CN108664122A - A kind of attitude prediction method and apparatus - Google Patents
A kind of attitude prediction method and apparatus Download PDFInfo
- Publication number
- CN108664122A CN108664122A CN201810300681.1A CN201810300681A CN108664122A CN 108664122 A CN108664122 A CN 108664122A CN 201810300681 A CN201810300681 A CN 201810300681A CN 108664122 A CN108664122 A CN 108664122A
- Authority
- CN
- China
- Prior art keywords
- network model
- neural network
- attitude
- training
- prediction
- Prior art date
- Legal status (The legal status 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 status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
- G06F3/01—Input arrangements or combined input and output arrangements for interaction between user and computer
- G06F3/011—Arrangements for interaction with the human body, e.g. for user immersion in virtual reality
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- User Interface Of Digital Computer (AREA)
Abstract
The invention discloses a kind of attitude prediction method and apparatus, method includes：According to preset sample frequency, the attitude data that attitude transducer generates when to user movement samples, and obtains raw data set；The number of plies of neural network and each node layer number are determined according to raw data set, obtain the neural network model of training completion；During an attitude prediction, the attitude data in the predetermined amount of time of acquisition is input in neural network model, obtains the corresponding attitude data of object time after the current time of neural network model output；Wherein, predetermined amount of time is the period determined by the predetermined instant before current time and current time.As it can be seen that the neural network model that technical scheme of the present invention is trained using machine learning, which carries out attitude prediction, not only meets efficiency requirements, computation complexity is reduced.Moreover, the small motor pattern in history posture can be captured, the precision of prediction is high, improves the experience such as the feeling of immersion of user.
Description
Technical field
The present invention relates to wear display equipment technical field, and in particular to a kind of attitude prediction method and apparatus.
Background technology
VR (Virtual Reality, virtual reality) and AR (Augmented Reality, augmented reality) technology pass through
Controller allows user to obtain the immersion experience that can be interacted.Controller can be located at HMD, and (Head Mounted Display, wear
Show equipment) in, it may be alternatively located in handle.Controller, which crosses high latency, can destroy feeling of immersion, be transported particularly with this heads of HMD-eye
For the equipment of dynamic feedback, excessively high delay can cause serious sense of discomfort and motion sickness, influence the physiological health of user.One
Perfect VR/AR systems usually require to combine a variety of optimization means to reduce the delay of controller.In the controller of VR/AR usually
The attitude transducers such as IMU (Inertial Measurement Unit, Inertial Measurement Unit is referred to as) are integrated with, and utilize posture
Sensor obtains the attitude data of user, and the system delay that user is perceived is reduced by attitude data prediction.
Or but it is existing based on attitude transducer carry out attitude prediction scheme error it is larger, precision does not reach requirement,
Computation complexity is high, implements extremely difficult, cannot meet actual demand, user experience is bad.
Invention content
Larger, the computation complexity that carries out pre- time determination error to solve existing virtual reality or augmented reality to posture
Height, the bad technical problem of user experience, an embodiment of the present invention provides a kind of attitude prediction method and apparatus.
According to an aspect of the invention, there is provided a kind of attitude prediction method, including：
According to preset sample frequency, the attitude data that attitude transducer generates when to user movement samples, and obtains original
Beginning data set；
The number of plies of initial neural network model and the node of each layer are determined according to raw data set and specified prediction duration
Number, and trained using the initial neural network model of the initial data set pair, to obtain the neural network model of training completion；
During an attitude prediction, the attitude data in the predetermined amount of time of acquisition is input to the god of training completion
The corresponding posture number of object time after the current time of neural network model output through in network model, obtaining training completion
According to；
Wherein, predetermined amount of time is the period determined by the predetermined instant before current time and current time.
Optionally, the number of plies of initial neural network model and each layer are determined according to raw data set and specified prediction duration
Number of nodes, and include using the initial data set pair neural metwork training：
Raw data set is arranged, the posture of the attitude data and object time in the indicating predetermined period is generated
The mapping relations collection of the mapping relations of data；
The data that mapping relations are concentrated are input in initial neural network model, and using supervised learning training nerve
Network model.
Optionally, method further includes：
During using supervised learning training neural network model, neural network model is adjusted according to training result
Specified parameter；Specified parameter includes：The number of plies of neural network model, the number of nodes of each layer, activation primitive and loss function.
Optionally, include according to the specified parameter of training result adjustment neural network model：
The loss function of neural network model is set to according to training result practical attitude data value and the prediction of acquisition
The mean square deviation of attitude data value.
Optionally, method further includes：
After the completion of neural network model is trained, the specified parameter of the neural network model of training completion is extracted and exported
Value, to run neural network model on designated terminal platform.
Optionally, according to preset sample frequency, the attitude data that attitude transducer generates when to user movement samples,
Obtaining raw data set includes：
According to preset sample frequency, the attitude data that attitude transducer generates under normal operating condition when to user movement
It is sampled, the time series of attitude data is obtained, as raw data set；
The time series of attitude data indicates one or more in following user's head or hand exercise：
Smooth pivotal accelerates rotation, underdrive, predetermined time to turn to stopping, predetermined time rotation after predeterminated position
To precalculated position again to back rotation.
According to another aspect of the present invention, a kind of attitude prediction device is provided, applied to display equipment is worn, is wrapped
It includes：
Sampling unit, for attitude data that according to preset sample frequency, attitude transducer generates when to user movement into
Row sampling, obtains raw data set；
Training unit, for according to raw data set and specified prediction duration determine the number of plies of initial neural network model with
And the number of nodes of each layer, and trained using the initial neural network model of the initial data set pair, to obtain the god of training completion
Through network model；
Predicting unit, for during an attitude prediction, the attitude data in the predetermined amount of time of acquisition to be inputted
To training complete neural network model in, obtain training completion neural network model output current time after target when
Carve corresponding attitude data；
Wherein, predetermined amount of time is the period determined by the predetermined instant before current time and current time.
Optionally, training unit generates the appearance in the indicating predetermined period specifically for being arranged to raw data set
The mapping relations collection of the mapping relations of the attitude data of state data and object time；
The data that mapping relations are concentrated are input in initial neural network model, and using supervised learning training nerve
Network model.
Optionally, further include：Training optimization unit, for using supervised learning training neural network model process
In, the specified parameter of neural network model is adjusted according to training result；Specified parameter includes：It is the number of plies of neural network model, each
Number of nodes, activation primitive and the loss function of layer.
Optionally, training optimization unit is specifically used for, and the loss function of neural network model is arranged according to training result
For the mean square deviation of the practical attitude data value and prediction attitude data value of acquisition；
Device further includes：Platform expanding element, the specified ginseng of the neural network model for extracting and exporting training completion
Several value, to run the neural network model that training is completed on designated terminal platform.
Advantageous effect：The attitude prediction method and apparatus of the embodiment of the present invention, attitude transducer when by user movement
The attitude data of generation is sampled to obtain raw data set, and neural network model is trained using raw data set, pre- in posture
During survey, using the neural network model, the attitude data in input past predetermined amount of time can predict following a certain
Moment (i.e. object time) is corresponding, accurate attitude data, and the accuracy of the attitude data of prediction is high, ensure that wear-type is aobvious
Show that equipment can calculate the scenic picture that will be presented in advance according to the attitude data predicted and user is watched virtual
Image and user's head or the movement which matches of hand reduce delay to conveniently wear display equipment, improve user experience.More
Importantly, the technical solution of the present embodiment predicts user movement posture using the neural network model of machine learning, nerve
Network algorithm is suitably based on spy of the motor pattern of user during headset equipment use in the time series of attitude data
It levies and high-precision prediction, and the neural network model that machine learning trains is made to following posture, computationally show
For a series of vector-matrix multiplication and addition, specific complexity depends primarily on the size and the number of plies of matrix and vector,
And operand maximum Function Fitting carries out when can be placed on off-line training, it ensure that computation complexity when attitude prediction compared with
It is low, to take into account two indexs of computation complexity and precision of prediction of athletic posture prediction, solves existing attitude prediction and calculate
Or method error is larger, required precision or computation complexity height is not achieved and realizes difficult technical problem, improve wear it is aobvious
Show the competitiveness of equipment.
Description of the drawings
Fig. 1 is the prediction result comparison diagram using prior art polynomial fitting method；
The flow chart of the attitude prediction method of Fig. 2 one embodiment of the invention；
Fig. 3 is the prediction result comparison diagram using the attitude prediction method of one embodiment of the invention；
Fig. 4 is the block diagram of the attitude prediction device of one embodiment of the invention；
Fig. 5 is the block diagram of the electronic equipment of one embodiment of the invention.
Specific implementation mode
Industry generally believes that the delay of HMD (Head Mounted Display wear display equipment) should be less than 20ms,
Even lower than 15ms or 10ms just can guarantee smooth experience.But these indexs are one to existing VR/AR software and hardware systems
Very high challenge, in VR/AR software views, a kind of important optimization means are the attitude transducers to being integrated on controller
(IMU/Optic) posture of following a period of time is predicted, is responded in advance with this and prepares to examine to reduce user
The system delay felt.
Current attitude prediction method mainly has：
(1) newtonian motion method.The algorithm will current (i.e. time shaft coordinate for 0 at) attitude data (position, direction,
Speed, acceleration etc.) it is updated in newtonian motion formula, to extrapolate the posture at future time point 0+ Δs T.The algorithm is false
The attitude data determined in [0, Δ T] time interval is identical as the attitude data at 0 point.In practice, if the speed of attitude transducer
When degree and angular speed change suddenly in the section [0, Δ T] or even generation direction is turned back, at the 0+ Δs T which predicts
Posture will have serious deviation.And an important indicator of attitude prediction algorithm is error size and error distribution.If
Error is excessive to cause perceiving for user, it will the sense of discomfort of more serious aggravation user, such case ratio do not do attitude prediction
Also want bad.
From the foregoing, it will be observed that newtonian motion method mainly the disadvantage is that computational accuracy is excessively poor, be only utilized and work as in prediction
Attitude data at preceding time point 0 has ignored influence of the historical data to Future Data so that the precision of prediction is very poor.And it is real
On border 0 point for the previous period, i.e., the motion state after 0 point of motion state pair in [- Δ T, 0] has a great impact.
(2) polynomial fitting method.Connect aforementioned, the data in [- Δ T, 0] section contain deeper movable information,
For example, the derivative etc. of acceleration, these additional information can be used to improve the accuracy of prediction.Therefore some schemes, which propose, adopts
With polynomial fitting method, the posture in the section [- Δ T, 0] is changed into the multinomial that curve matching goes out high-order, followed by obtaining
Multinomial calculate the attitude value at 0+ Δs T.Polynomial fitting method also has some mutation, such as, it is believed that the history number before 0 point
In, bigger is influenced on the motion state after 0 point closer to 0 point of data, therefore it is past from 0 point that sliding window algorithm may be used
Front slide simultaneously fits multiple multinomials, is finally weighted again to a plurality of curve by weight average.
Polynomial fitting method and its prediction effect of mutation are promoted than newtonian motion method, and polynomial fitting method utilizes
The variation tendency of posture curve in history, therefore when attitudes vibration is slower, compared to newtonian motion method have compared with
High accuracy, but the limitation improved.Polynomial fitting method disadvantage mainly when acute variation occurs for posture, for example passes
Sensor quickly turns to certain point and at once when being rotated to negative direction, or in the changed place of directional velocity, multinomial
The prediction error of formula fitting process is very big.
Fig. 1 is the prediction result comparison diagram using prior art polynomial fitting method；Referring to Fig. 1, horizontal axis indicates the time, indulges
Axis indicates that, according to the calculated Eulerian angles roll values of IMU initial data, unit (degree), 101 representative polynomial fitting process postures are pre-
Measured value, 102 be the posture actual value that attitude transducer generates；As shown in Figure 1, polynomial fitting method and the attitude data of prediction with
Actual attitude data error is very big.
Present inventor thinks that reason may be, and polynomial fitting method is only simply to historical data curve
It is fitted, in this way in the case where trend direction is constant, trend that the curve that fits rises or falls in future time point
The movement tendency for generally conforming to history also complies with the variation tendency of the following a bit of time.But in direction when acute variation, fitting
Though the curve gone out is identical as historical trend direction in the trend that future time point rises or falls, with actual trend direction
On the contrary, to cause prediction error larger.
Meanwhile polynomial fitting method computation complexity is very high (especially average weighted mutation method), it may be said that multinomial
Formula fitting is that a complexity and time-consuming process, according to the times N of sliding, need for average weighted mutation method
N times fitting of a polynomial is done, this is a very time-consuming process.In the weaker platform of the computing capabilitys such as mobile platform or microcontroller
On, which implements extremely difficult, is unfavorable for practicing on a large scale.
In this regard, the present invention propose a kind of posture at the new following a certain moment of the attitude transducer to wearing display equipment into
The scheme of row prediction, to promote precision of prediction and reduce computation complexity.The attitude prediction method of the embodiment of the present invention thinks, by
It is that the head of user or hand is followed to be moved in wearing the controller of display equipment, and the movement on the head of people and hand
It is limited by muscle, joint, posture curve when movement in space can form some specific patterns.It is adopted in attitude transducer
In the case of sample rate is sufficiently high, these specific patterns can be captured by attitude transducer, and be embodied in attitude transducer
In the attitude data returned.
Based on this, the embodiment of the present invention proposes a kind of completely new attitude prediction method, introduces machine learning and depth god
It through network, and acquires practical attitude data of the big quantity sensor in VR/AR applications and neural network is trained, make the nerve
Network can identify the pattern in attitudes vibration curve, to make high-precision prediction to the following posture sometime.
To make the object, technical solutions and advantages of the present invention clearer, below in conjunction with attached drawing to embodiment party of the present invention
Formula is described in further detail.
Fig. 2 is the flow chart of the attitude prediction method of one embodiment of the invention, and referring to Fig. 2, the posture of the present embodiment is pre-
Survey method, including：
Step S201, according to preset sample frequency, the attitude data that attitude transducer generates when to user movement is adopted
Sample obtains raw data set；Here the attitude data before raw data set is, for example, current time in some period, than
Such as, the attitude data that corresponding countershaft on time arranges from the old to the new in [- Δ T, 0] section, 0 indicates current time (time coordinate
At 0 point of axis), negative sign "-" represents direction, and at the time of indicating older than current time, Δ T indicates time span, such as 6 seconds.
Step S202 determines the number of plies of initial neural network model and each according to raw data set and specified prediction duration
The number of nodes of layer, and trained using the initial neural network model of the initial data set pair, to obtain the nerve net of training completion
Network model；
Here, first according to the complexity of raw data set, and to predicted time length demand (such as, it is desirable to it is pre-
The posture after 5 milliseconds is surveyed, then 5 milliseconds here are prediction duration) determine an initial neural network model, then, with
It is trained afterwards using initial data set pair model, obtains the neural network model eventually for prediction.
It should be noted that in practical application, predict that the demand of duration is different, for example, in the case of one kind, to predict 5 millis
Posture after second, in another case, to predict the posture after 2 milliseconds.Predict that the demand of duration is different, in raw data set phase
With in the case of, initial neural network model also different (the mainly number of plies of neural network model and each node layers of structure
Number is different).
In the present embodiment, neural network model is trained using raw data set, since raw data set is represented over
In a period of time (that is, using certain moment as starting point, the time interval pair that is determined as terminal using a certain moment in the negative direction of time shaft
The attitude data answered) attitude data, and original number is utilized not merely with the attitude data that some time carves in prediction in this way
According to the deeper movable information that collection includes, the accuracy of attitude prediction is improved.
Attitude data in the predetermined amount of time of acquisition is input to instruction by step S203 during an attitude prediction
Practice in the neural network model completed, obtains object time pair after the current time of the neural network model output of training completion
The attitude data answered；
Wherein, predetermined amount of time is the period determined by the predetermined instant before current time and current time.
As shown in Figure 2 it is found that the attitude prediction method of the present embodiment, neural network model is trained using machine learning,
To by it is larger it is complexity optimized be transferred in off-line model training, ensure that computation complexity is low when actual prediction, meets
Efficiency requirements are realized simple.In addition, some motor patterns may be frequently occurred during using head-mounted display apparatus,
Then the acceleration rotation of such as user's head, quick rotation to certain position stop isotype suddenly, these motor patterns are in posture
There is respective feature in the time series of data, and neural network model can capture the small fortune in history attitude data
Dynamic model formula, after being trained based on the small motor pattern captured in history attitude data, even if the posture of user
Acute variation occurs for direction, and neural network model is maintained to higher precision of prediction, the prediction with polynomial fitting method
Precision is compared, and polynomial fitting method is only simply fitted historical movement change curve, is occurred in this way in trend direction
When acute variation, the trend that the curve that fits rises or falls in future time point is still identical as historical trend direction, and with
Actual trend direction is on the contrary, to predict that error is larger.In this way, accurately being predicted using the neural network model of machine learning
Posture, facilitates to wear display equipment and shown calculating wear-type in advance according to the higher attitude data of the accuracy predicted and sets
Standby picture makes the picture of display and the movement which matches of user, to reduce delay, promotes the feeling of immersion experience of user.
In one embodiment of the invention, attitude prediction method includes three parts, respectively：(I) attitude data is acquired,
(II) training neural network model, (III) realize attitude prediction using the neural network model trained, are said individually below
It is bright.
First, attitude data is acquired.
In the present embodiment, according to preset sample frequency, the attitude data that attitude transducer generates when to user movement carries out
Sampling, obtains raw data set.It specifically includes：According to preset sample frequency, attitude transducer normally makes when to user movement
It is sampled with the attitude data generated under state, obtains the time series (Time Series) of attitude data, preserve,
As raw data set；The time series of attitude data indicates one or more in following user's head or hand exercise：It is flat
Trackslip it is dynamic, accelerate rotation, underdrive, predetermined time to turn to predeterminated position after stopping, predetermined time turn to precalculated position
Again to back rotation.Here predetermined time stops after turning to predeterminated position, for example is that head goes to the right side from initial position in 3 seconds
It is maintained at angle certain time behind 45 degree of side.Predetermined time turns to precalculated position again to back rotation, than if so, head in 3 seconds
Head is gone to initial position by portion in 2 seconds again after going to 45 degree of the right from initial position.Wherein, the sample frequency of attitude data
E.g. 1000Hz, then raw data set is once to sample obtained time series every 1ms progress.
Secondly, training neural network.
In the present embodiment, the number of plies of neural network and the section of each layer are determined according to the raw data set that aforementioned sample obtains
Points.It specifically includes：Raw data set is arranged, the attitude data and object time in the indicating predetermined period are generated
Attitude data mapping relations mapping relations collection；The data that mapping relations are concentrated are input to initial neural network model
In, and use supervised learning training neural network model.
In practical application, a deep neural network is first designed.It is appreciated that the number of plies of neural network and each layer
Depending on number of nodes is by actual conditions such as the complexity of data set, the length in historical time section, the length of predicted time, this
Embodiment is not restricted this.
Followed by the data set collected, neural network is allowed to exercise supervision formula study, after trained in this way, can obtained
The neural network model for being used for attitude prediction to one.Inputting arbitrary length to the neural network model in actual prediction is
ΔT1Posture array, for example the time where the array the last one element is Z, then, the neural network model, that is, exportable Z
+ΔT2The posture of place prediction.
In the present embodiment, by above step, training obtains lower array function：
Z+T2=f ([Z- Δs T1,Z])
The function may migrate on each platform and be realized, specifically, after the completion of neural network model is trained, extraction
And export specified parameter (number of plies of neural network model, the number of nodes of each layer, the activation letter of the neural network model of training completion
Number, loss function etc.) value, to running neural network model on designated terminal platform.Here designated terminal platform packet
Include but be not limited to PC platforms, mobile platform, microcontroller etc., it is known that, the method for the present embodiment is low to Platform Requirements, convenient extensive
It promotes and applies.
It is emphasized that the present embodiment, during using supervised learning training neural network model, meeting is according to instruction
Practice the specified parameter of result adjustment neural network model；Specified parameter includes：The node of the number of plies of neural network model, each layer
Number, activation primitive and loss function.The parameter for extracting the model trained realizes corresponding prediction letter on required platform
Number.Here the specified parameter that neural network model why is adjusted according to training result, mainly determines according to the result of test
Whether prediction result has reached the error requirements of application and whether neural network model produces over-fitting or poor fitting, from
And precision of prediction is improved, reduce error.For example, theoretically, when training neural network using supervised learning, with the increasing of iteration
Add, the accuracy of training set and test set should rise that (in spite of the presence of the error of fitting, the accuracy of test set does not have
The accuracy of training set is so high).But it in practice, is can be well realized when neural network fitting training set, loss function is very
Small, accuracy is very big；But loss function is very big when being fitted test set, accuracy fluctuates in a relatively low range,
I.e. training set prediction effect is good, and test set prediction effect is poor, intends in this way when initial training and test result show to have occurred
When conjunction, the data of node in neural network model middle layer or middle layer are reduced, or dropout is added, train again and tests,
Adjusting repeatedly can terminate to train until error is met the requirements.Here dropout refers to the instruction in deep learning neural network
During white silk, for neural network unit, it is temporarily abandoned from network according to certain probability.
It is emphasized that in the present embodiment, the loss function of neural network model is set as adopting according to training result
The mean square deviation of the practical attitude data value of collection and prediction attitude data value.That is, loss function is practical attitude value and prediction posture
The mean square deviation of value, which ensures that the difference minimums between predicted value and actual value, to which error is smaller.Here loss function
(loss function) is also cost function (cost function), is the object function of Neural Network Optimization, neural network
The process of training or optimization is exactly to minimize the process of loss function, and loss function value is smaller, the result of corresponding prediction and very
The value of real result is with regard to closer.
Finally, the model realization attitude prediction trained is utilized.
Prediction process is：Attitude data in the predetermined amount of time of acquisition is input in neural network model, god is obtained
Object time (such as 0+ Δs T after current time through network model output2) corresponding attitude data.Predetermined amount of time is such as
[-ΔT1, 0] and (negative sign here indicates direction, is by current time (such as 0 point) and the predetermined instant before current time
The period of (23 points) determinations.Here [- Δ T1, 0] and what is stored in time interval is the time series of aforementioned attitude data, be
Countershaft arranges from the old to the new on time, and each data, which represents, once samples obtained posture numerical value, between front and back two data
Time interval be sampling period (inverse of sample frequency).
So far, in one embodiment of the invention, the attitude prediction method passes through the data collection mould in controller
Block reads and preserves the initial data that the attitude controllers such as HMD are returned and (includes the number of the acquisitions such as gyroscope, accelerometer
According to).Then, the data being collected into are arranged, is generated a series of shaped like [- Δ T1,0]→ΔT2Mapping relations data.It connects
It, the mapping relations data of above-mentioned generation is updated in neural network model and are trained.Nerve is adjusted according to training result
The structure of network, the accuracy of Optimization Prediction, until precision of prediction is met the requirements.In training process, the parameter packet that mainly adjusts
Include the number of plies of neural network, Batch Size (total sample number in a Batch), Epoch (complete data set), training set
With the important parameters such as the size of test set.Finally, the neural network model that application training is completed realizes Accurate Prediction.
It should be noted that in neural network, when a complete data set has passed through neural network once and returns
It has returned once, this process is known as an epoch.And when cannot data disposably be passed through neural network, it is necessary to will count
It is divided into several batch according to collection.
The attitude prediction method of the present embodiment realizes following advantageous effect：Prediction side based on machine learning neural network
Method, computation complexity when prediction is relatively low, between newtonian motion method and polynomial fitting method.What machine learning trained
Neural network model, computationally shows as a series of vector-matrix multiplication and addition, specific complexity mainly depend on
The size and the number of plies of matrix and vector.Although complexity of the neural network in training is higher, by operation in the present embodiment
It measures maximum Function Fitting to carry out in off-line training, computation complexity when predicting in real time in this way is relatively low, in a way may be used
It is considered complexity optimized to be transferred to larger in off-line training.
In addition, computational accuracy is more preferable compared with polynomial fitting method.The prediction technique based on neural network model of the present embodiment
When acute variation occurs for posture direction, the precision of prediction is still higher.It is gone through the reason is that neural network can be captured in training
Small motor pattern in history posture, such as the smooth pivotal on head, acceleration rotation, underdrive, quick rotation to certain position
Then again to back rotation etc., these may be frequently during using equipment for stopping, quick rotation to some position suddenly
The motor pattern of appearance has respective feature, neural network can be according to these features in the time series of attitude data
More acurrate, unified data prediction is carried out, to ensure that the precision of prediction of the present embodiment.
Fig. 3 is the prediction result comparison diagram for the attitude prediction method for applying one embodiment of the invention, in figure 3 horizontal axis table
Show that time, the longitudinal axis indicate to indicate attitude transducer according to the calculated Eulerian angles roll values of IMU initial data, unit (degree), 302
Actual attitude value, 301 indicate the predicted value of the present embodiment attitude prediction method, from the figure 3, it may be seen that predicted value and posture actual value
Closely, this also illustrates the attitude prediction method precision of prediction of the present embodiment is higher, practical application request is met, favorably
Delay is reduced in wearing display equipment, the user experience is improved.
Fig. 4 is the block diagram of the attitude prediction device of one embodiment of the invention, and referring to Fig. 4, attitude prediction device 400 is applied
In wearing display equipment, including：
Sampling unit 401 is used for according to preset sample frequency, the attitude data that attitude transducer generates when to user movement
It is sampled, obtains raw data set；
Training unit 402, the layer for determining initial neural network model according to raw data set and specified prediction duration
The number of nodes of several and each layer, and trained using the initial neural network model of the initial data set pair, to obtain training completion
Neural network model；
Predicting unit 403 is used for during an attitude prediction, and the attitude data in the predetermined amount of time of acquisition is defeated
Enter in the neural network model completed to training, obtains target after the current time of the neural network model output of training completion
Moment corresponding attitude data；
Wherein, predetermined amount of time is the period determined by the predetermined instant before current time and current time.
In one embodiment of the invention, training unit 402 are generated specifically for being arranged to raw data set
The mapping relations collection of the mapping relations of the attitude data of attitude data and object time in the indicating predetermined period；
The data that mapping relations are concentrated are input in initial neural network model, and using supervised learning training nerve
Network model.
In one embodiment of the invention, attitude prediction device 400 further includes：
Training optimization unit, is used for during using supervised learning training neural network model, according to training result
Adjust the specified parameter of neural network model；Specified parameter includes：The number of plies of neural network model, the number of nodes of each layer, activation
Function and loss function.
In one embodiment of the invention, training optimization unit is specifically used for, according to training result by neural network mould
The loss function of type is set as the mean square deviation of the practical attitude data value and prediction attitude data value of acquisition.
In one embodiment of the invention, attitude prediction device 400 further includes：Platform expanding element, for extracting simultaneously
The value of the specified parameter for the neural network model that output training is completed, to run the nerve that training is completed on designated terminal platform
Network model.
It should be noted that the attitude prediction device of the present embodiment is corresponding with aforementioned attitude prediction method, thus
The content not described to attitude prediction device in the present embodiment can be found in the explanation in preceding method embodiment, no longer superfluous here
It states.
Fig. 5 is the block diagram of the electronic equipment of one embodiment of the invention, and electronic equipment includes：Memory 501 and processor
502, it is communicated and is connected by internal bus 503 between the memory 501 and the processor 502, the memory 501 stores
There are the program instruction that can be executed by the processor 502, described program instruction that can be realized when being executed by the processor 502
In previous embodiment the step of attitude prediction method.
It should be understood by those skilled in the art that, the embodiment of the present invention can be provided as method, system or computer program
Product.Therefore, complete hardware embodiment, complete software embodiment or reality combining software and hardware aspects can be used in the present invention
Apply the form of example.Moreover, the present invention can be used in one or more wherein include computer usable program code computer
The computer program production implemented in usable storage medium (including but not limited to magnetic disk storage, CD-ROM, optical memory etc.)
The form of product.
The above description is merely a specific embodiment, under the above-mentioned introduction of the present invention, those skilled in the art
Other improvement or deformation can be carried out on the basis of the above embodiments.It will be understood by those skilled in the art that above-mentioned tool
Body description only preferably explains that the purpose of the present invention, protection scope of the present invention should be subject to the protection scope in claims.
Claims (10)
1. a kind of attitude prediction method, which is characterized in that including：
According to preset sample frequency, the attitude data that attitude transducer generates when to user movement samples, and obtains original number
According to collection；
The number of plies of initial neural network model and the node of each layer are determined according to the raw data set and specified prediction duration
Number, and trained using the initial neural network model of the initial data set pair, to obtain the neural network model of training completion；
During an attitude prediction, the attitude data in the predetermined amount of time of acquisition is input to the god that the training is completed
The corresponding appearance of object time after current time through in network model, obtaining the neural network model output that the training is completed
State data；
Wherein, the predetermined amount of time is the period determined by the predetermined instant before current time and current time.
2. according to the method described in claim 1, it is characterized in that, being determined according to the raw data set and specified prediction duration
The initial number of plies of neural network model and the number of nodes of each layer, and use the initial data set pair neural metwork training packet
It includes：
The raw data set is arranged, the attitude data and object time indicated in the predetermined amount of time is generated
The mapping relations collection of the mapping relations of attitude data；
The data that the mapping relations are concentrated are input in initial neural network model, and using supervised learning training nerve
Network model.
3. according to the method described in claim 2, it is characterized in that, further including：
During using supervised learning training neural network model, the specified of neural network model is adjusted according to training result
Parameter；
The specified parameter includes：The number of plies of neural network model, the number of nodes of each layer, activation primitive and loss function.
4. according to the method described in claim 3, it is characterized in that, the finger for adjusting neural network model according to training result
Determining parameter includes：
The loss function of neural network model is set to according to training result the practical attitude data value and prediction posture of acquisition
The mean square deviation of data value.
5. according to the method described in claim 1, it is characterized in that, further including：
After the completion of neural network model is trained, the value of the specified parameter of the neural network model of training completion is extracted and exports,
To run the neural network model on designated terminal platform.
6. according to the method described in claim 1, it is characterized in that, according to preset sample frequency, posture passes when to user movement
The attitude data that sensor generates is sampled, and obtaining raw data set includes：
According to preset sample frequency, the attitude data that attitude transducer generates under normal operating condition when to user movement carries out
Sampling, obtains the time series of attitude data, as raw data set；
The time series of the attitude data indicates one or more in following user's head or hand exercise：
Smooth pivotal, accelerate rotation, underdrive, predetermined time to turn to predeterminated position after stopping, predetermined time turn to it is pre-
Positioning is set again to back rotation.
7. a kind of attitude prediction device, which is characterized in that applied to wearing display equipment, including：
Sampling unit, for according to preset sample frequency, the attitude data that attitude transducer generates when to user movement to be adopted
Sample obtains raw data set；
Training unit, for according to the raw data set and specified prediction duration determine the number of plies of initial neural network model with
And the number of nodes of each layer, and trained using the initial neural network model of the initial data set pair, to obtain the god of training completion
Through network model；
Predicting unit, for during an attitude prediction, the attitude data in the predetermined amount of time of acquisition to be input to institute
In the neural network model for stating training completion, mesh after the current time for the neural network model output that the training is completed is obtained
Mark moment corresponding attitude data；
Wherein, the predetermined amount of time is the period determined by the predetermined instant before current time and current time.
8. device according to claim 7, which is characterized in that
The training unit is generated and is indicated in the predetermined amount of time specifically for being arranged to the raw data set
The mapping relations collection of the mapping relations of the attitude data of attitude data and object time；
The data that the mapping relations are concentrated are input in initial neural network model, and using supervised learning training nerve
Network model.
9. device according to claim 8, which is characterized in that further include：
Training optimization unit, for during using supervised learning training neural network model, being adjusted according to training result
The specified parameter of neural network model；
The specified parameter includes：The number of plies of neural network model, the number of nodes of each layer, activation primitive and loss function.
10. device according to claim 9, which is characterized in that the training optimization unit is specifically used for, and is tied according to training
Fruit sets the loss function of neural network model to the mean square deviation of the practical attitude data value and prediction attitude data value of acquisition；
Device further includes：
Platform expanding element, the value of the specified parameter of the neural network model for extracting and exporting training completion, with specified
The neural network model that the training is completed is run on terminal platform.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810300681.1A CN108664122A (en) | 2018-04-04 | 2018-04-04 | A kind of attitude prediction method and apparatus |
| PCT/CN2018/112854 WO2019192172A1 (en) | 2018-04-04 | 2018-10-31 | Attitude prediction method and apparatus, and electronic device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810300681.1A CN108664122A (en) | 2018-04-04 | 2018-04-04 | A kind of attitude prediction method and apparatus |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108664122A true CN108664122A (en) | 2018-10-16 |
Family
ID=63782107
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810300681.1A Pending CN108664122A (en) | 2018-04-04 | 2018-04-04 | A kind of attitude prediction method and apparatus |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN108664122A (en) |
| WO (1) | WO2019192172A1 (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109794935A (en) * | 2019-01-09 | 2019-05-24 | 中国工程物理研究院激光聚变研究中心 | Assembling process power and torque forecasting system and method based on neural network analysis |
| CN110068326A (en) * | 2019-04-29 | 2019-07-30 | 京东方科技集团股份有限公司 | Computation method for attitude, device, electronic equipment and storage medium |
| CN110248178A (en) * | 2019-06-18 | 2019-09-17 | 深圳大学 | Utilize the viewport prediction technique and system of object tracking and historical track panoramic video |
| WO2019192172A1 (en) * | 2018-04-04 | 2019-10-10 | 歌尔股份有限公司 | Attitude prediction method and apparatus, and electronic device |
| CN111771141A (en) * | 2019-01-30 | 2020-10-13 | 百度时代网络技术（北京）有限公司 | LIDAR positioning in autonomous vehicles using 3D CNN networks for solution inference |
| CN112100841A (en) * | 2020-09-09 | 2020-12-18 | 中铁二十局集团有限公司 | Shield tunneling machine attitude prediction method and device, terminal equipment and storage medium |
| CN112649815A (en) * | 2019-10-10 | 2021-04-13 | 华为技术有限公司 | Method and device for processing data |
| CN112818898A (en) * | 2021-02-20 | 2021-05-18 | 北京字跳网络技术有限公司 | Model training method and device and electronic equipment |
| CN112947180A (en) * | 2021-02-04 | 2021-06-11 | 中国地质大学（武汉） | Heavy machinery operation state identification and prediction method, device, equipment and storage medium |
| WO2021190421A1 (en) * | 2020-03-27 | 2021-09-30 | 海信视像科技股份有限公司 | Virtual reality-based controller light ball tracking method on and virtual reality device |
| CN113486214A (en) * | 2021-07-23 | 2021-10-08 | 广州酷狗计算机科技有限公司 | Music matching method and device, computer equipment and storage medium |
| CN113537489A (en) * | 2021-07-09 | 2021-10-22 | 厦门大学 | Elbow angle prediction method, terminal device and storage medium |
| CN113616494A (en) * | 2020-05-08 | 2021-11-09 | 未来穿戴技术有限公司 | Massage control method, massage control device, computer equipment and computer readable storage medium |
| CN113792484A (en) * | 2021-09-07 | 2021-12-14 | 珠海格力电器股份有限公司 | Cargo impact attitude model construction method and platform |
| CN113804189A (en) * | 2021-09-06 | 2021-12-17 | 国家海洋局北海预报中心((国家海洋局青岛海洋预报台)(国家海洋局青岛海洋环境监测中心站)) | INS/CNS-based integrated navigation method |
| CN113836795A (en) * | 2021-09-07 | 2021-12-24 | 珠海格力电器股份有限公司 | Method and platform for building initial cargo placement attitude model |
| CN114296539A (en) * | 2020-10-07 | 2022-04-08 | 财团法人工业技术研究院 | Direction prediction method, virtual reality device and non-transitory computer readable medium |
| US20220137184A1 (en) * | 2020-10-30 | 2022-05-05 | KaiKuTek Inc. | Impulse-like Gesture Recognition Method, and Impulse-like Gesture Recognition System |
| CN113836795B (en) * | 2021-09-07 | 2024-06-07 | 珠海格力电器股份有限公司 | Method and platform for constructing initial goods placement attitude model |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111027692A (en) * | 2019-11-04 | 2020-04-17 | 中国电子科技集团公司第十四研究所 | Target motion situation prediction method and device |
| CN111343340A (en) * | 2020-03-11 | 2020-06-26 | 杭州十域科技有限公司 | Posture distinguishing method of low-end hardware equipment based on big data |
| CN111311714A (en) * | 2020-03-31 | 2020-06-19 | 北京慧夜科技有限公司 | Attitude prediction method and system for three-dimensional animation |
| CN112486318A (en) * | 2020-11-26 | 2021-03-12 | 北京字跳网络技术有限公司 | Image display method, image display device, readable medium and electronic equipment |
| CN113268859B (en) * | 2021-04-25 | 2023-07-14 | 北京控制工程研究所 | Simulation method, system and storage medium for spacecraft on-orbit game |
| CN113221681B (en) * | 2021-04-26 | 2023-09-26 | 平安科技（深圳）有限公司 | Model training method, action gesture generating method, device, equipment and medium |
| CN113569327B (en) * | 2021-08-10 | 2024-02-09 | 同济大学 | Method and system for optimizing posture of shield tunneling machine and method and system for training model |
| CN113806172B (en) * | 2021-09-14 | 2024-02-06 | 北京百度网讯科技有限公司 | Method and device for processing equipment state parameters, electronic equipment and storage medium |
| CN113919236A (en) * | 2021-10-29 | 2022-01-11 | 中国国家铁路集团有限公司 | Model generation method for predicting track irregularity and prediction method for track irregularity |
| CN114096010B (en) * | 2021-11-01 | 2024-05-07 | 江苏科技大学 | Monitoring-oriented underwater sensor network energy continuous management method |
| CN114187360B (en) * | 2021-12-14 | 2024-02-06 | 西安交通大学 | Head pose estimation method based on deep learning and quaternion |
| CN116321286B (en) * | 2023-03-13 | 2023-09-22 | 湖北华中电力科技开发有限责任公司 | Data deviation identification system applying intelligent algorithm |
| CN117560638B (en) * | 2024-01-10 | 2024-03-22 | 山东派蒙机电技术有限公司 | Converged communication method, device and equipment applied to mobile terminal communication system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778579A (en) * | 2016-12-07 | 2017-05-31 | 电子科技大学 | A kind of head pose estimation method based on accumulative attribute |
| CN107403440A (en) * | 2016-05-18 | 2017-11-28 | 株式会社理光 | For the method and apparatus for the posture for determining object |
| WO2018005433A1 (en) * | 2016-06-27 | 2018-01-04 | Robin Young | Dynamically managing artificial neural networks |
| CN107688773A (en) * | 2017-07-07 | 2018-02-13 | 北京联合大学 | A kind of gesture identification method based on deep learning |
| CN107766844A (en) * | 2017-11-13 | 2018-03-06 | 杭州有盾网络科技有限公司 | Method, apparatus, equipment of a kind of reticulate pattern according to recognition of face |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103984416B (en) * | 2014-06-10 | 2017-02-08 | 北京邮电大学 | Gesture recognition method based on acceleration sensor |
| US10691698B2 (en) * | 2014-11-06 | 2020-06-23 | International Business Machines Corporation | Automatic near-real-time prediction, classification, and notification of events in natural language systems |
| CN106682733B (en) * | 2016-11-07 | 2018-10-19 | 上海资誉电子科技有限公司 | Unmanned plane motion state analysis method and device |
| CN108664122A (en) * | 2018-04-04 | 2018-10-16 | 歌尔股份有限公司 | A kind of attitude prediction method and apparatus |
-
2018
- 2018-04-04 CN CN201810300681.1A patent/CN108664122A/en active Pending
- 2018-10-31 WO PCT/CN2018/112854 patent/WO2019192172A1/en active Application Filing
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107403440A (en) * | 2016-05-18 | 2017-11-28 | 株式会社理光 | For the method and apparatus for the posture for determining object |
| WO2018005433A1 (en) * | 2016-06-27 | 2018-01-04 | Robin Young | Dynamically managing artificial neural networks |
| CN106778579A (en) * | 2016-12-07 | 2017-05-31 | 电子科技大学 | A kind of head pose estimation method based on accumulative attribute |
| CN107688773A (en) * | 2017-07-07 | 2018-02-13 | 北京联合大学 | A kind of gesture identification method based on deep learning |
| CN107766844A (en) * | 2017-11-13 | 2018-03-06 | 杭州有盾网络科技有限公司 | Method, apparatus, equipment of a kind of reticulate pattern according to recognition of face |
Cited By (32)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019192172A1 (en) * | 2018-04-04 | 2019-10-10 | 歌尔股份有限公司 | Attitude prediction method and apparatus, and electronic device |
| CN109794935B (en) * | 2019-01-09 | 2022-01-14 | 中国工程物理研究院激光聚变研究中心 | Neural network analysis-based assembly process force and moment prediction system and method |
| CN109794935A (en) * | 2019-01-09 | 2019-05-24 | 中国工程物理研究院激光聚变研究中心 | Assembling process power and torque forecasting system and method based on neural network analysis |
| CN111771141A (en) * | 2019-01-30 | 2020-10-13 | 百度时代网络技术（北京）有限公司 | LIDAR positioning in autonomous vehicles using 3D CNN networks for solution inference |
| CN111771141B (en) * | 2019-01-30 | 2024-04-09 | 百度时代网络技术（北京）有限公司 | LIDAR positioning for solution inference using 3D CNN network in autonomous vehicles |
| CN110068326A (en) * | 2019-04-29 | 2019-07-30 | 京东方科技集团股份有限公司 | Computation method for attitude, device, electronic equipment and storage medium |
| CN110068326B (en) * | 2019-04-29 | 2021-11-30 | 京东方科技集团股份有限公司 | Attitude calculation method and apparatus, electronic device, and storage medium |
| CN110248178B (en) * | 2019-06-18 | 2021-11-23 | 深圳大学 | Viewport prediction method and system using object tracking and historical track panoramic video |
| CN110248178A (en) * | 2019-06-18 | 2019-09-17 | 深圳大学 | Utilize the viewport prediction technique and system of object tracking and historical track panoramic video |
| CN112649815A (en) * | 2019-10-10 | 2021-04-13 | 华为技术有限公司 | Method and device for processing data |
| WO2021190421A1 (en) * | 2020-03-27 | 2021-09-30 | 海信视像科技股份有限公司 | Virtual reality-based controller light ball tracking method on and virtual reality device |
| CN113616494A (en) * | 2020-05-08 | 2021-11-09 | 未来穿戴技术有限公司 | Massage control method, massage control device, computer equipment and computer readable storage medium |
| CN113616494B (en) * | 2020-05-08 | 2023-07-04 | 未来穿戴技术有限公司 | Massage control method, massage control device, computer equipment and computer readable storage medium |
| CN112100841B (en) * | 2020-09-09 | 2024-04-19 | 中铁二十局集团有限公司 | Method and device for predicting attitude of shield tunneling machine, terminal equipment and storage medium |
| CN112100841A (en) * | 2020-09-09 | 2020-12-18 | 中铁二十局集团有限公司 | Shield tunneling machine attitude prediction method and device, terminal equipment and storage medium |
| CN114296539A (en) * | 2020-10-07 | 2022-04-08 | 财团法人工业技术研究院 | Direction prediction method, virtual reality device and non-transitory computer readable medium |
| CN114296539B (en) * | 2020-10-07 | 2024-01-23 | 财团法人工业技术研究院 | Direction prediction method, virtual reality device and non-transitory computer readable medium |
| US11892562B2 (en) * | 2020-10-30 | 2024-02-06 | KaiKuTek Inc. | Impulse-like gesture recognition method, and impulse-like gesture recognition system |
| US20220137184A1 (en) * | 2020-10-30 | 2022-05-05 | KaiKuTek Inc. | Impulse-like Gesture Recognition Method, and Impulse-like Gesture Recognition System |
| CN112947180A (en) * | 2021-02-04 | 2021-06-11 | 中国地质大学（武汉） | Heavy machinery operation state identification and prediction method, device, equipment and storage medium |
| CN112947180B (en) * | 2021-02-04 | 2022-06-24 | 中国地质大学（武汉） | Heavy machinery operation state identification and prediction method, device, equipment and storage medium |
| CN112818898B (en) * | 2021-02-20 | 2024-02-20 | 北京字跳网络技术有限公司 | Model training method and device and electronic equipment |
| CN112818898A (en) * | 2021-02-20 | 2021-05-18 | 北京字跳网络技术有限公司 | Model training method and device and electronic equipment |
| CN113537489A (en) * | 2021-07-09 | 2021-10-22 | 厦门大学 | Elbow angle prediction method, terminal device and storage medium |
| CN113537489B (en) * | 2021-07-09 | 2024-03-19 | 厦门大学 | Elbow angle prediction method, terminal equipment and storage medium |
| CN113486214A (en) * | 2021-07-23 | 2021-10-08 | 广州酷狗计算机科技有限公司 | Music matching method and device, computer equipment and storage medium |
| CN113804189A (en) * | 2021-09-06 | 2021-12-17 | 国家海洋局北海预报中心((国家海洋局青岛海洋预报台)(国家海洋局青岛海洋环境监测中心站)) | INS/CNS-based integrated navigation method |
| CN113804189B (en) * | 2021-09-06 | 2024-05-14 | 国家海洋局北海预报中心((国家海洋局青岛海洋预报台)(国家海洋局青岛海洋环境监测中心站)) | INS/CNS-based integrated navigation method |
| CN113836795A (en) * | 2021-09-07 | 2021-12-24 | 珠海格力电器股份有限公司 | Method and platform for building initial cargo placement attitude model |
| CN113792484A (en) * | 2021-09-07 | 2021-12-14 | 珠海格力电器股份有限公司 | Cargo impact attitude model construction method and platform |
| CN113792484B (en) * | 2021-09-07 | 2024-02-27 | 珠海格力电器股份有限公司 | Cargo impact attitude model construction method and platform |
| CN113836795B (en) * | 2021-09-07 | 2024-06-07 | 珠海格力电器股份有限公司 | Method and platform for constructing initial goods placement attitude model |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2019192172A1 (en) | 2019-10-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108664122A (en) | A kind of attitude prediction method and apparatus | |
| CN108115682B (en) | Rote learning device, robot system and learning by rote | |
| Popović et al. | Motion sketching for control of rigid-body simulations | |
| CN111402290B (en) | Action restoration method and device based on skeleton key points | |
| Cheema et al. | Predicting mid-air interaction movements and fatigue using deep reinforcement learning | |
| CN109064487B (en) | Human body posture comparison method based on Kinect skeleton node position tracking | |
| CN107851216A (en) | Train intensified learning neutral net | |
| CN106548675A (en) | Virtual military training method and device | |
| CN108803499A (en) | control device and machine learning device | |
| CN108230429A (en) | Real-time whole body posture reconstruction method based on head and two-hand positions and posture | |
| CN107791246A (en) | Robot simulation apparatus | |
| US11731052B2 (en) | Animation and physics synchronization | |
| US11644892B2 (en) | Synthesizing haptic and sonic feedback for textured materials in interactive virtual environments | |
| US11238634B2 (en) | Motion model refinement based on contact analysis and optimization | |
| JPWO2020049847A1 (en) | Estimator, learning device, estimation method, learning method and program | |
| CN108051001A (en) | A kind of robot movement control method, system and inertia sensing control device | |
| CN111459280A (en) | VR space expansion method, device, equipment and storage medium | |
| CN101952879A (en) | Methods and apparatus for designing animatronics units from articulated computer generated characters | |
| CN114296539B (en) | Direction prediction method, virtual reality device and non-transitory computer readable medium | |
| US20220057857A1 (en) | Computer-implemented simulation of a virtual hand | |
| CN115035909A (en) | Music visual display method and device | |
| US7583262B2 (en) | Optimization of time-critical software components for real-time interactive applications | |
| US9275488B2 (en) | System and method for animating a body | |
| Zhou et al. | A mixed reality training system for hand-object interaction in simulated microgravity environments | |
| Clarke et al. | The effects of simulated inertia and force prediction on delayed telepresence |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181016 |