CN106778639A - A kind of exercise data search method based on the description of attitude relative space-time characteristic statisticses - Google Patents
A kind of exercise data search method based on the description of attitude relative space-time characteristic statisticses Download PDFInfo
- Publication number
- CN106778639A CN106778639A CN201611204708.4A CN201611204708A CN106778639A CN 106778639 A CN106778639 A CN 106778639A CN 201611204708 A CN201611204708 A CN 201611204708A CN 106778639 A CN106778639 A CN 106778639A
- Authority
- CN
- China
- Prior art keywords
- action
- feature
- retrieval
- relative space
- user
- 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
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/42—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
Abstract
The invention discloses a kind of exercise data search method based on the description of attitude relative space-time characteristic statisticses, point following lower step:3 d pose relative space-time feature extraction, extract 3 d pose in joint formed point, line, surface geometric element between relative tertiary location and its change measurement as attitude character representation;Action statistics description generation, the statistics description for extracting attitude relative space-time feature represents that the statistic of selection includes as the characteristic vector of action, average, extreme difference, variance and skewness totally 4 statistical variables;Similarity mode:The similarity degree of action and retrieval action in storehouse is calculated using Euclidean distance, candidate result is acted after being sorted from high to low to final similarity is returned;Relevant feedback, the feedback by linear SVM (SVM Support Vector Machine) for user sets up disaggregated model, so as to optimize character subset and its weight combination so that the performance of retrieval is optimal.
Description
Technical field
The present invention relates to human body movement data retrieval method, belong to computer three-dimensional animation technology and multimedia-data procession
Field, is a kind of human body movement data retrieval method based on the description of attitude relative space-time characteristic statisticses specifically.
Background technology
Due to various applications in the urgent need to and business capture device be widely popularized, occurred in that at present increasingly
Human action storehouse (the http in many large-scale three dimensional human action storehouses, such as Carnegie Mellon University of the U.S.://
Mocap.cs.cmu.edu) etc..Human body movement data required for how obtaining user from human action storehouse has become dynamic
Make the data effectively key issue that utilizes, the retrieval technique of Case-based Reasoning using character representation motion content, by feature it
Between measuring similarity realize the matching of content similar movement, preferably overcome traditional based on text marking search method
Deficiency, has become the study hotspot of motion capture data searching field.Effective character representation and corresponding similarity mode
Mechanism is the basic and key issue of the retrieval of Case-based Reasoning.
Feature will carry out complete and effective expression to the content moved.In existing posture feature is represented, close
The three-dimensional coordinate of section and its coded representation, quaternary number and Eulerian angles and its coded representation are conventional method for expressing, but research
Show that they are only applicable to the similar exercise data retrieval of numerical value, can not effectively represent due to style, duration equal difference
Abnormity into the similar motion of logic.M ü ller et al. are in document M ü ller M, et al.Efficient content based
retrieval of motion capture data.ACM Transactions on Graphics,2005,24(3):677-
685) boolean's geometric properties motion retrieval similar to realize logic is proposed in, but limited feature quantity and two-value boolean's shape
State causes that it is poor for the action separating capacity of subdivision.In document Pan Hong, the such as Xiao Jun, Wu Fei is based on key frame to Pan Hong et al.
3 d human motion retrieval [J], CAD and graphics journal, 2009,21 (2):Proposed in 214-222 [77]
Using the angle between appendicular skeleton and middle cardiac skeleton as posture feature, but there is the unrelated part of numerous and middle cardiac skeleton
Action.Tang et al. is in document Tang J T, et al.Retrieval of logically relevant 3D human
motions by Adaptive Feature Selection with Graded Relevance Feedback[J]
.Pattern Recognition Letters,2012,33(4):Point out that relative position relation is that extraction logic contains in 420-430
Effective expression of justice, and use joint to the distance between as attitude character representation, but it have ignored straight line and plane
The two important geometric elements, additionally, this feature can not directly express joint rotation content.Chen et al. is in document Chen
C,et al.Perceptual 3D pose distance estimation by boosting relational
geometric features[J].Computer Animation and Virtual Worlds,2009,20(2-3):267-
Defined in 277 attitude with respect to geometric properties collection to evaluate the similitude of perceptually attitude, but different joints formed point,
Line, any combination in face so that feature quantity has reached 560,000 multidimensional, it is necessary to be subject to proper restraint to reduce for concrete application
Intrinsic dimensionality.
Attitude is the Component units of motion, some researchers based on the character representation of attitude, by dimensionality reduction and poly-
Class method is represented obtaining the low-dimensional of attitude;By Singular Value Decomposition Using, the humorous conversion of ball, tensor algebra Subspace Decomposition and
Keyframe techniques are obtained to be redescribed on higher level to motion segments.But in the secondary of dimensionality reduction, cluster and action
During feature extraction distance is calculated, reflect that the attitude local feature of motion component is assigned with specific weight, these local features
Weight can not during retrieval with the difference of retrieval example enter Mobile state adjust, result in features described above and represent
Can not effectively support that the local space of motion is retrieved.
In sum, existing motion component character representation still can not effectively support that exercise data is retrieved, the present invention
A kind of exercise data case retrieval methods based on the description of attitude relative space-time characteristic statisticses are proposed, it is relatively special in attitude space-time
Levy on the basis of content representation, quadratic character of the method further using statistics description extraction action is vectorial, in the mistake of retrieval
Cheng Zhong, the method is also used based on SVM relevant feedbacks to improve retrieval performance.It is text with the inventive method method relatively
Offer (the case retrieval methods CADs of the national dance exercise datas such as Chen Songle and graphics journal, 2014,26
(2):198-210) with document (Songle Chen, et al.Relevance Feedback for Human Motion
Retrieval Using a Boosting Approach.Multimedia tools and applications,2016,75
(2):The method proposed in 787-817), but both approaches are directly using attitude relative space-time feature as character representation,
And using dynamic time warping algorithm as the similarity calculating method for acting, computation complexity is too high, it is difficult to meet extensive
The performance requirement of motion capture data library searching.
The content of the invention
Goal of the invention:The technical problems to be solved by the invention are directed to the deficiencies in the prior art, it is proposed that one kind is based on
The exercise data search method of attitude relative space-time characteristic statisticses description.
Technical scheme:A kind of exercise data retrieval side based on the description of attitude relative space-time characteristic statisticses disclosed by the invention
Method, comprises the following steps:
Step 1,3 d pose relative space-time feature extraction:The point, line, surface geometric element collection that joint is formed in 3 d pose
Conjunction is the minimum Component units of the corresponding regional area of different patterns.Joint is formed in present invention extraction 3 d pose
The measurement of relative tertiary location and its change between point, line, surface geometric element as attitude character representation, by different offices
The weight of the different type feature that portion region includes combines to express extensive gesture mode.Included for everything in storehouse
Each attitude, its relative space-time feature extraction comprises the following steps that:
1) 3 D human body joint model is defined, several joints of selection most important of which represent as 3 d pose;
2) geometric element set is built, the joint of selection constitutes the point set in geometric element set, and point concentrates any at 2 points
Straight line is formed, then constitute plane at any 3 points;
3) extract each 3 d pose space feature, including joint adjust the distance feature, joint and bone distance feature,
Joint is with plan range feature, bone to angle feature, bone and plane included angle feature, plane and plane included angle feature, joint
Hyperspin feature;
4) each 3 d pose relative time feature, including joint angle velocity and acceleration feature are extracted.
Step 2, action statistics description generation:Action is formed by several attitude consecutive variations, and the present invention extracts attitude phase
To space-time characteristic statistics description as action characteristic vector represent that the statistic of selection includes, average, extreme difference, variance with
And skewness totally 4 statistical variables.For each action in storehouse, what action statistics description was generated comprises the following steps that:
1) the relative space-time feature of each 3 d pose included in loading action, and using each feature as one-dimensional random
Variable, each attitude in action has corresponded to the value of the stochastic variable;
2) average of each relative space-time feature is calculated;
3) extreme difference of each relative space-time feature is calculated;
4) variance of each relative space-time feature is calculated;
5) skewness of each relative space-time feature is calculated;
6) will in order be arranged simultaneously after the average of each relative space-time feature extraction, extreme difference, variance, skewness normalization
Preserve, the characteristic vector for forming each action is represented.
Step 3, similarity mode:The feature for extracting the retrieval action example that user submits to using step 1 and step 2 is retouched
State, and and then using Euclidean distance calculate storehouse in action and retrieval action similarity degree, final similarity is arranged from high to low
Result action is returned after sequence.Similarity mode process is comprised the following steps that:
1) user submits to retrieval to act example;
2) the relative space-time characteristic statisticses description of the retrieval action that user submits to is calculated using step 1 and step 2;
3) for each action in storehouse, based on the description of relative space-time characteristic statisticses, calculated using Euclidean distance and user
The distance of the retrieval action of submission, the weight of each feature is identical;
4) distance of the retrieval action submitted to each action and user in storehouse is ranked up;
5) Top-20 minimum action of distance is returned into user.
Step 4, relevant feedback:Each feature employs identical weight in the Similarity Measure of step 3.It is actually right
A specific character subset and its weight combination are all there is in the retrieval each time of user so that the performance of retrieval reaches most
It is excellent.In order to the retrieval for preferably catching user is intended to and the optimal character subset of Step wise approximation and its weight combination, the present invention is logical
Cross SVMs (SVM-Support Vector Machine) related feedback method to optimize retrieval result, based on SVM's
Relevant feedback is comprised the following steps that:
1) for step 3 or the result of last round of relevant feedback, user annotation returns to example and the retrieval of its submission is moved
Whether related make;
2) submit user annotation to the related sample of retrieval action as positive example sample to it, by user annotation and its
Retrieval is submitted to act incoherent sample as negative data;
3) linear kernel is selected, negative data is aligned using SVM and is learnt, obtain svm classifier model;
4) each action in storehouse is input to svm classifier model successively, calculates score;
5) score is ranked up, Top-20 action of highest scoring is returned into user;
If 6) user is satisfied with to result, retrieving terminates, otherwise return to step 1), carry out next round feedback.
Beneficial effect
1) present invention is described using attitude space-time relative characteristic as 3 d pose, and joint in attitude is extracted in this feature description
The measurement of relative tertiary location and its change between the point, line, surface geometric element of formation is represented as movement content.Knuckle shaped
Into point, line, surface geometric element set be the corresponding regional area of different patterns minimum Component units, and point, line, surface
Angle between geometric element and the relative tertiary location between minimum Component units is reflected from different aspect apart from isometry
Relation, the weight combination of the different type feature included by different regional areas can express extensive pattern.
2) present invention is described as the character representation for acting, average, pole using the statistics of 3 d pose space-time relative characteristic
Difference, variance, skewness statistical variable reflect benchmark and situation of change of each feature in action, being capable of effective expression action fortune
Dynamic content.Meanwhile, the action unification of time upper different length is converted to the characteristic vector of uniform length for statistics description, therefore can
To carry out Similarity Measure and relevant feedback using methods such as Euclidean distance and Linear SVMs, calculating effect is drastically increased
Rate and the validity of retrieval.
3) present invention uses Linear SVM to carry out relevant feedback and retrieves performance with further raising, and SVM has small sample in itself
The good advantage of study generalization ability, therefore general study side caused by the sample size due to user annotation can be overcome the shortcomings of
The problem of method generalization ability difference.And allow to explain right according to the weight of each feature in disaggregated model using linear kernel function
In the different important procedures that this retrieves each feature, the interpretation of retrieval result is improve.
Brief description of the drawings
Fig. 1 is flow chart of the method for the present invention.
Fig. 2 is the human synovial model schematic used in the present invention.
Fig. 3 is the relative space-time feature schematic diagram of present invention definition.
Fig. 4 is the retrieval result to retrieval input action " jump " that embodiment is obtained by similarity mode.
Fig. 5 is that embodiment acts the result after " jump " carries out retrieving 4 iteration to retrieval by SVM relevant feedbacks.
Specific embodiment
The present invention is done with reference to the accompanying drawings and detailed description further is illustrated.
A kind of exercise data search method based on the description of attitude relative space-time characteristic statisticses, comprises the following steps:
Step 1,3 d pose relative space-time feature extraction:The point, line, surface geometric element collection that joint is formed in 3 d pose
Conjunction is the minimum Component units of the corresponding regional area of different patterns.Joint is formed in present invention extraction 3 d pose
The measurement of relative tertiary location and its change between point, line, surface geometric element as attitude character representation, by different offices
The weight combination of the different type feature that portion region includes, attitude space-time relative characteristic can express extensive gesture mode.It is right
Each attitude that everything is included in storehouse, its relative space-time feature extraction is comprised the following steps that:
1) 3 D human body joint model is defined, the 3 D human body joint model that the present invention is used as shown in Fig. 2 contain altogether
18 joints.According to each joint in the specific translation of sampling instant or rotation numerical, each joint of sampling instant is calculated
Three-dimensional coordinate (x, y, z);
2) geometric element set is built, the joint of selection constitutes the point set in geometric element set, and point concentrates any at 2 points
Straight line is formed, then constitute plane at any 3 points.The present invention is reasonably constrained straight line and plane the two geometric elements, only
Consider that the straight line and plane of adjacent segment formation, about to subtract the dimension of posture feature, are wrapped altogether in final attitude geometry element set
The straight line that 17 bones are formed, and 10 planes that adjacent segment is formed are contained;
3) extract each 3 d pose space feature, including joint adjust the distance feature, joint and bone distance feature,
Joint is with plan range feature, bone to angle feature, bone and plane included angle feature, plane and plane included angle feature, joint
Hyperspin feature.The 3 d pose space feature of extraction is as shown in figure 3, specific calculating process is as follows;
A) adjust the distance feature F in jointj,j,d, it is of the invention using between joint pair in Euclidean distance computational geometry element set
Distance, if joint j in attitude1、j2Three-dimensional coordinate be respectively (x1,y1,z1)、(x2,y2,z2), then joint to the distance between meter
Calculating formula is:
B) joint and bone distance feature Fj,l,d, the distance in joint to bone calculated by triangle area formula, if
d12、d13、d23Respectively joint j1、j2、j3The distance between, p=(d12+d23+d13)/2, then joint j1With joint j2、j3Formed
The distance between straight line be:
Fj,l,d=2p (p-d12)(p-d13)(p-d23)/d23;
C) joint and plan range feature Fj,p,d, the distance in joint to plane a little formed by joint and plane meaning of taking up an official post
Vector sum plane normal vector between dot product try to achieve, if n be joint j2、j3、j4The normal vector of the plane of formation, v is joint
j1、j3The vector of formation, then joint j1To j2、j3、j4The distance of the plane of formation is:
Fj,p,d=nv/ | | n | |;
D) bone is to angle feature Fl,l,a, bone is calculated with the angle of bone by dot product formula, if joint j1、
j2Form vector va, joint j3、j4Form vector vb, then the angle calcu-lation formula between bone be:
Fl,l,a=arccos (va·vb/(||va||×||vb||));
E) bone and plane included angle feature Fl,p,a, the angle between bone and plane is by bone and the point of plane normal vector
Product formula is calculated, if n is joint j3、j4、j5The normal vector of the plane P of formation, v is joint j1、j2The vector of formation, then close
Section j1、j2The bone of formation is with the angle of P:
Fl,p,a=arccos (nv/ (| | n | | × | | v | |));
F) plane and plane included angle feature Fp,p,a, the dot product public affairs of the normal vector that the angle between plane and plane passes through plane
Formula is calculated, if n1It is joint j1、j2、j3The plane P of formation1Normal vector, n2It is joint j4、j5、j6The plane P of formation2's
Normal vector, then P1With P2Angle be:
Fp,p,a=arccos (n1·n2/(||n1||×||n2||));
G) joint hyperspin feature Feuler, above configuration space feature only has one-dimension information, can not reflect the adjacent of three-dimensional
Joint rotation information, the present invention represents the rotation information of adjacent segment from Eulerian angles.
4) each 3 d pose relative time feature, including joint angle velocity and acceleration feature are extracted.The present invention is used
Kim et al. is in document Kim T H, Park S I, Shin S Y.Rhythmic-motion synthesis based on
motion-beat analysis[J].Acm Transactions on Graphics,2003,22(3):Proposed in 392-401
Method calculate the angular speed and acceleration in joint.The angular speed of extraction is as shown in Figure 3 with acceleration signature schematic diagram.Assuming that
Joint j is expressed as q in the rotation quaternary number of moment i-1 and ijAnd q (i-1)jI (), sampling interval duration is Δ t.
A) joint j is in the angular speed of moment i:
B) after angular velocity is calculated, acceleration can be tried to achieve according to the change of angular speed, i.e.,:
Step 2, action statistics description generation:Action is formed by several attitude consecutive variations, and the present invention extracts attitude phase
To space-time characteristic statistics description as action characteristic vector represent that the statistic of selection includes, average, extreme difference, variance with
And skewness totally 4 statistical variables.For each action in storehouse, what action statistics description was generated comprises the following steps that:
1) the relative space-time feature of each 3 d pose included in loading action, and using each feature as one-dimensional random
Variable, each attitude in action has corresponded to the value of the stochastic variable, and n is the attitude number included in acting, and is included in action
N attitude on a value for attitude relative space-time feature be x1,x2,...,xn;
2) average of each relative space-time feature is calculated:
3) extreme difference of each relative space-time feature is calculated:
R=max (x1, x2..., xn)-min(x1, x2..., xn);
4) variance of each relative space-time feature is calculated:
5) skewness of each relative space-time feature is calculated, skewness is the measurement of data distribution shape, and computational methods are:
6) by the average of each relative space-time feature extraction, extreme difference, variance, skewness, it is normalized, then by suitable
Sequence is arranged and preserved, and the characteristic vector for forming each action is represented.
Step 3, similarity mode:The feature for extracting the retrieval action example that user submits to using step 1 and step 2 is retouched
State, and and then using Euclidean distance calculate storehouse in action and retrieval action similarity degree, final similarity is arranged from high to low
Result action is returned after sequence.Similarity mode process is comprised the following steps that:
1) user submits to retrieval to act example;
2) the relative space-time characteristic statisticses description of the retrieval action that user submits to is calculated using step 1 and step 2;
3) for each action in storehouse, based on the description of relative space-time characteristic statisticses, calculated using Euclidean distance and user
The distance of the retrieval action of submission, the weight of each feature is identical.If the retrieval action of user is i, the candidate actions in storehouse are
J, NfIt is 3 d pose relative space-time feature sum, fi,k,lAnd fj,k,lRespectively action i and action j are special on attitude relative space-time
The value of the statistics description k of f is levied, then retrieval action is that the candidate actions in i and storehouse are the distance between j circular
For:
4) distance of the retrieval action submitted to each action and user in storehouse is ranked up;
5) Top-20 minimum action of distance is returned into user.
Step 4, relevant feedback:Each feature employs identical weight in the Similarity Measure of step 3.It is actually right
A specific character subset and its weight combination are all there is in the retrieval each time of user so that the performance of retrieval reaches most
It is excellent.In order to the retrieval for preferably catching user is intended to and the optimal character subset of Step wise approximation and its weight combination, the present invention is logical
Cross SVMs (SVM-Support Vector Machine) related feedback method to optimize retrieval result, based on SVM's
Relevant feedback is comprised the following steps that:
1) for step 3 or the result of last round of relevant feedback, user annotation returns to example and the retrieval of its submission is moved
Whether related make;
2) submit user annotation to the related sample of retrieval action as positive example sample to it, by user annotation and its
Retrieval is submitted to act incoherent sample as negative data;
3) linear kernel is selected, negative data is aligned using SVM and is learnt, obtain svm classifier model;
4) each action in storehouse is input to svm classifier model successively, calculates score;
5) score is ranked up, Top-20 action of highest scoring is returned into user;
If 6) user is satisfied with to result, retrieving terminates, otherwise return to step 1), carry out next round feedback.
The motion retrieval system realized using this programme acts the effect retrieved as shown in Figure 4 and Figure 5 to " jump ", figure
4 is using the result of similarity mode, because each feature uses the weight combination of identical weight, feature not pass through
Optimization, so in 9 retrieval results action that homepage shows, only 4 is related.Fig. 5 is related anti-by Linear SVM
Result after 4 iteration of feedback, optimizes feature weight and combines by on-line study disaggregated model, and retrieval performance has obtained significantly carrying
Height, shows that 9 retrieval results act all relevant actions in homepage.
It is specific real the invention provides a kind of exercise data search method based on the description of attitude relative space-time characteristic statisticses
Now the method and approach of the technical scheme are a lot, and the above is only the preferred embodiment of the present invention, it is noted that for this
For the those of ordinary skill of technical field, under the premise without departing from the principles of the invention, some improvement and profit can also be made
Decorations, these improvements and modifications also should be regarded as protection scope of the present invention.Each part being not known in the present embodiment can use
Prior art is realized.
Claims (5)
1. it is a kind of based on attitude relative space-time characteristic statisticses description exercise data search method, it is characterised in that including following
Step:
Step 1,3 d pose relative space-time feature extraction:Extract 3 d pose in joint formed point, line, surface geometric element it
Between relative tertiary location and its change measurement as attitude character representation, the inhomogeneity included by different regional areas
The weight of type feature combines to express extensive gesture mode;
Step 2, action statistics description generation:The statistics description of attitude relative space-time feature is extracted as the characteristic vector table of action
Show, the statistic of selection includes:Value, extreme difference, variance and skewness totally 4 statistical variables;
Step 3, similarity mode:The feature for extracting the retrieval action example that user submits to using step 1 and step 2 is described, and
And then using the similarity degree of action and retrieval action in Euclidean distance calculating storehouse, will after being sorted from high to low to final similarity
Result action is returned;
Step 4, relevant feedback:Each feature employs identical weight in the Similarity Measure of step 3, in practice for
The retrieval each time at family all has a specific character subset and its weight combination, by SVMs relevant feedback side
Method optimizes retrieval result, is intended to and the optimal character subset of Step wise approximation and its weight group with the retrieval for preferably catching user
Close.
2. method according to claim 1, it is characterised in that the step 1 includes:
3 D human body joint model 1-1) is defined, several joints of selection most important of which represent as 3 d pose;
Geometric element set 1-2) is built, the joint of selection constitutes the point set in geometric element set, and point concentrates any 2 shapes
It is in line, any 3 points then constitute plane;
1-3) extract each 3 d pose space feature, including joint adjust the distance feature, joint and bone distance feature, close
Section is with plan range feature, bone to angle feature, bone and plane included angle feature, plane and plane included angle feature, joint rotation
Turn feature;
1-4) extract each 3 d pose relative time feature, including joint angle velocity and acceleration feature.
3. method according to claim 1, it is characterised in that the step 2 includes:
2-1) the relative space-time feature of each 3 d pose included in loading action, and become using each feature as one-dimensional random
Amount, each attitude in action has corresponded to the value of the stochastic variable;
2-2) calculate the average of each relative space-time feature;
2-3) calculate the extreme difference of each relative space-time feature;
2-4) calculate the variance of each relative space-time feature;
2-5) calculate the skewness of each relative space-time feature;
2-6) will in order be arranged and protected after the average of each relative space-time feature extraction, extreme difference, variance, skewness normalization
Deposit, the characteristic vector for forming each action is represented.
4. method according to claim 1, it is characterised in that the step 3 includes:
3-1) user submits to retrieval to act example;
The relative space-time characteristic statisticses description of the retrieval action that user submits to 3-2) is calculated using step 1 and step 2;
3-3) for each action in storehouse, based on the description of relative space-time characteristic statisticses, calculated using Euclidean distance and user carries
The distance of the retrieval action of friendship, the weight of each feature is identical;
3-4) distance of the retrieval action submitted to each action and user in storehouse is ranked up;
Top-20 minimum action of distance 3-5) is returned into user.
5. method according to claim 1, it is characterised in that the step 4 includes:
4-1) for step 3 or the result of last round of relevant feedback, user annotation returns to example and the retrieval action of its submission
It is whether related;
4-2) user annotation is submitted to the related sample of retrieval action as positive example sample to it, user annotation is carried with it
Retrieval is handed over to act incoherent sample as negative data;
Linear kernel 4-3) is selected, negative data is aligned using SVM and is learnt, obtain svm classifier model;
Each action in storehouse 4-4) is input to svm classifier model successively, score is calculated;
4-5) score is ranked up, Top-20 action of highest scoring is returned into user;
If 4-6) user is satisfied with to result, retrieving terminates, otherwise return to step 4-1), carry out next round feedback.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611204708.4A CN106778639A (en) | 2016-12-23 | 2016-12-23 | A kind of exercise data search method based on the description of attitude relative space-time characteristic statisticses |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611204708.4A CN106778639A (en) | 2016-12-23 | 2016-12-23 | A kind of exercise data search method based on the description of attitude relative space-time characteristic statisticses |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106778639A true CN106778639A (en) | 2017-05-31 |
Family
ID=58897723
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611204708.4A Pending CN106778639A (en) | 2016-12-23 | 2016-12-23 | A kind of exercise data search method based on the description of attitude relative space-time characteristic statisticses |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106778639A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109241909A (en) * | 2018-09-06 | 2019-01-18 | 闫维新 | A kind of long-range dance movement capture evaluating system based on intelligent terminal |
CN109543054A (en) * | 2018-10-17 | 2019-03-29 | 天津大学 | A kind of Feature Dimension Reduction method for searching three-dimension model based on view |
CN109993818A (en) * | 2017-12-31 | 2019-07-09 | 中国移动通信集团辽宁有限公司 | Three-dimensional (3 D) manikin moves synthetic method, device, equipment and medium |
CN110362843A (en) * | 2018-11-20 | 2019-10-22 | 莆田学院 | A kind of visual human's entirety posture approximation generation method based on typical posture |
CN114267087A (en) * | 2022-02-28 | 2022-04-01 | 成都考拉悠然科技有限公司 | Action registration method and system based on hand sample machine learning model |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102298649A (en) * | 2011-10-09 | 2011-12-28 | 南京大学 | Space trajectory retrieval method of body movement data |
US20150186374A1 (en) * | 2013-12-27 | 2015-07-02 | Nuctech Company Limited | Retrieving system, retrieving method, and security inspection device based on contents of fluoroscopic images |
-
2016
- 2016-12-23 CN CN201611204708.4A patent/CN106778639A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102298649A (en) * | 2011-10-09 | 2011-12-28 | 南京大学 | Space trajectory retrieval method of body movement data |
US20150186374A1 (en) * | 2013-12-27 | 2015-07-02 | Nuctech Company Limited | Retrieving system, retrieving method, and security inspection device based on contents of fluoroscopic images |
Non-Patent Citations (4)
Title |
---|
TANG J K T ET AL: "《Retrieval of logically relevant 3D human motions by Adaptive Feature Selection》", 《PATTERN RECOGNITION LETTERS》 * |
XIANG SEAN ZHOU ET AL: "《relevance feedback in image retrieval:comprehensive review》", 《MULTIMEDIA SYSTEMS》 * |
胡莹等: "《一种改进的SVM相关反馈图像检索方法》", 《计算机应用研究》 * |
陈松乐等: "《民族舞蹈运动数据的实例检索方法》", 《计算机辅助设计与图形学学报》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109993818A (en) * | 2017-12-31 | 2019-07-09 | 中国移动通信集团辽宁有限公司 | Three-dimensional (3 D) manikin moves synthetic method, device, equipment and medium |
CN109993818B (en) * | 2017-12-31 | 2023-09-19 | 中国移动通信集团辽宁有限公司 | Method, device, equipment and medium for synthesizing motion of three-dimensional human body model |
CN109241909A (en) * | 2018-09-06 | 2019-01-18 | 闫维新 | A kind of long-range dance movement capture evaluating system based on intelligent terminal |
CN109543054A (en) * | 2018-10-17 | 2019-03-29 | 天津大学 | A kind of Feature Dimension Reduction method for searching three-dimension model based on view |
CN109543054B (en) * | 2018-10-17 | 2022-12-09 | 天津大学 | View-based feature dimension reduction three-dimensional model retrieval method |
CN110362843A (en) * | 2018-11-20 | 2019-10-22 | 莆田学院 | A kind of visual human's entirety posture approximation generation method based on typical posture |
CN114267087A (en) * | 2022-02-28 | 2022-04-01 | 成都考拉悠然科技有限公司 | Action registration method and system based on hand sample machine learning model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Liu et al. | Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding | |
Punnakkal et al. | BABEL: Bodies, action and behavior with english labels | |
CN106778639A (en) | A kind of exercise data search method based on the description of attitude relative space-time characteristic statisticses | |
Wang et al. | Modeling temporal dynamics and spatial configurations of actions using two-stream recurrent neural networks | |
Zhang et al. | On geometric features for skeleton-based action recognition using multilayer lstm networks | |
Cai et al. | Effective active skeleton representation for low latency human action recognition | |
CN107742102B (en) | Gesture recognition method based on depth sensor | |
Wan et al. | Explore efficient local features from RGB-D data for one-shot learning gesture recognition | |
Zhu et al. | Evaluating spatiotemporal interest point features for depth-based action recognition | |
Sheng et al. | GreenSea: visual soccer analysis using broad learning system | |
Kang et al. | Review of action recognition and detection methods | |
Feng et al. | Depth-projection-map-based bag of contour fragments for robust hand gesture recognition | |
CN102298649B (en) | Space trajectory retrieval method of body movement data | |
CN105574510A (en) | Gait identification method and device | |
CN109460734A (en) | The video behavior recognition methods and system shown based on level dynamic depth projection difference image table | |
CN108897791A (en) | A kind of image search method based on depth convolution feature and semantic similarity amount | |
Zhang et al. | Human pose estimation and tracking via parsing a tree structure based human model | |
CN107169117A (en) | A kind of manual draw human motion search method based on autocoder and DTW | |
CN108205684A (en) | Image disambiguation method, device, storage medium and electronic equipment | |
CN107479693A (en) | Real-time hand recognition methods based on RGB information, storage medium, electronic equipment | |
Ye et al. | 3D sketch-based 3D model retrieval with convolutional neural network | |
CN102855488A (en) | Three-dimensional gesture recognition method and system | |
Valcik et al. | Assessing similarity models for human‐motion retrieval applications | |
CN104881651B (en) | A kind of personage's Activity recognition method based on accidental projection and Fisher vector | |
Li et al. | Augmenting bag-of-words: a robust contextual representation of spatiotemporal interest points for action recognition |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170531 |
|
WD01 | Invention patent application deemed withdrawn after publication |