object is predicted.

In one embodiment, the training unit is also used to believe based on the first state information and second state Breath, obtains state change information；

According to the mapping relations between the state change information, state change and excitation value, excitation value is calculated；

Wherein, status information includes at least one of: the attribute value of target object, the attribute value of interactive object, target The object run result of object.

In one embodiment, the training unit is also used to adjust the deep neural network mould according to the excitation value The parameter of type；

According to the deep neural network model after adjusting parameter, third state information is obtained；

Excitation value is calculated based on second status information and the third state information；

Aforesaid operations are repeated, until the deep neural network model restrains.

In one embodiment, the recognition unit is also used to be in the non-object interaction stage in response to the virtual scene, The image recognition that at least one dimension is carried out to the image of the virtual scene obtains corresponding recognition result；

The determination unit is also used to the trigger condition met according to the recognition result, and it is described non-right to obtain corresponding to As the object run in interaction stage.

In one embodiment, the recognition unit is also used to execute at least one of to the image of the virtual scene:

Mobile instruction identification is carried out to the image of the virtual scene, is used for determining to whether there is in the virtual scene Indicate that target object carries out mobile mobile sign；

Operation instruction identification is carried out to the image of the virtual scene, is used for determining to whether there is in the virtual scene It indicates to control the target object and/or carries out the operation instruction mark of scene switching.

In one embodiment, the determination unit is also used to be identified as mobile instruction identification in response to described image, works as knowledge Other result is that there are when the mobile sign, determine that the object run is to refer to based on the movement in the virtual scene Indicating, which is known, carries out corresponding route movement；

When recognition result is that the mobile sign is not present in the virtual scene, determine that the object run is It is mobile that corresponding route is carried out according to preset direction.

In one embodiment, the determination unit, be also used to identify in response to described image include mobile instruction identification and Operation instruction identification, when recognition result is that the mobile sign is not present in the virtual scene and exists to be used to indicate When controlling the operation instruction mark of scene switching, the object run is determined to carry out virtual scene switching.

In one embodiment, the recognition unit is also used to carry out Object identifying to the image of the virtual scene；

When recognition result characterizes in the virtual scene, there is only when target object, determine the virtual scene in non-right As the interaction stage；

When recognition result characterizes the interactive object existed in addition to the target object in the virtual scene, institute is determined It states virtual scene and is in the object interaction stage.

In one embodiment, the recognition unit is also used to the image using algorithm of target detection to the virtual scene Object identifying is carried out, to determine in the virtual scene with the presence or absence of the interactive object in addition to the target object.

The embodiment of the invention also provides a kind of storage medium, it is stored with executable program, at the executable code When managing device execution, the data processing method of the above-mentioned virtual scene of the embodiment of the present invention is realized.

It need to be noted that: above is referred to the descriptions of the data processing equipment of virtual scene, with above-mentioned virtual scene Data processing method description be it is similar, with method beneficial effect describe, do not repeat them here.For virtual field of the present invention Undisclosed technical detail in the data processing equipment embodiment of scape, please refers to the description of embodiment of the present invention method.

The above, only the embodiment of the present invention, are not intended to limit the scope of the present invention.It is all in this hair Made any modifications, equivalent replacements, and improvements etc. within bright spirit and scope, be all contained in protection scope of the present invention it It is interior.

Claims

1. a kind of data processing method of virtual scene, which is characterized in that the described method includes:

It is determined by the way of calling machine learning model or condition triggering according to different phase locating for the virtual scene The object run of different phase locating for the corresponding virtual scene；

2. the method as described in claim 1, which is characterized in that the different phase according to locating for the virtual scene is adopted With the mode for calling machine learning model or condition triggering, the target behaviour of different phase locating for the corresponding virtual scene is determined Make, comprising:

It is in the object interaction stage in response to the virtual scene, the image of the virtual scene comprising target object is inputted into depth Neural network model obtains the target object in the object run in the object interaction stage.

3. method according to claim 2, which is characterized in that the image by the virtual scene comprising target object inputs Deep neural network model obtains the target object in the object run in the object interaction stage, comprising:

The multiframe described image being sequentially arranged is inputted into the deep neural network model, obtains the object interaction rank The evaluation of estimate of each candidate operations in multiple candidate operations of section；

4. method according to claim 2, which is characterized in that the method also includes:

The first state information that the virtual scene is in the object interaction stage is obtained, and corresponds to the first state information First object operation, the first object operation is one of the multiple candidate operations；

It controls the target object and executes the first object operation, obtain the second status information；

Excitation value is calculated based on the first state information and second status information；

Based on the excitation value and second status information, the training deep neural network model is according to the object interaction stage Status information, predict the performance of the object run of the target object.

5. method as claimed in claim 4, which is characterized in that be based on the first state information and second status information Excitation value is calculated, comprising:

Based on the first state information and second status information, state change information is obtained；

Wherein, status information includes at least one of: the attribute value of target object, the attribute value of interactive object, target object Object run result.

6. method as claimed in claim 4, which is characterized in that be based on the excitation value and second status information, training The deep neural network model predicts the property of the object run of the target object according to the status information in object interaction stage It can, comprising:

According to the excitation value, the parameter of the deep neural network model is adjusted；

7. the method as described in claim 1, which is characterized in that the different phase according to locating for the virtual scene is adopted With the mode for calling machine learning model or condition triggering, the target behaviour of different phase locating for the corresponding virtual scene is determined Make, comprising:

It is in the non-object interaction stage in response to the virtual scene, at least one dimension is carried out to the image of the virtual scene Image recognition, obtain corresponding recognition result；

According to the trigger condition that the recognition result is met, the object run for corresponding to the non-object interaction stage is obtained.

8. the method for claim 7, which is characterized in that the image to the virtual scene carries out at least one dimension The image recognition of degree obtains corresponding recognition result, comprising:

At least one of is executed to the image of the virtual scene:

Mobile instruction identification is carried out to the image of the virtual scene, is used to indicate with determining to whether there is in the virtual scene Target object carries out mobile mobile sign；

Operation instruction identification is carried out to the image of the virtual scene, is used to indicate with determining to whether there is in the virtual scene It controls the target object and/or carries out the operation instruction mark of scene switching.

9. method according to claim 8, which is characterized in that the trigger condition met according to the recognition result, Obtain corresponding to the object run in the non-object interaction stage, comprising:

It is identified as mobile instruction identification in response to described image, when recognition result is that there are the movements to refer in the virtual scene When indicating is known, determine that the object run is that corresponding route movement is carried out based on the mobile sign；

When recognition result is that the mobile sign is not present in the virtual scene, determine that the object run is foundation It is mobile that preset direction carries out corresponding route.

10. method according to claim 8, which is characterized in that the trigger condition met according to the recognition result, Obtain corresponding to the object run in the non-object interaction stage, comprising:

It include mobile instruction identification and operation instruction identification in response to described image identification, when recognition result is the virtual scene In there is no the mobile sign and exist be used to indicate control scene switching the operation instruction mark when, determine The object run is to carry out virtual scene switching.

11. method as described in any one of claim 1 to 10, which is characterized in that the image to virtual scene is known Not, with the stage locating for the determination virtual scene, comprising:

Object identifying is carried out to the image of the virtual scene；

When recognition result characterizes in the virtual scene, there is only when target object, determine that the virtual scene be in non-object friendship The mutual stage；

When recognition result characterizes the interactive object existed in addition to the target object in the virtual scene, the void is determined Quasi- scene is in the object interaction stage.

12. method as claimed in claim 11, which is characterized in that the image to the virtual scene carries out object knowledge Not, comprising:

Object identifying is carried out to the image of the virtual scene using algorithm of target detection, with determine in the virtual scene whether In the presence of the interactive object in addition to the target object.

13. a kind of data processing equipment of virtual scene, which is characterized in that described device includes:

Determination unit is touched for the different phase according to locating for the virtual scene using calling machine learning model or condition The mode of hair determines the object run of different phase locating for the corresponding virtual scene；

14. a kind of data processing equipment of virtual scene, which is characterized in that described device includes:

Memory, for storing executable instruction；

Processor when for executing the executable instruction stored in the memory, is realized described in any one of claim 1 to 12 Virtual scene data processing method.

15. a kind of storage medium, which is characterized in that be stored with executable instruction, the executable instruction is for causing memory Perform claim requires the data processing method of 1 to 12 described in any item virtual scenes.