CN111553935A - Human motion form obtaining method based on increment dimension reduction projection position optimization - Google Patents

Human motion form obtaining method based on increment dimension reduction projection position optimization Download PDF

Info

Publication number
CN111553935A
CN111553935A CN202010406264.2A CN202010406264A CN111553935A CN 111553935 A CN111553935 A CN 111553935A CN 202010406264 A CN202010406264 A CN 202010406264A CN 111553935 A CN111553935 A CN 111553935A
Authority
CN
China
Prior art keywords
dimensional data
dimensional
motion form
low
human motion
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
Application number
CN202010406264.2A
Other languages
Chinese (zh)
Other versions
CN111553935B (en
Inventor
李万益
谢海蓉
张谦
邬依林
徐海蛟
陈强
陈国明
张菲菲
陈勇昌
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guangdong second normal university
Original Assignee
Guangdong second normal university
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guangdong second normal university filed Critical Guangdong second normal university
Priority to CN202010406264.2A priority Critical patent/CN111553935B/en
Publication of CN111553935A publication Critical patent/CN111553935A/en
Application granted granted Critical
Publication of CN111553935B publication Critical patent/CN111553935B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • G06T7/251—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/20—Movements or behaviour, e.g. gesture recognition
    • G06V40/23—Recognition of whole body movements, e.g. for sport training
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Health & Medical Sciences (AREA)
  • Social Psychology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Psychiatry (AREA)
  • Evolutionary Computation (AREA)
  • Human Computer Interaction (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Graphics (AREA)
  • Geometry (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses a human motion form obtaining method based on increment dimension reduction projection position optimization, which comprises the following steps: (1) acquiring two different motion form image sequences, processing the first motion form image sequence and acquiring a high-dimensional data sample Y of the three-dimensional human motion form of the first motion form image sequenceIAnd obtaining low-dimensional data X after dimension reduction1And corresponding training parameters; (2) corresponding training parameters and low-dimensional data X1Training again to obtain updated low-dimensional data X1And corresponding training parameters, and obtaining a mapping relation f; (3) processing the second motion form image sequence to obtain a high-dimensional data sample Y of the three-dimensional human motion formIIObtaining a mapping relation g after training; the human motion form obtaining method based on the incremental dimension reduction projection position optimization has the characteristics of less time consumption, accurate estimation and high efficiency.

Description

Human motion form obtaining method based on increment dimension reduction projection position optimization
Technical Field
The invention relates to the field of three-dimensional human motion, in particular to a human motion form obtaining method based on incremental dimension reduction projection position optimization.
Background
Three-dimensional human motion morphology acquisition has been widely used in various fields such as medical diagnosis, animation, 3D game production, and the like. How to rapidly acquire the three-dimensional human motion form is the key for manufacturing corresponding multimedia image products. The three-dimensional human motion form is described by high-dimensional data, a series of human motion three-dimensional models represent motion postures, a plurality of motion postures form a human motion form sequence, and the human motion form sequence is a complete motion process which is also called gait.
The three-dimensional human motion morphology has been rapidly studied in China and is now the subject of intense research. There are various methods for obtaining the three-dimensional human motion profile. At present, after a two-dimensional image is preprocessed, a heuristic intelligent calculation method is used for obtaining the two-dimensional image, the method consumes long time, the obtained three-dimensional human motion form is easily influenced by the preprocessing quality, the obtaining accuracy rate is low, and the efficiency is low. Secondly, the method for acquiring the three-dimensional human motion form by using the dimension reduction model is a method with higher efficiency, but when the dimension reduction model learns a high-dimensional data sample, only low-dimensional visualization processing can be carried out on the data sample. Some improved dimension reduction models can generate new low-dimensional data samples from the reduced low-dimensional space, then generate corresponding new high-dimensional data samples through the generation of the new low-dimensional data samples and the mapping relation of the new low-dimensional data samples, and then obtain a new human motion three-dimensional model. The improved dimension reduction algorithms are only limited to high-dimensional data samples of the same type of motion morphology during learning and acquisition of the high-dimensional data samples. The above methods are all limited to the same type of motion profile acquisition. For the acquisition of different types of human motion forms, the improvement research on dimension reduction models in some documents is only limited to the fitting of model data samples, and low-dimensional space processing is also needed, so that the difficulty of the required three-dimensional human motion three-dimensional model is increased, the three-dimensional human motion three-dimensional model cannot be directly input to obtain a three-dimensional model of another motion form, and the method cannot well process how to acquire different types of human motion forms from the same type of motion forms.
Disclosure of Invention
The invention aims to provide a human motion form acquisition method based on incremental dimension reduction projection position optimization, which can generate a new motion type human motion three-dimensional model from a motion type human motion three-dimensional model and has the characteristics of less time consumption, accurate estimation and high efficiency.
The technical scheme adopted by the invention is as follows:
a human motion form obtaining method based on increment dimension reduction projection position optimization comprises the following steps:
(1) acquiring two different motion form image sequences, processing the first motion form image sequence and acquiring a high-dimensional data sample Y of the three-dimensional human motion form of the first motion form image sequenceIAnd obtaining low-dimensional data X after dimension reduction1And corresponding training parameters;
(2) corresponding training parameters obtained in the step (1) are addedStep (2) processed low-dimensional data X1Training again to obtain updated low-dimensional data X1And corresponding training parameters, and obtaining a mapping relation f;
(3) processing the second motion form image sequence to obtain a high-dimensional data sample Y of the three-dimensional human motion formIIAnd with the low-dimensional data X1Establishing a mapping relation, and obtaining a mapping relation g after training;
(4) high-dimensional data sample Y in three-dimensional human motion formIOr extracting motion shape samples y from real measurement dataI' obtaining a new second motion pattern sample y through the mapping relation f and the mapping relation gII’。
Compared with the prior art, the invention has the beneficial effects that:
the invention discloses a human motion form acquisition method based on increment dimension reduction projection position optimizationIAnd obtaining low-dimensional data X after dimension reduction1And corresponding training parameters; corresponding training parameters and processed low-dimensional data X1Training again to obtain updated low-dimensional data X1And corresponding training parameters, and obtaining a mapping relation f; processing the second motion form image sequence to obtain a high-dimensional data sample Y of the three-dimensional human motion formIIAnd with the low-dimensional data X1Establishing a mapping relation, and obtaining a mapping relation g after training; high-dimensional data sample Y in three-dimensional human motion formIOr extracting motion shape samples y from real measurement dataI' obtaining a new second motion pattern sample y by the mapping relation f and the mapping relation gII'. The human motion form acquisition method based on the increment dimension reduction projection position optimization can generate a human motion stereo model of another motion type from the human motion stereo model of one motion type, the acquisition accuracy and efficiency are high, the acquired three-dimensional human motion form has smooth visual effect, and errors are smoothLower.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the invention and not to limit the invention. In the drawings:
FIG. 1 is a schematic diagram of projection position optimization;
FIG. 2 is a diagram of the visual effects of three-dimensional human running motion morphology obtained using the IDRPPO algorithm and the IDRNPPO algorithm;
FIG. 3 is a diagram of the visual effect of a three-dimensional missing frame of the walking movement pattern of a human body obtained by using the IDRPPO algorithm and the IDRNPPO algorithm;
FIG. 4 is a diagram showing data in a corresponding low-dimensional space where a three-dimensional human walking motion pattern missing frame is obtained by using the IDRPPO algorithm and the IDRNPPO algorithm;
fig. 5 is a graph for comparing and analyzing acquisition errors of the IDRPPO algorithm and the IDRPPO algorithm.
Detailed Description
The technical solution of the present invention will be described in further detail with reference to the following embodiments, but the present invention is not limited thereto.
The invention discloses a human motion form obtaining method based on increment dimension reduction projection position optimization, which comprises the following steps of:
(1) acquiring two different motion form image sequences, processing the first motion form image sequence and acquiring a high-dimensional data sample Y of the three-dimensional human motion form of the first motion form image sequenceIAnd obtaining low-dimensional data X after dimension reduction1And corresponding training parameters.
Wherein, the high-dimensional data sample Y of the three-dimensional human motion formIThe calculation formula for performing the dimension reduction processing and obtaining the corresponding training parameters is as follows:
Figure BDA0002491398650000031
wherein the content of the first and second substances,
Figure BDA0002491398650000032
Figure BDA0002491398650000033
y is a high-dimensional data sequence, and Y is [ Y ═ Y%1,...,yi,...,yN]T∈RN×D，yi∈RDX is a low-dimensional data sequence, X ═ X1,...,xi,...,xN]T∈RN×q，xi∈Rq,KYIn the form of a kernel matrix, the kernel matrix,
Figure BDA0002491398650000034
having a nuclear parameter of
Figure BDA0002491398650000035
KXIs a kernel matrix, KX∈R(N-1)×(N-1),
Figure BDA0002491398650000036
Having a nuclear parameter of
Figure BDA0002491398650000037
W is a scale parameter, and W is a scale parameter,
Figure BDA0002491398650000038
X2:N＝[x2,x3,..,xN]T，x1obeying to a q-dimensional gaussian distribution,
Figure BDA0002491398650000039
satisfy the requirement of
Figure BDA00024913986500000310
Figure BDA00024913986500000311
Satisfy the requirement of
Figure BDA00024913986500000312
Also comprises the step (1.0) of judging the high-dimensional data sample Y of the three-dimensional human motion formIIf the missing frame exists, using the low-dimensional data X obtained in the step (1) of projection position optimization processing1。
Wherein, the low dimensional data X obtained in the step (1)1The calculation formula for optimizing the projection position is as follows:
Figure BDA00024913986500000313
Figure BDA0002491398650000041
wherein c is a preset parameter of the distance from the point of the missing frame to the projection point.
The projection position optimization is carried out after dimension reduction is needed for the study of the human motion form of incomplete gait, and
Figure BDA0002491398650000042
is in a vector
Figure BDA00024913986500000410
Projection operation, a is the known low-dimensional data preceding the first frame of the missing motion shape (missing frame), B is the known low-dimensional data of the last frame of the missing motion shape (missing frame), Ci,i＝1,2,...NmissIs the low dimensional data of the missing frame. According to fig. 1, i have:
Figure BDA0002491398650000044
Figure BDA0002491398650000045
after dimension reduction, c in the formula (7) is a preset parameter of the distance from the point of the missing frame to the projection point. The position of the obtained missing frame should satisfy the positional relationship of the expressions (6) and (7). Therefore, the model can be trained only when the secondary parameter training is carried out, namely the formula (1). An optimization objective function and a gradient function thereof can be obtained according to the formula (6) and the formula (7) as shown in the formula (4);
in formula (5), there are
Figure BDA0002491398650000046
The product operation representing the corresponding position element in the matrix. The solution of the formula (4) is not unique, but the position of the missing frame in the training process is guaranteed by any solution, so that the required position of the missing low-dimensional data sample can be obtained by performing the secondary training formula. Equation (4) can be solved using some common gradient optimization algorithms.
(2) Corresponding training parameters obtained in the step (1) and the processed low-dimensional data X obtained in the step (2)1Training again to obtain updated low-dimensional data X1And corresponding training parameters, and obtaining a mapping relation f.
Corresponding training parameters obtained in the step (1) and the processed low-dimensional data X obtained in the step (2)1The calculation formula for the training again is as follows:
Figure BDA0002491398650000047
due to the fact that
Figure BDA0002491398650000048
y∈RD,x∈RqThe mapping relationship from the high-dimensional data space to the low-dimensional data space can be established by the following formula, that is, the calculation formula of the mapping relationship f is:
Figure BDA0002491398650000049
if more than 2 low-dimensional data spaces to high-dimensional data spaces need to be established, we can train equation (1) more, but start with the second mapping and fix the low-dimensional data obtained by the first mapping training equation.
(3) Processing the second motion pattern image sequence to obtain the three-dimensional human motion patternHigh dimensional data sample YIIAnd with the low-dimensional data X1And establishing a mapping relation, and obtaining a mapping relation g after training.
Wherein, the high-dimensional data sample Y of the three-dimensional human motion formIIAnd low-dimensional data X1The calculation formula for establishing the mapping relation is as follows:
X＝ΦWD(9)
wherein, phi ∈ RN×NkIn order to be a radial basis function,
Figure BDA0002491398650000051
WD∈RNk×qis a weight matrix, and Nk is less than or equal to N.
Figure BDA0002491398650000052
Estimate W for least squaresD，
Figure BDA0002491398650000053
Then, y*∈RDRepresenting new high-dimensional data samples, x*∈RDRepresenting the corresponding low dimensional data. If b is known, from y*To x*The mapping relationship of (a) can be established by the following formula:
the calculation formula of the mapping relation g is as follows:
Figure BDA0002491398650000054
wherein, phi (y)*)＝[φ(y*,c1),φ(y*,c2),...,φ(y*,cNk)]. Then, we can get:
Figure BDA0002491398650000055
in formula (11), e ∈ RN×NkError matrix, order
Figure BDA0002491398650000056
Then, there are
Figure BDA0002491398650000057
Figure BDA0002491398650000058
Can be decomposed to make
Figure BDA0002491398650000059
Is a diagonal matrix
Figure BDA00024913986500000510
Is a reversible matrix (
Figure BDA00024913986500000511
Figure BDA00024913986500000512
Then, the user can use the device to perform the operation,
Figure BDA00024913986500000513
order to
Figure BDA00024913986500000514
We can derive:
Figure BDA00024913986500000515
therefore, equation (12) can be written as:
Figure BDA00024913986500000516
according to the nature of the least-squares method,
Figure BDA00024913986500000517
we have:
Figure BDA00024913986500000518
then, can obtain
Figure BDA00024913986500000519
When training, Nk orthogonal vectors can be replaced, and we can then get:
Figure BDA00024913986500000520
formula (15) is equivalent to:
Figure BDA0002491398650000061
a compound of the formula (16),
Figure RE-GDA0002534724230000062
and
Figure RE-GDA0002534724230000063
are all sets of orthogonal vectors, SwIs subset S'w，S′wIs that
Figure RE-GDA0002534724230000064
The set of vectors is then used to generate a set of vectors,
Figure RE-GDA0002534724230000065
for a set of orthogonal vectors
Figure RE-GDA0002534724230000066
Elements, however
Figure RE-GDA0002534724230000067
When it is satisfied with
Figure RE-GDA0002534724230000068
And is1Training is done for a sufficiently small positive number, which is equivalent to selecting as few vectors as possible
Figure RE-GDA0002534724230000069
I.e., minimizing the value of Nk to complete the training of the incremental mapping relationship.
(4) High-dimensional data sample Y in three-dimensional human motion formIOr is realExtracting motion form samples y from measurement dataI' obtaining a new second motion pattern sample y through the mapping relation f and the mapping relation gII’。
The human motion form acquisition method based on the increment dimension reduction projection position optimization can generate a human motion stereo model of another motion type from the human motion stereo model of one motion type, the acquisition accuracy and efficiency are high, the acquired three-dimensional human motion form has smooth visual effect and low error.
Test examples
Since most of heuristic intelligent algorithms and dimension reduction models cannot realize the acquisition of one type of motion form from another type of motion form, one of the keys of the human motion form acquisition method based on the incremental dimension reduction projection position optimization for acquiring the motion form is the projection position optimization characteristic. Therefore, a comparison is made using the unused Projection Position Optimization algorithm and using the Projection Position Optimization (IDRPPO) method. An algorithm that does not use Projection Position Optimization, referred to herein as an Incremental Dimension Reduction and non-Projection Position Optimization algorithm (IDRNPPO), i.e., IDRNPPO algorithm. In the simulation test, the obtained missing frame of the first type of three-dimensional human motion form (walking of the human body) and the obtained visual effect and error of the second type of three-dimensional human motion form (running of the human body) are respectively compared.
A first type of three-dimensional human body walking motion form image sequence is obtained, wherein missing frames exist in the three-dimensional human body walking motion form image sequence, and a high-dimensional data sample of the three-dimensional human body walking motion form image sequence is obtained after the three-dimensional human body walking motion form image sequence is processed, as shown in fig. 2 (a).
1. Visual effect comparison
The high-dimensional data samples of the three-dimensional human walking movement form are respectively input into an IDRPPO algorithm and an IDRNPPO algorithm, the high-dimensional data samples of the three-dimensional running human movement form are obtained, the high-dimensional data samples of the three-dimensional running human movement form output from the IDRPPO algorithm are shown in fig. 2(b), and the high-dimensional data samples of the three-dimensional running human movement form output from the IDRNPPO algorithm are shown in fig. 2 (c). Fig. 3(a) shows high-dimensional data samples of the three-dimensional body walking motion pattern missing frame output from the IDRPPO algorithm, and fig. 3 (b) shows high-dimensional data samples of the three-dimensional body walking motion pattern missing frame output from the IDRPPO algorithm. Fig. 4(a) shows low-dimensional data of a three-dimensional body walking motion pattern missing frame output from the IDRPPO algorithm, and fig. 4(b) shows low-dimensional data of a three-dimensional body walking motion pattern missing frame output from the IDRPPO algorithm.
As can be seen from fig. 2, the visual effect of the three-dimensional human running exercise form obtained by the IDRPPO algorithm is better than that of the IDRPPO algorithm, the 30 th, 35 th, 40 th, 45 th, 48 th, 52 th and 58 th frames are almost repeated by the IDRPPO algorithm, and the corresponding frames have less visual fluency than the IDRPPO algorithm, and cannot embody the running posture which the human body should have during running. Moreover, the original input human walking motion form has missing frames, and the missing frames need to be acquired first. Two missing frame acquisition cases for the input data samples can be seen in fig. 3. In fig. 3, the missing frame obtained by the IDRPPO algorithm shows a smooth three-dimensional human body movement walking shape, which is a visually smooth walking process. On the contrary, the three-dimensional human body walking motion form displayed by the IDRNPPO algorithm through acquiring the missing frame is single, and the walking process with smooth vision is not displayed. Fig. 4 shows the low-dimensional data display of the three-dimensional human walking movement morphology learned by the IDRPPO algorithm and the IDRNPPO algorithm. The IDRPPO algorithm uses projection position optimization, the processed low-dimensional data missing frame part can be well fused into a non-missing frame to form a smooth manifold curve, and the IDRNPPO algorithm does not optimize the low-dimensional data of the missing frame, so that the low-dimensional data of the missing frame is randomly displayed and even overlapped with the low-dimensional data of other frames. Fig. 4 also illustrates the reason why the IDRPPO algorithm gets the missing frames visually smoother than the IDRPPO algorithm from another perspective. As can be seen from fig. 2, fig. 3 and fig. 4, the IDRPPO algorithm has better acquisition performance than the idrnpo algorithm.
2. Obtaining an error comparison
Next, the IDRPPO algorithm and the IDRNPPO algorithm are compared with each other from the error of obtaining the three-dimensional human body running motion form and the error of obtaining the three-dimensional human body walking motion form missing frame, wherein the error analysis of obtaining the three-dimensional human body running motion form is shown in fig. 5(a), and the error analysis of obtaining the three-dimensional human body walking motion form missing frame is shown in fig. 5 (b).
As can be seen from fig. 4, no matter the missing frames (8 frames) of the three-dimensional body running motion form or the three-dimensional body walking motion form are obtained, the error of each frame is the minimum of the idrpp algorithm overall. It is normal that some frames in the two methods shown in fig. 5(a) have closer acquisition errors, because some frames in the IDRNPPO algorithm are also part of the process that can normally display the running movement form of the human body, but the total error is measured by the average error. It is found from fig. 3 that the average error of the running exercise form and the walking exercise form is also the lowest in the idrpp algorithm. The results of fig. 5 again demonstrate that the performance of the IDRPPO algorithm for obtaining three-dimensional human motion morphology is better than that of the IDRNPPO algorithm.
Therefore, the test results show that the human motion form acquisition method based on the incremental dimension reduction projection position optimization can learn the three-dimensional human motion form of incomplete gait and then better acquire another three-dimensional human motion form of different types. The defects of some current unsupervised learning algorithms are made up.
The above description is only exemplary of the invention, and any modification, equivalent replacement, and improvement made within the spirit and scope of the present invention should be considered within the scope of the present invention.

Claims (6)

1. A human motion form obtaining method based on increment dimension reduction projection position optimization is characterized by comprising the following steps:
(1) acquiring two different motion form image sequences, processing the first motion form image sequence and acquiring a high-dimensional data sample Y of the three-dimensional human motion form of the first motion form image sequenceIAnd obtaining low-dimensional data X after dimension reduction1And corresponding training parameters;
(2) corresponding training parameters obtained in the step (1) and the processed low-dimensional data X obtained in the step (2)1Training again to obtain updated low-dimensional data X1And corresponding training parameters, and obtaining a mapping relation f;
(3) processing the second motion form image sequence to obtain a high-dimensional data sample Y of the three-dimensional human motion formIIAnd with the low-dimensional data X1Establishing a mapping relation, and obtaining a mapping relation g after training;
(4) high-dimensional data sample Y in three-dimensional human motion formIOr extracting motion shape samples y from real measurement dataI' obtaining a new second motion pattern sample y through the mapping relation f and the mapping relation gII’。
2. The method for acquiring the human motion form based on the position optimization of the incremental dimension reduction projection according to claim 1, wherein in the step (1), the high-dimensional data sample Y of the three-dimensional human motion form is obtainedIThe calculation formula for performing the dimension reduction processing and obtaining the corresponding training parameters is as follows:
Figure FDA0002491398640000011
wherein the content of the first and second substances,
Figure FDA0002491398640000012
Figure FDA0002491398640000013
y is a high-dimensional data sequence, and Y is [ Y ═ Y%1,...,yi,...,yN]T∈RN×D，yi∈RDX is a low-dimensional data sequence, X ═ X1,...,xi,...,xN]T∈RN×q，xi∈Rq,KYIs a kernel matrix, KY∈RN×N，
Figure FDA0002491398640000014
Having a nuclear parameter of
Figure FDA0002491398640000015
KXIs a kernel matrix, KX∈R(N-1)×(N-1),
Figure FDA0002491398640000016
Having a nuclear parameter of
Figure FDA0002491398640000017
W is a scale parameter, and W is a scale parameter,
Figure FDA0002491398640000018
wm＞0，κ＝10-3，X2:N＝[x2,x3,..,xN]T，x1obeying to a q-dimensional gaussian distribution,
Figure FDA0002491398640000019
satisfy the requirement of
Figure FDA00024913986400000110
Figure FDA00024913986400000111
Satisfy the requirement of
Figure FDA00024913986400000112
3. The method for acquiring the human motion form based on the increment dimension reduction projection position optimization according to claim 1, characterized by further comprising the step (1.0) of judging the high-dimensional data sample Y of the three-dimensional human motion form after the step (1)IWhether there is a missing frame or not,if the data exists, the low-dimensional data X obtained in the step (1) is processed by using the projection position optimization1。
4. The method for acquiring the human motion morphology based on the position optimization of the incremental dimension reduction projection according to claim 3, wherein in the step (1.0), the low-dimensional data X obtained in the step (1) is subjected to1The calculation formula for optimizing the projection position is as follows:
Figure FDA0002491398640000021
Figure FDA0002491398640000022
wherein c is a preset parameter of the distance from the point of the missing frame to the projection point,
Figure FDA0002491398640000023
Figure FDA0002491398640000024
Figure FDA0002491398640000025
Figure FDA0002491398640000026
is in a vector
Figure FDA0002491398640000027
Projection operation, a is the known low-dimensional data before the first frame of the missing motion profile (missing frame), B is the known low-dimensional data of the last frame of the missing motion profile (missing frame), Ci,i＝1,2,...NmissLow dimensional data for missing frames;
the product operation representing the corresponding position element in the matrix.
5. The method for acquiring the human motion morphology based on the position optimization of the incremental dimension reduction projection as claimed in claim 2, wherein in the step (2), the corresponding training parameters obtained in the step (1) and the processed low-dimensional data X obtained in the step (2) are used1The formula of the calculation for the retraining is:
Figure FDA0002491398640000028
the calculation formula of the mapping relation f is as follows:
Figure FDA0002491398640000029
6. the method for acquiring the human motion form based on the position optimization of the incremental dimension reduction projection according to claim 2, wherein in the step (3), the high-dimensional data sample Y of the three-dimensional human motion form is obtainedIIAnd low-dimensional data X1The calculation formula for establishing the mapping relation is as follows:
X＝ΦWD；
wherein, phi ∈ RN×NkIn order to be a radial basis function,
Figure FDA0002491398640000031
WD∈RNk×qis a weight matrix, and Nk is less than or equal to N.
Figure FDA0002491398640000032
Estimate W for least squaresD，
Figure FDA0002491398640000033
y*∈RDRepresenting new high-dimensional data samples, x*∈RDRepresenting the corresponding low dimensional data;
the calculation formula of the mapping relation g is as follows:
Figure FDA0002491398640000034
wherein, phi (y)*)＝[φ(y*,c1),φ(y*,c2),...,φ(y*,cNk)]。
CN202010406264.2A 2020-05-14 2020-05-14 Human motion form obtaining method based on increment dimension reduction projection position optimization Active CN111553935B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010406264.2A CN111553935B (en) 2020-05-14 2020-05-14 Human motion form obtaining method based on increment dimension reduction projection position optimization

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010406264.2A CN111553935B (en) 2020-05-14 2020-05-14 Human motion form obtaining method based on increment dimension reduction projection position optimization

Publications (2)

Publication Number Publication Date
CN111553935A true CN111553935A (en) 2020-08-18
CN111553935B CN111553935B (en) 2020-12-15

Family

ID=72004729

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010406264.2A Active CN111553935B (en) 2020-05-14 2020-05-14 Human motion form obtaining method based on increment dimension reduction projection position optimization

Country Status (1)

Country Link
CN (1) CN111553935B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104952105A (en) * 2014-03-27 2015-09-30 联想(北京)有限公司 Method and apparatus for estimating three-dimensional human body posture
CN108596027A (en) * 2018-03-18 2018-09-28 西安电子科技大学 The detection method of unknown sorting signal based on supervised learning disaggregated model
CN109635754A (en) * 2018-12-17 2019-04-16 四川师范大学 Gait feature fusion method based on Non-linear coupling metric learning
CN110415270A (en) * 2019-06-17 2019-11-05 广东第二师范学院 A kind of human motion form evaluation method based on double study mapping increment dimensionality reduction models
CN110648276A (en) * 2019-09-25 2020-01-03 重庆大学 High-dimensional image data dimension reduction method based on manifold mapping and dictionary learning

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104952105A (en) * 2014-03-27 2015-09-30 联想(北京)有限公司 Method and apparatus for estimating three-dimensional human body posture
CN108596027A (en) * 2018-03-18 2018-09-28 西安电子科技大学 The detection method of unknown sorting signal based on supervised learning disaggregated model
CN109635754A (en) * 2018-12-17 2019-04-16 四川师范大学 Gait feature fusion method based on Non-linear coupling metric learning
CN110415270A (en) * 2019-06-17 2019-11-05 广东第二师范学院 A kind of human motion form evaluation method based on double study mapping increment dimensionality reduction models
CN110648276A (en) * 2019-09-25 2020-01-03 重庆大学 High-dimensional image data dimension reduction method based on manifold mapping and dictionary learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
WANYI LI等: "Projection Analysis Optimization for Human Transition Motion Estimation", 《INTERNATIONAL JOURNAL OF DIGITAL MULTIMEDIA BROADCASTING》 *
李万益等: "基于双学习映射增量降维模型的人体运动形态估计", 《科学技术与工程》 *

Also Published As

Publication number Publication date
CN111553935B (en) 2020-12-15

Similar Documents

Publication Publication Date Title
Chin et al. Incremental kernel principal component analysis
Oliva et al. Cross entropy based thresholding for magnetic resonance brain images using Crow Search Algorithm
CN107229914B (en) Handwritten digit recognition method based on deep Q learning strategy
Bolkart et al. A groupwise multilinear correspondence optimization for 3d faces
CN108154499B (en) Woven fabric texture flaw detection method based on K-SVD learning dictionary
CN108121950B (en) Large-pose face alignment method and system based on 3D model
CN105426929B (en) Object shapes alignment device, object handles devices and methods therefor
CN114266939B (en) Brain extraction method based on ResTLU-Net model
CN113658721B (en) Alzheimer disease progress prediction method
Krebs et al. Probabilistic motion modeling from medical image sequences: application to cardiac cine-MRI
Bhattacharjee et al. Single index Fréchet regression
CN114170088A (en) Relational reinforcement learning system and method based on graph structure data
CN110543916A (en) Method and system for classifying missing multi-view data
Guo et al. Occluded joints recovery in 3d human pose estimation based on distance matrix
CN104866905B (en) A kind of learning method of the sparse tensor dictionary of nonparametric based on beta processes
O'Reilly et al. Pre-trained vs. random weights for calculating fréchet inception distance in medical imaging
CN111553250B (en) Accurate facial paralysis degree evaluation method and device based on face characteristic points
CN111553935B (en) Human motion form obtaining method based on increment dimension reduction projection position optimization
CN110415270B (en) Human motion form estimation method based on double-learning mapping incremental dimension reduction model
CN111553249B (en) H-B grading-based accurate facial paralysis degree evaluation method and device under CV
CN115527097A (en) CNN interpretation graph object correlation analysis method based on ablation analysis
CN115481256A (en) Inverse relation rotation embedding knowledge representation method and system based on convolution quaternion
Zhang et al. Latent multi-relation reasoning for gan-prior based image super-resolution
Olieman et al. Fitness-based linkage learning in the real-valued gene-pool optimal mixing evolutionary algorithm
Bannister et al. Detecting 3D syndromic faces as outliers using unsupervised normalizing flow models

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