CN110163892A - Learning rate Renewal step by step method and dynamic modeling system based on estimation interpolation - Google Patents
Learning rate Renewal step by step method and dynamic modeling system based on estimation interpolation Download PDFInfo
- Publication number
- CN110163892A CN110163892A CN201910375212.0A CN201910375212A CN110163892A CN 110163892 A CN110163892 A CN 110163892A CN 201910375212 A CN201910375212 A CN 201910375212A CN 110163892 A CN110163892 A CN 110163892A
- Authority
- CN
- China
- Prior art keywords
- frame
- module
- learning rate
- model
- mdl
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Television Systems (AREA)
- Processing Or Creating Images (AREA)
Abstract
The invention discloses a kind of learning rate Renewal step by step method and dynamic modeling system based on estimation interpolation, this method is to utilize the target motion information in video data, the motion vector field of target is calculated in real time, estimation interpolation is used to original video data, generate Augmented Data frame, and Augmented Data frame smoothly more new model is used, to obtain accurate prediction model.Using prediction model, combined with learning rate dynamic adjustment module, when target quickly changes, increase learning rate, increase quick variation targets to the influence power of model, increases model adaptability, when blocking interference, reduce learning rate, background is reduced using learning rate dynamic adjustment to interfere object module bring, increases the robustness of model, in terms of two kinds, the problem of mitigating the middle object module degeneration of target dynamic modeling problem, so that target dynamic modeling work is more accurate.
Description
Technical field
The present invention relates to target dynamics to model field, progressive more particularly, to a kind of learning rate based on estimation interpolation
Update method and dynamic modeling system.
Background technique
In recent years, in target tracking domain, many algorithms are all closely bound up with the dynamic modeling of tracked target.
The main problem that they are faced is:
1. the model accuracy for modeling the preferable dynamic modelling method of real-time is insufficient.Recently generally using tool in tracking field
There is differentiation correlation filter (the Histograms of oriented gradients for human detection. of feature
Conference on Computer Vision and Pattern Recognition), since MOSSE tracker
(Visual object tracking using adaptive correlation filters. In:2010 IEEE
Computer Society Conference on Computer Vision and Pattern
Recognition.pp.2544-2550).Core correlation filter (KCF) (High-speed tracking with
kernelized correlation filters. IEEE Transactions on Pattern Analysis and
Machine Intelligence 37 (3), 583-596) promotion for computational efficiency of loop structure is further studied,
However loop structure also results in several the problem of being referred to as boundary effect, has an impact to accuracy.
2. the real-time for modeling accurate dynamic modelling method is lacking:
(Correlation filters with limited boundaries.In: 2015IEEE Conference on
Computer Vision and Pattern Recognition (CVPR) .pp.4630-4638) finite boundary CF
(CFLB)), spatial regularization CF(SRDCF) (Learning spatially regularized correlation
filters for visual tracking.In: 2015 IEEE International Conference on
Computer Vision.pp. 4310-431) and background correlation filter (BACF) (Learning background-
aware correlation filters for visual tracking. In: 2017 IEEE International
Conference on Computer Vision (ICCV) .pp. 1144-1152) all in order to be obtained more from training sample
Important information proposes different methods, to realize better trace model.It is proposed the CF(CFLB of finite boundary)
(Correlation filters with limited boundaries.In: 2015 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR) .pp.4630-4638) method can learn boundary
The less CF of effect.However, the feature that it is used is only based on pixel, this is proved to be not enough to expression rule in the picture
(Multi-channel correlation filters.In:2013 IEEE International Conference on
Computer Vision.pp.3072–3079) .Spatial regularization CF algorithm (SRDCF) has modified the optimization aim of tracker,
To retain more information around the center of image packet.Although this method has good performance in terms of prediction, it
Have the shortcomings that one it is main: specifically, the optimization cost of this method is very high since the optimization aim does not have analytic solutions.In addition,
One group of hyper parameter for being used to form regularization weight is desirable to carefully tune.If fail to adjust these hyper parameters may result in
Track performance is bad.Background correlation filter (BACF) is by the inclusion of more background informations from frame and generates negative training sample
Method, to promote classifier to separate background and foreground zone.Although BACF shows impressive, change training process
Structure lead to no analytic solutions, requirement of real-time is not achieved in calculating speed.
Depth convolutional network achieves major progress in terms of dynamic modeling recently, develops on this basis more and more
Track algorithm replace conventional model (the Long-term correlation tracking.2015 based on manual feature
IEEE Conference on Computer Vision and Pattern Recognition pp. 4847–4856).So
And due to the high calculation amount in training process, most of method (Learning multi-domain based on deep learning
convolutional neural networks for visual tracking.2016 IEEE Conference on
Computer Vision and Pattern Recognition. pp. 4293-4302) it is limited to high calculating cost, this makes
It obtains them and is not suitable for real-time objects modeling.In addition, the demand of GPU also increases the cost of equipment, which has limited depth methods
Application in real time embedded system.
The analysis of domestic and international related patents can be concluded that currently without similar utilization estimation interpolation and
Learning rate Renewal step by step is to mitigate the application on object module degenerate problem.
Summary of the invention
It is built for using the status and existing target dynamic that generate intermediate data and learning rate dynamic update method rareness
The deficiency of mould method carries out target the first purpose of the invention is to provide a kind of combining target motion information, to video data
Dynamic modeling, the learning rate Renewal step by step based on estimation interpolation to reduce model degradation, promote dynamic modeling accuracy
Method.
A second object of the present invention is to provide a kind of dynamic modeling systems for learning rate Renewal step by step method.
The first purpose of this invention is achieved in that
A kind of learning rate Renewal step by step method based on estimation interpolation, be characterized in: specific step is as follows:
It A, the use of adjacent two frame of original video data (t frame, t+1 frame) is one group of (ft, ft+1), carry out estimation: movement
Estimation (ME) passes through the identical entity in two consecutive frames of matching and calculates motion vector field (MVF) to complete;For t frame ft
With t+1 frame ft+1Between motion vector field ut, enable δtIndicate time interval, then generated on time t+λ δ t
Sports ground c may be expressed as: uλ(c+λut(c))=ut(c);For each hole region in interpolated movements field, by most adjacent
Close motion vector filling;
It B, the use of adjacent two frame of original video data (t frame, t+1 frame) is one group of (ft, ft+1) and step A in calculate institute
The motion vector field u obtainedt, augmentation obtains the sample { f of one group of carrying space time informationλ, by t frame ftWith t+1 frame frame ft+1It inserts
The λ interpolation results that value generates are represented as: fλ(c)=(1-λ)ft(c-λuλ(c)+λft+1(c+(1-λ)uλ(c));
C, using the data after augmentation, input as Target Modeling task: according to given target response Rt, extract
The existing model mdl of t framet;
D, according to object module, next frame is detected, calculates response Rt, RtPeak value of response max (R) is as the sound predicted
Answer Rt+1;
E, the response R of prediction is utilizedt+1Generate t frame prediction model Tmpt;
F, t frame prediction model Tmp is utilizedtWith the existing model mdl of t frametThe degenerate case of model is assessed, dynamic
Selected learning rate α;
G, according to t frame prediction model Tmpt, and selected learning rate α, update the existing model mdl of t framet, generate t frame
Object module mdlt+1, indicate are as follows:
mdlt+1= αTmpt+(1-α)mdlt;
H, step C to step G is repeated, R is responded according to t frametWith the object module mdl of t framet+1, by the target mould of t frame
Type mdlt+1Existing model mdl as t+1 framet+1Generate Target Modeling result ptWith the object module mdl of t+1 framet+2。
In step A into step B: original video data is with consecutive frame (t frame, t+1 frame) for one group of (ft, ft+1), quilt
For two aspects, in a first aspect, in step, according to (ft, ft+1), calculate motion vector field ut;Second aspect, in step B
In, according to (ft, ft+1) and A in the u that calculatest, calculate Augmented Data frame { fλ}。
In stepb: using motion vector field to legacy data augmentation, not needing additionally to demarcate, the data of generation are also made
It is inputted for new sample.
In stepb: Augmented Data frame { fλOriginal video data is contained in definition, in special circumstances, as λ=0, fλ
= ft, as λ=1, fλ= ft+1。
The step C is into step the G: object module can be converted into existing model with time change, such as t frame
Object module mdlt+1In t+1 frame, as existing model mdlt+1 。
Second object of the present invention is achieved in that
A kind of dynamic modeling system for learning rate Renewal step by step method, is characterized in: including sport interpolation estimation module and
The target dynamic modeling module of habit rate Renewal step by step, in which:
Sport interpolation estimation module includes motion estimation module and data augmentation module;
The target dynamic modeling module of learning rate Renewal step by step includes that object module extraction module, response computation module, model move back
Change evaluation module and model modification module;
By the data that sport interpolation estimation module generates, the target dynamic modeling module of learning rate Renewal step by step is inputted;That is:
A, sport interpolation estimation module carries out estimation to video contiguous frames and generates movement according to the original video data of input
Vector field ut, and then generate Augmented Data frame { fλ, by ut, { fλInput learning rate Renewal step by step target dynamic modeling module;
The motion estimation module: being one group of (f from adjacent two frame (t frame, t+1 frame) of original video datat, ft+1), root
According to (ft, ft+1) calculate all pixels motion vector field ut, for hole region, it is filled, will be calculated using nearest neighbor algorithm
Obtained motion vector field utInput data augmentation module;
The data augmentation module: being one group of (f from adjacent two frame (t frame, t+1 frame) of original video datat, ft+1), with
And the motion vector field u of motion estimation module inputt, to (ft, ft+1) video data progress data augmentation, utilize t frame frame ft
With t+1 frame frame ft+1, interpolation goes out the smallest Augmented Data frame { f of errorλ, and by Augmented Data frame { fλIncoming learning rate is progressive
The target dynamic modeling module of update;
B, the target dynamic modeling module of the learning rate Renewal step by step, according to Augmented Data frame { fλ, in conjunction with the fortune of contiguous frames
Dynamic information carries out learning rate dynamic and adjusts, and generates dynamic modeling result;Wherein:
The object module extraction module: as t=0, according to initially given response R0, calculate existing model mdl0Work as t > 0
When, the existing model mdl for the t frame for directly taking update to obtaint, input response computation module;
The response computation module: the existing model mdl of t frame is usedt, with t+1 frame ft+1In response RtNeighbouring region meter
Correlation is calculated, response R is obtainedt+1, and result R according to responset+1, calculate prediction model tmpt;
The model degradation evaluation module: calculated prediction model tmp is usedt, with existing model mdltIt is compared, for
A possibility that model deviates the case where central area, and target is fast-moving target is big, according to motion vector field utIt is moved
Rate increases learning rate α.For prediction model tmptCompared to existing model mdltAbsolute value declines obvious situation, then target shape
Become or to encounter a possibility that blocking big, uses Augmented Data frame { f at this timeλIn relevant range F and prediction model tmptIn frequency domain
It is multiplied and obtains response RF;If prediction model tmptWith assessment area F calculate gained response reduce present it is regional, then for
It blocks, according to serious shielding situation, learning rate α is accordingly decreased, if without regionality, for deformation, according to deformation degree,
Correspondingly increase learning rate α;Model degradation evaluation module is by α input model update module;
The model modification module: the learning rate α more new model obtained according to model degradation evaluation module, updated model
mdlt+1=(1-α)mdlt+αtmpt。
For original video data after input motion Interpolate estimation module, sport interpolation estimation module carries out video contiguous frames
Estimation generates intermediate interpolation frame, inputs the target dynamic modeling module of learning rate Renewal step by step together, and learning rate is progressive more
Video data after the augmentation that new target dynamic modeling module is inputted according to target dynamic modeling module, in conjunction with the fortune of contiguous frames
Dynamic information carries out learning rate dynamic and adjusts, and generates dynamic modeling result, it may be assumed that
The model modification module: the learning rate α more new model obtained according to model degradation evaluation module, updated model are
mdlt+1=(1-α)mdlt+αtmpt。
Compared with prior art, the present invention has the effect that
The present invention calculates the motion vector field of target, to original video counts using the target motion information in video data in real time
According to estimation interpolation is used, Augmented Data frame is generated, and uses Augmented Data frame smoothly more new model, it is accurate to obtain
Prediction model.It using prediction model, is combined with learning rate dynamic adjustment module, when target quickly changes, increases learning rate,
Increase quick variation targets to the influence power of model, increase model adaptability, when blocking interference, reduces learning rate, benefit
Background is reduced with learning rate dynamic adjustment to interfere object module bring, increases the robustness of model, in terms of two kinds, is mitigated
The problem of middle object module of target dynamic modeling problem is degenerated, balances real-time and accuracy, and slightly reducing, modeling is real
In the case where when property, the accuracy of dynamic modeling is improved with learning rate Renewal step by step method.
Detailed description of the invention
Fig. 1 is the structural representation of the target dynamic modeling module relationship of interpolation-movement estimation module and learning rate Renewal step by step
Figure;
Fig. 2 is the schematic diagram of interpolation-movement estimation module;
Fig. 3 is the structural schematic diagram of the target dynamic modeling module example of learning rate Renewal step by step.
Specific embodiment
Below against embodiment and in conjunction with attached drawing, the present invention is further illustrated.Following embodiment will be helpful to ability
The technical staff in domain further understands the present invention, but the invention is not limited in any way.It should be pointed out that this field
For those of ordinary skill, without departing from the inventive concept of the premise, various modifications and improvements can be made.These all belong to
In protection scope of the present invention.
A kind of learning rate Renewal step by step method based on estimation interpolation, the specific steps are as follows:
Step 1: in order to generate the Augmented Data frame for correcting learning rate, the method proposed uses estimation interpolation side
Method generates the sample of one group of carrying space time information.Preferably, by depending on Block- matching or Feature Points Matching, two phases are matched
Identical entity in adjacent frame simultaneously calculates motion vector field to complete.Two consecutive frames are divided into the square with length first, and
Pass through match block and record the movement between two frames, calculates MVF.For the block with center c and size pp, pass through minimum
Change the energy difference between the block from two frames to complete Block- matching: the form of energy difference depends on cost function, it is preferable that makes
With mean absolute difference (MAD) or mean square error (MSE).It, can be raw by each piece (δ c) of estimation by minimizing cost
Each piece of estimation is recorded as MVF u (c), MVF.In order to calculate interpolated frame, interpolation field is calculated first.From frame ft and
Sports ground ut between frame ft+1.δ t is enabled to indicate time interval, then the sports ground c generated on time t+λ δ t can
It is represented as: uλ(c+λut(c))=ut(c) it for each hole region in interpolated movements field, is sweared by closest movement
Amount filling.It is represented as by the λ interpolation results that frame ft and frame ft+1 interpolation generate: fλ(c)=(1-λ)ft(c-λuλ(c)+λ
ft+1(c+(1-λ)uλ(c));
Step 2: the input using Augmented Data frame, as Target Modeling task.According to given modeling target position p0, mention
Take out the model mdl of target0, it is preferable that object module is chosen using correlation filter;
Step 3: detecting according to object module to next frame, response R is calculated0, utilize response R0Generate a prediction bits
The object module Tmp set1;
Step 4: using calculated prediction model tmpt, with existing model mdltIt is compared, center is deviateed for model
A possibility that the case where domain, target is fast-moving target, is big, according to movement rate, increases learning rate α, for prediction model phase
Than declining obvious situation in original model absolute value, then target deformation or to encounter a possibility that blocking big, it is preferable that use view
Relevant range F, which is multiplied with prediction model in frequency domain, in frequency obtains response RFIt is rung if prediction model and assessment area calculate gained
Reduction should be worth, regionality is presented, then to block, according to serious shielding situation, learning rate α is accordingly decreased, if without region
Property, then it is deformation, according to deformation degree, correspondingly increases learning rate α;
Step 5: prediction model tmptNew training sample is provided for fresh target model;
Step 6: according to updated learning rate α, updating original object module according to the object module of predicted position, generating mesh
Mark model mdlt+1。
For t frame, second step is repeated to the 6th step, R is responded according to t frametWith the existing model mdl of t frametIt gives birth to frame by frame
At response RtWith object module mdlt+1。
The function that the present embodiment described device is realized is as follows:
(1) estimated using motion information, generate the intermediate video data comprising space-time consistency information, learnt for Renewal step by step
Rate.
(2) data of associative learning rate dynamic adjustment and augmentation, assess model degradation situation, progressive to learning rate
Formula updates, and realizes target dynamic modeling.
To sum up, using the target motion information in video data, the motion vector field of target is calculated in real time, to original video
Data carry out interpolation, generate intermediate state frame, for generating learning rate adjustable strategies, then carry out dynamic in real time to learning rate and adjust
It is whole, when target quickly changes, increase learning rate, increase model adaptability, when target is disturbed such as blocks, reduces study
Rate increases model robustness, mitigates the model degradation problem in target dynamic modeling work.
Specific embodiments of the present invention are described above.It is to be appreciated that the invention is not limited to above-mentioned
Particular implementation, those skilled in the art can make various deformations or amendments within the scope of the claims, this not shadow
Ring substantive content of the invention.
Claims (6)
1. a kind of learning rate Renewal step by step method based on estimation interpolation, it is characterised in that: specific step is as follows:
It A, the use of adjacent two frame of original video data (t frame, t+1 frame) is one group of (ft, ft+1), carry out estimation: movement
Estimation (ME) passes through the identical entity in two consecutive frames of matching and calculates motion vector field (MVF) to complete;For t frame ft
With t+1 frame ft+1Between motion vector field ut, enable δtIndicate time interval, then generated on time t+λ δ t
Sports ground c may be expressed as: uλ(c+λut(c))=ut(c);For each hole region in interpolated movements field, by most adjacent
Close motion vector filling;
It B, the use of adjacent two frame of original video data (t frame, t+1 frame) is one group of (ft, ft+1) and step A in calculate institute
The motion vector field u obtainedt, augmentation obtains the sample { f of one group of carrying space time informationλ, by t frame ftWith t+1 frame frame ft+1It inserts
The λ interpolation results that value generates are represented as: fλ(c)=(1-λ)ft(c-λuλ(c)+λft+1(c+(1-λ)uλ(c));
C, using the data after augmentation, input as Target Modeling task: according to given target response Rt, extract t
The existing model mdl of framet;
D, according to object module, next frame is detected, calculates response Rt, RtPeak value of response max (R) is as the sound predicted
Answer Rt+1;
E, the response R of prediction is utilizedt+1Generate t frame prediction model Tmpt;
F, t frame prediction model Tmp is utilizedtWith the existing model mdl of t frametThe degenerate case of model is assessed, dynamic
Selected learning rate α;
G, according to t frame prediction model Tmpt, and selected learning rate α, update the existing model mdl of t framet, generate t frame
Object module mdlt+1, indicate are as follows:
mdlt+1= αTmpt+(1-α)mdlt;
H, step C to step G is repeated, R is responded according to t frametWith the object module mdl of t framet+1, by the object module of t frame
mdlt+1Existing model mdl as t+1 framet+1Generate Target Modeling result ptWith the object module mdl of t+1 framet+2。
2. the learning rate Renewal step by step method according to claim 1 based on estimation interpolation, it is characterised in that: in step
Rapid A is into step B: original video data is with consecutive frame (t frame, t+1 frame) for one group of (ft, ft+1), it is used for two sides
Face, in a first aspect, in step, according to (ft, ft+1), calculate motion vector field ut;Second aspect, in stepb, according to
(ft, ft+1) and A in the u that calculatest, calculate Augmented Data frame { fλ}。
3. the learning rate Renewal step by step method according to claim 1 based on estimation interpolation, it is characterised in that: in step
In rapid B: using motion vector field to legacy data augmentation, not needing additionally to demarcate, the data of generation are also used as new sample defeated
Enter.
4. the learning rate Renewal step by step method according to claim 1 based on estimation interpolation, it is characterised in that: in step
In rapid B: Augmented Data frame { fλOriginal video data is contained in definition, in special circumstances, as λ=0, fλ= ft, when λ=1
When, fλ= ft+1。
5. the learning rate Renewal step by step method according to claim 1 based on estimation interpolation, it is characterised in that: described
Step C is into step the G: object module can be converted into existing model with time change, in the object module mdl of t framet+1
In t+1 frame, as existing model mdlt+1 。
6. a kind of dynamic modeling system for learning rate Renewal step by step method, it is characterised in that: estimate mould including sport interpolation
The target dynamic modeling module of block and learning rate Renewal step by step, in which:
Sport interpolation estimation module includes motion estimation module and data augmentation module;
The target dynamic modeling module of learning rate Renewal step by step includes that object module extraction module, response computation module, model move back
Change evaluation module and model modification module;
By the data that sport interpolation estimation module generates, the target dynamic modeling module of learning rate Renewal step by step is inputted;That is:
A, sport interpolation estimation module carries out estimation to video contiguous frames and generates movement according to the original video data of input
Vector field ut, and then generate Augmented Data frame { fλ, by ut, { fλInput learning rate Renewal step by step target dynamic modeling module;
The motion estimation module: being one group of (f from adjacent two frame (t frame, t+1 frame) of original video datat, ft+1), root
According to (ft, ft+1) calculate all pixels motion vector field ut, for hole region, it is filled, will be calculated using nearest neighbor algorithm
Obtained motion vector field utInput data augmentation module;
The data augmentation module: being one group of (f from adjacent two frame (t frame, t+1 frame) of original video datat, ft+1), with
And the motion vector field u of motion estimation module inputt, to (ft, ft+1) video data progress data augmentation, utilize t frame frame ft
With t+1 frame frame ft+1, interpolation goes out the smallest Augmented Data frame { f of errorλ, and by Augmented Data frame { fλIncoming learning rate is progressive
The target dynamic modeling module of update;
B, the target dynamic modeling module of the learning rate Renewal step by step, according to Augmented Data frame { fλ, in conjunction with the movement of contiguous frames
Information carries out learning rate dynamic and adjusts, and generates dynamic modeling result;Wherein:
The object module extraction module: as t=0, according to initially given response R0, calculate existing model mdl0As t > 0,
The existing model mdl for the t frame for directly update being taken to obtaint, input response computation module;
The response computation module: the existing model mdl of t frame is usedt, with t+1 frame ft+1In response RtNeighbouring region meter
Correlation is calculated, response R is obtainedt+1, and result R according to responset+1, calculate prediction model tmpt;
The model degradation evaluation module: calculated prediction model tmp is usedt, with existing model mdltIt is compared, for
A possibility that model deviates the case where central area, and target is fast-moving target is big, according to motion vector field utIt is moved
Rate increases learning rate α;For prediction model tmptCompared to existing model mdltAbsolute value declines obvious situation, then target shape
Become or to encounter a possibility that blocking big, uses Augmented Data frame { f at this timeλIn relevant range F and prediction model tmptIn frequency domain
It is multiplied and obtains response RF;If prediction model tmptWith assessment area F calculate gained response reduce present it is regional, then for
It blocks, according to serious shielding situation, learning rate α is accordingly decreased, if without regionality, for deformation, according to deformation degree,
Correspondingly increase learning rate α;Model degradation evaluation module is by α input model update module;
The model modification module: the learning rate α more new model obtained according to model degradation evaluation module, updated model
mdlt+1=(1-α)mdlt+αtmpt;
For original video data after input motion Interpolate estimation module, sport interpolation estimation module moves video contiguous frames
Estimation generates intermediate interpolation frame, inputs the target dynamic modeling module of learning rate Renewal step by step together, learning rate Renewal step by step
Video data after the augmentation that target dynamic modeling module is inputted according to target dynamic modeling module is believed in conjunction with the movement of contiguous frames
Breath carries out learning rate dynamic and adjusts, and generates dynamic modeling result, it may be assumed that
The model modification module: the learning rate α more new model obtained according to model degradation evaluation module, updated model are
mdlt+1=(1-α)mdlt+αtmpt。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910375212.0A CN110163892B (en) | 2019-05-07 | 2019-05-07 | Learning rate progressive updating method based on motion estimation interpolation and dynamic modeling system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910375212.0A CN110163892B (en) | 2019-05-07 | 2019-05-07 | Learning rate progressive updating method based on motion estimation interpolation and dynamic modeling system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110163892A true CN110163892A (en) | 2019-08-23 |
CN110163892B CN110163892B (en) | 2023-06-20 |
Family
ID=67633542
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910375212.0A Active CN110163892B (en) | 2019-05-07 | 2019-05-07 | Learning rate progressive updating method based on motion estimation interpolation and dynamic modeling system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110163892B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11889227B2 (en) | 2020-10-05 | 2024-01-30 | Samsung Electronics Co., Ltd. | Occlusion processing for frame rate conversion using deep learning |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060285590A1 (en) * | 2005-06-21 | 2006-12-21 | Docomo Communications Laboratories Usa, Inc. | Nonlinear, prediction filter for hybrid video compression |
CN102868879A (en) * | 2011-07-05 | 2013-01-09 | 北京大学 | Method and system for converting video frame rate |
CN103402098A (en) * | 2013-08-19 | 2013-11-20 | 武汉大学 | Video frame interpolation method based on image interpolation |
CN103888767A (en) * | 2014-03-31 | 2014-06-25 | 山东大学 | Frame rate improving method with UMH block matching motion estimation and optical flow field motion estimation combined |
EP2805306A1 (en) * | 2012-01-19 | 2014-11-26 | Thomson Licensing | Method and device for generating a motion field for a video sequence |
CN106056540A (en) * | 2016-07-08 | 2016-10-26 | 北京邮电大学 | Video time-space super-resolution reconstruction method based on robust optical flow and Zernike invariant moment |
CN107396111A (en) * | 2017-07-13 | 2017-11-24 | 河北中科恒运软件科技股份有限公司 | The compensation method of automatic video frequency interleave and system in mediation reality |
CN109151474A (en) * | 2018-08-23 | 2019-01-04 | 复旦大学 | A method of generating new video frame |
-
2019
- 2019-05-07 CN CN201910375212.0A patent/CN110163892B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060285590A1 (en) * | 2005-06-21 | 2006-12-21 | Docomo Communications Laboratories Usa, Inc. | Nonlinear, prediction filter for hybrid video compression |
CN102868879A (en) * | 2011-07-05 | 2013-01-09 | 北京大学 | Method and system for converting video frame rate |
EP2805306A1 (en) * | 2012-01-19 | 2014-11-26 | Thomson Licensing | Method and device for generating a motion field for a video sequence |
CN103402098A (en) * | 2013-08-19 | 2013-11-20 | 武汉大学 | Video frame interpolation method based on image interpolation |
CN103888767A (en) * | 2014-03-31 | 2014-06-25 | 山东大学 | Frame rate improving method with UMH block matching motion estimation and optical flow field motion estimation combined |
CN106056540A (en) * | 2016-07-08 | 2016-10-26 | 北京邮电大学 | Video time-space super-resolution reconstruction method based on robust optical flow and Zernike invariant moment |
CN107396111A (en) * | 2017-07-13 | 2017-11-24 | 河北中科恒运软件科技股份有限公司 | The compensation method of automatic video frequency interleave and system in mediation reality |
CN109151474A (en) * | 2018-08-23 | 2019-01-04 | 复旦大学 | A method of generating new video frame |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11889227B2 (en) | 2020-10-05 | 2024-01-30 | Samsung Electronics Co., Ltd. | Occlusion processing for frame rate conversion using deep learning |
Also Published As
Publication number | Publication date |
---|---|
CN110163892B (en) | 2023-06-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110191299B (en) | Multi-frame interpolation method based on convolutional neural network | |
CN102202164B (en) | Motion-estimation-based road video stabilization method | |
CN110163887B (en) | Video target tracking method based on combination of motion interpolation estimation and foreground segmentation | |
KR100985805B1 (en) | Apparatus and method for image stabilization using adaptive Kalman filter | |
CN115210716A (en) | System and method for multi-frame video frame interpolation | |
CN110796010A (en) | Video image stabilization method combining optical flow method and Kalman filtering | |
CN113011329B (en) | Multi-scale feature pyramid network-based and dense crowd counting method | |
CN104820997B (en) | A kind of method for tracking target based on piecemeal sparse expression Yu HSV Feature Fusion | |
CN112801043A (en) | Real-time video face key point detection method based on deep learning | |
CN102741884A (en) | Mobile body detection device and mobile body detection method | |
CN103139568A (en) | Video image stabilizing method based on sparseness and fidelity restraining | |
CN110210304A (en) | Method and system for target detection and tracking | |
CN111161309B (en) | Searching and positioning method for vehicle-mounted video dynamic target | |
CN102917220B (en) | Dynamic background video object extraction based on hexagon search and three-frame background alignment | |
CN108615241A (en) | A kind of quick estimation method of human posture based on light stream | |
Wang et al. | Video stabilization: A comprehensive survey | |
Kong et al. | Weakly supervised crowd-wise attention for robust crowd counting | |
CN110163892A (en) | Learning rate Renewal step by step method and dynamic modeling system based on estimation interpolation | |
CN112270691A (en) | Monocular video structure and motion prediction method based on dynamic filter network | |
Cai et al. | Leveraging intra-domain knowledge to strengthen cross-domain crowd counting | |
CN102970527B (en) | Video object extraction method based on hexagon search under five-frame-background aligned dynamic background | |
CN102917222B (en) | Mobile background video object extraction method based on self-adaptive hexagonal search and five-frame background alignment | |
CN112819849B (en) | Mark point-free visual motion capture method based on three eyes | |
CN110827324B (en) | Video target tracking method | |
Li et al. | Grid-based retargeting with transformation consistency smoothing |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |