CN109858430A - A kind of more people's attitude detecting methods based on intensified learning optimization - Google Patents
A kind of more people's attitude detecting methods based on intensified learning optimization Download PDFInfo
- Publication number
- CN109858430A CN109858430A CN201910080912.7A CN201910080912A CN109858430A CN 109858430 A CN109858430 A CN 109858430A CN 201910080912 A CN201910080912 A CN 201910080912A CN 109858430 A CN109858430 A CN 109858430A
- Authority
- CN
- China
- Prior art keywords
- people
- training
- target
- human body
- frame
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The present invention provides a kind of more people's attitude detecting methods based on intensified learning optimization.It cannot be bonded human body well for some encirclement frames that object detector in conventional method is positioned, the detection accuracy of gesture detector is caused to decline, and then influence the precision of entire more people's attitude detection algorithms, the target refined model based on intensified learning is proposed, is adjusted for inaccurate encirclement frame.Target refined model enables encirclement frame to be more bonded human body, reduces the redundancy for surrounding image in frame, the detection accuracy of gesture detector can be improved.
Description
Technical field
The invention belongs to the technical fields of image procossing, and in particular to a kind of more people's postures inspection based on intensified learning optimization
Survey method.
Background technique
Intensified learning is the machine learning branch inspired by game theory and behaviour psychology, is a kind of object-oriented
How study is taken some movements so that target optimizes, as learnt for a problem in training by algorithm, the algorithm
How to operate to obtain highest score in gaming.Intensified learning can be since a randomly selected state, not
It can achieve the ability more than human levels in disconnected training.It will receive and punish when doing the prediction to make mistake in the training of intensified learning
Penalize, it is on the contrary then will receive reward, intensified learning be be constantly awarded with learn how to select to act after punishment so that
Maximized process must be rewarded.
At present for the judgement of portrait posture in image, the detection in particular for more people's postures mainly uses two steps to detect
Algorithm: target is positioned with object detector, with gesture detector to all target detection postures.It is detected and is schemed using object detector
Portrait as in, then detects the posture for judging portrait using gesture detector, but just carries out in this way to portrait
Judgement, since some encirclement frames that object detector is positioned cannot sometimes be bonded human body well, that is, it is only determined
Position human body, human body have part surrounding in frame, some is then surrounding outer frame portion, are allowed for when human posture judges in this way
The detection accuracy of gesture detector declines, and affects entire more people's postures there are more inaccurate when multiple people in image
The precision of detection algorithm.
So need to make improvement to existing more people's attitude detections, to improve arithmetic accuracy, in accurate judgement image
Multiple people human body attitude.
Summary of the invention
The purpose of the present invention is to solve the above problems, provide a kind of more people's attitude detections based on intensified learning optimization
Method.It cannot be bonded human body well for some encirclement frames that object detector in conventional method is positioned, posture is caused to be examined
The detection accuracy decline of device is surveyed, and then influences the precision of entire more people's attitude detection algorithms, is proposed based on intensified learning
Target refined model is adjusted for inaccurate encirclement frame.Target refined model enables encirclement frame to be more bonded people
Body reduces the redundancy for surrounding image in frame, the detection accuracy of gesture detector can be improved.
In order to achieve the above object of the invention, the invention adopts the following technical scheme:
A kind of more people's attitude detecting methods based on intensified learning optimization, which comprises the following steps:
S1. multiple more people's pictures are acquired and are handled, more personal data collection and single data set are generated, by it is described one with
More personal data collection are split as training set and test set by preset ratio respectively, obtain more people's training sets, more people's test sets, single instruction
Practice collection, single test set；
S2. establish object detector for positioning target, for the target refined model of adjustment package peripheral frame, for detecting
The gesture detector of human body attitude, three form more people's attitude detection algorithm structures；
S3. using the characteristic extraction part in single training set training objective refined model, as target refined model
Pre-training parameter, and tested using single test set to prevent model over-fitting；
S4. more people's training set training objective detectors are utilized, using single training set training gesture detector, and using each
From test set (i.e. more people's training sets, single training set) carry out each self-test respectively to prevent model over-fitting；Use target
The training precision of detector and gesture detector, which generates the single image data of square and assembles for training, practices target refined model, in training
It is initialized using the refined model of pre-training parameters on target described in S3；
S5. input needs the more people's pictures detected, is positioned in the more people's pictures for needing to detect using object detector
Multiple human body targets, be adjusted using encirclement frame of the target refined model to multiple human body targets, and utilize attitude detection
Device detects the posture of multiple human body targets respectively.
Invention increases target refined models, for adjusting the encirclement frame of object detector positioning, so that surrounding frame more
The human body being bonded in more people's pictures reduces the redundancy for surrounding image in frame, and the detection essence of gesture detector can be improved
Degree.
Further, more personal data collection in the step S1 include that more people's pictures and human body surround box label.
Further, the single data set in the step S1 is including the use of surrounding the single picture after frame is cut and every
The body joint point coordinate of a human body.
Further, the step S1 specifically includes following procedure:
S11. more people's pictures are acquired, multiple human body targets in more people's pictures are positioned using frame is surrounded, and saves
The encirclement frame coordinate of multiple human body targets surrounds frame coordinate and is made of upper left angle point and bottom right angle point, forms more personal data collection；
S12. the initial body joint point coordinate of each of each of more people's pictures body target is positioned, with each human body
The encirclement frame of target, which corresponds, saves each initial body joint point coordinate, and the initial body joint point coordinate is by single coordinate points
It constitutes；S13. human body target is cut to obtain single picture according to the encirclement frame that more personal datas are concentrated, by the list after cutting
People's picture mends into the single picture of square for the length that side length is single humanoid figure piece long side by way of zero padding around；It will be described more
Body joint point coordinate in people's picture maps in the single picture of square, saves body joint point coordinate data, forms single number
According to collection；
S14. randomly selecting for total quantity 10% is carried out to more personal data collection and single data set, as more people's test sets and
Single test set, remaining picture is as more people's training sets and single training set.Here 10% can voluntarily be set according to practical
It sets.
Further, when surrounding's zero padding in the present invention refers to that picture is non-square, RGB brightness is supplemented around picture
The pixel for being zero becomes square picture.
Further, the step S4 specifically includes following procedure:
S41. more people's training set training objective detectors are utilized, and after training with more people's test sets to object detector
It is tested to prevent its over-fitting；Human body target detection is carried out to more people's pictures in more people's training sets and test set, to every
Human body target in more people's pictures positions and saves all encirclements frame coordinates using surrounding frame, the encirclement frame coordinate and
Every more people's pictures correspond；
S42. the more people's training sets obtained through object detector in S41 is utilized to surround frame to described with test set human body target
More people's pictures in more people's training sets and test set are cut to obtain single picture, use surrounding according to the long side of single picture
The mode of zero padding mends into the single picture of square, and the multiple single picture of square forms the single picture of square after cutting
Training set and test set, image credit are corresponding with more people's training sets and test set；
S43. using single training set training gesture detector, and after training to single test set to target detection
Device is tested to prevent its over-fitting；Using trained gesture detector to the single picture training set of square and test set
It carries out attitude detection and saves the detection body joint point coordinate that detection obtains, the detection body joint point coordinate and the single picture of square
It corresponds；
S44. the detection artis of the single image data collection (including training set and test set) of square in S43 is calculated one by one
Coordinate precision corresponding with the human body encirclement frame coordinate label of all single pictures of square gone out from single more people's croppings,
Calculation method is detection artis number identical with initial body joint point coordinate in one single picture of square of statistics, divided by institute
State the artis number in the single picture of square；Choosing the highest label of precision is humanoid mesh corresponding to more people's pictures
Mark, and save as original precision；
S45. start after being initialized using the characteristic extraction part of pre-training parameters on target refined model described in S3
Training, target refined model read the encirclement frame coordinate of the object detector and are adjusted to it, form people adjusted
Body surrounds frame；
S46. attitude detection is carried out using the human body target that gesture detector surrounds in frame the human body adjusted, and
And compared with original precision, obtain the reward that intensified learning intelligent body needs.
Human body in step s44 of the invention surrounds the corresponding precision of frame coordinate label, refers to object detector and posture
The training precision of detector, in particular to more people's pictures in data set pass through the target detection of object detector, using appearance
Precision after the attitude detection of state detector.
Further, the step S5 includes following procedure:
S51. people's picture more than one is read using object detector, carry out identification and confines a human body target with surrounding；
S52. the magnitude range for surrounding frame is adjusted using target refined model；
S53. attitude detection is carried out to the human body target in encirclement frame adjusted using gesture detector；
S54. the result of attitude detection is mapped back into more people's pictures.
Further, the object detector, including for positioning human body target extraction character network and coordinate return
Network；
Or, the target refined model, including character network and Q network are extracted for adjustment package peripheral frame；
Or, the gesture detector, including the extraction character network and coordinate Recurrent networks for attitude detection.
It further, include convolutional layer in the structure of the object detector, BN layers, pond layer and full articulamentum；
Or, include convolutional layer in the structure of the target refined model, BN layers, pond layer and full articulamentum；
Or, the adjustment package peripheral frame, refers to and carries out four, upper and lower, left and right side to the coordinate for surrounding the frame upper left corner and the lower right corner
It is adjusted to totally eight kinds
Further, the target refined model further includes terminating the termination movement of adjustment to the adjustment movement for surrounding frame.
The long side in encirclement frame long side or single humanoid figure piece long side in the present invention is to refer to the longer side of side length in rectangle.
The present invention is optimized using testing result of the target refined model to object detector, realizes an extensive chemical
It practises, so that the encirclement frame determination of human body target is more accurate.
Compared with prior art, the present invention beneficial effect is: two step detection algorithms of traditional more people's attitude detections are based on
Object detector is built with gesture detector, since the encirclement frame precision that object detector is positioned not enough causes posture to be examined
Survey accuracy decline；After algorithm of the invention, using the target refined model based on intensified learning to more people's attitude detections
Method optimizes, so that surrounding frame is bonded human body more to promote detection accuracy.
In addition, the present invention uses pre-training network of the trained human body disaggregated model as target refined model,
Compared to traditional disaggregated model based on ImageNet, human body disaggregated model focuses more on the extraction to characteristics of human body,
It is more suitable for the pre-training model of attitude detection.
Detailed description of the invention
Fig. 1 is the schematic diagram of intensified learning of the present invention；
Fig. 2 is target refined model workflow schematic diagram；
Fig. 3 is target refined model structural schematic diagram；
Fig. 4 is target refined model training flow chart；
Fig. 5 is the optimum results schematic diagram using the more people's detection algorithms of the present invention.
Specific embodiment
Below by specific embodiment the technical scheme of the present invention will be further described explanation so that the technical program is more
Add clear, clear.
The present invention is based on more people's attitude detection algorithms of intensified learning optimization using target refined model to object detector
After the encirclement frame of positioning is adjusted, then attitude detection is carried out, effectively increases the precision of more people's attitude detections.
Multiple more people's pictures are acquired the present embodiment provides S1. and are handled, and more personal data collection and single data set are generated,
One and more personal data collection are split as training set and test set by preset ratio respectively, obtain more people's training sets, more people
Test set, single training set, single test set；
S2. establish object detector for positioning target, for the target refined model of adjustment package peripheral frame, for detecting
The gesture detector of human body attitude, three form more people's attitude detection algorithm structures；
S3. using the characteristic extraction part in single training set training objective refined model, as target refined model
Pre-training parameter, and tested using single test set to prevent model over-fitting；
S4. more people's training set training objective detectors are utilized, using single training set collection training gesture detector, and are used
Respective test set is tested respectively to prevent model over-fitting；Use the training precision of object detector and gesture detector
It generates the single image data of square and assembles for training and practice target refined model, pre-training parameters on target described in S3 is used in training
Refined model is initialized；
S5. input needs the more people's pictures detected, is positioned in the more people's pictures for needing to detect using object detector
Multiple human body targets, be adjusted using encirclement frame of the target refined model to multiple human body targets, and utilize attitude detection
Device detects the posture of multiple human body targets respectively.
Target refined model is by extraction character network OR1With Q network OR2Composition, target refined model (pass through intelligent body
Realize) status information can be obtained with environmental interaction, markov decision process is established, as shown in Figure 1.In each iteration,
Model needs acquire information to determine a deformed movement, in iteration next time, after model can be according to last time deformation
Information determine again the deformed movement of next iteration, until the number of iterations for determining that target is optimal or reach limitation is
Only.After movement executes every time, algorithm calculates the reward that the movement is executed under the state.
The movement A and state st of target refined model, are controlled by function Q (st, A), which can learn letter by Q
Number is estimated.Model can select that the movement of reward can be obtained by function.Q learning function uses following Bellman equation
Continuous iteration updates model parameter:
Q (st, A)=R+ γ maxa'Q(st',A')
Wherein st and A is current corresponding movement and state, and R is current reward, maxa'Q (st', A') is indicated not
The reward come, γ indicate discount factor.
Target refined model carries out decision by the feature extracted by convolutional neural networks, selects institute under current state
The movement that should be selected.Target refined model is by two kinds of movement: one is adjustment acts, the movement of the type can be adjusted
Surround the shape of frame；The second is termination acts, once being selected, adjustment process is terminated for the movement of the type.Adjustment therein
Amount of action has eight kinds, is four for surrounding the four direction translation of frame top left co-ordinate, and surrounding frame bottom right angular coordinate respectively
Direction translation.The reasons why designing in this way is that the everything possibility for surrounding frame is covered in this eight kinds movements, compared to general encirclement
The rule action of frame zooming and panning can be designed so that surrounding frame makes irregular movement in this way, and being more advantageous to makes to surround
Frame is close to human body.Model can be acted constantly according to current state selection in an iterative process, can be obtained after each adjustment package peripheral frame
Obtain state newly, the new movement of reselection, until being selected as termination movement.Model flow is as shown in Figure 2.
The selected movement of target refined model can generate new encirclement frame, and gesture detector PE can be according to new encirclement frame
Generate new precision acc1；Algorithm defines the precision acc for being added without two step detection frameworks of intensified learning0As true value.For
The adjustment of current state st, intelligent body selection act to obtain new state st', the new precision acc of generation1, if it is greater than true
Value acc0, then it can obtain a reward (1), it is on the contrary then a punishment (- 1) can be obtained.For termination movement, when termination most
Whole new precision acc1If more than acc0, a bigger reward can be obtained, otherwise one big punishment can be obtained.And for true
Real value is greater than the target of τ, and algorithms selection directly allows intelligent body selection termination to act, rewarded.It is as follows to reward formula:
Ra(st, st')=sign (acc1-acc0)
The present embodiment propose based on intensified learning optimization more people's attitude detecting methods, model structure as shown in figure 3,
Its structure includes:
1, feature extraction network OR1, it is made of multiple convolutional layers, is used for feature extraction；
2, fc layers, full articulamentum, for multidimensional characteristic to be mapped as one-dimensional characteristic vector；
3, Q network, the full articulamentum exported by two 512 form, since information maps；
4, act vector, by one 9 connect full articulamentum form, for export the movement containing nine elements to
Amount represents eight adjustment movements and acts with a termination.
The model training stage operational process of the present embodiment as shown in figure 4, its process the following steps are included:
1, by the characteristic extraction part OR of human body disaggregated modeloriFeature extraction network OR as target refined model1's
Pre-training model；
2, original image Img is inputtedMPIn original packet peripheral frame bbox to target refined model OR, Q network OR2According to feature
Extract network OR1The feature extracted is adjusted to frame is surrounded, and obtains newly surrounding frame bbox '；
3, small figure is cut out using the new frame that surrounds, attitude detection is carried out to it using gesture detector, obtains new precision acc1；
4, using reward formula to new precision acc1With original precision acc0Reward calculating is carried out, and updates Q network OR2's
Parameter.
Method actually uses process, inputs more people's picture ImgMPBy object detector OD, target refined model OR, posture
More people's attitude detection results can be obtained in detector PE.Obtained detection effect figure is as shown in Figure 5.
The target refined model of the present embodiment is to be bonded insufficient encirclement frame for human body to be adjusted.Modelling
It surrounds the frame upper left corner and two, lower right corner point respectively carries out eight adjustment movement and one of upper and lower, left and right four direction translation
A termination movement for stopping adjustment, is iterated adjustment to frame is surrounded by markov decision process, finally to wrap
Peripheral frame is bonded human body more to promote detection accuracy.
More people's attitude detecting methods based on intensified learning optimization of the present embodiment are mainly pressed further for refinement
Following steps carry out:
1. processing is used for the data set of more people's attitude detections, more people's picture Img are obtainedMPBox label is surrounded with human body
LabelbboxMore personal data collection DMP, original image is cut to obtain single picture according to frame is surrounded, further according to picture long side
Single picture after length will be cut mends into the single picture Img of square by way of zero padding aroundPAnd each human body
Body joint point coordinate LabelkpSingle data set DP；
2. establishing object detector OD positioning target, target refined model OR adjustment package peripheral frame, gesture detector PE detection
More people's attitude detection algorithm structures of posture, wherein object detector includes extracting character network OD1With coordinate Recurrent networks OD2,
Target refined model includes extracting character network OR1With Q network OR2, gesture detector includes extracting character network PE1And coordinate
Recurrent networks PE2；
3. feature extraction network OR1It is the convolutional neural networks model OR an of standardoriCharacteristic extraction part, use
Single data set DPBy ORoriTraining becomes an object-class model, the neuron number of the full articulamentum of the last layer of model
It is two, represents two classifications, be background classes and the mankind respectively；By the object-class model OR after the completion of trainingoriFull connection
Layer is deleted, and has obtained one for extracting the feature extraction network OR of characteristics of human body1；
4. using more personal data collection DMPObject detector OD is trained and is used single data set DPTo attitude detection
Device PE is trained, using object detector OD to more personal data collection D after the completion of trainingMPTarget detection is carried out, will test to obtain
Encirclement frame bbox cut to obtain single picture, and the single picture after cutting according to long side passes through surrounding zero padding
Mode mends into the single picture of square, then the single picture of square is carried out attitude detection with gesture detector PE；Calculate every
The precision corresponding with all labels in original image of the attitude detection result in small figure cut, choosing the highest label of precision is
Target corresponding to the small figure, and save as original precision acc0；
5. the encirclement frame bbox that object detector in step 4 detects is adjusted using target refined model OR.
Model is designed using intensified learning, nine kinds of movements of model output, respectively for the upper left corner and the right side for surrounding frame bbox
The coordinate of inferior horn carries out the totally eight kinds of movements of upper and lower, left and right four direction, and terminates the termination movement of adjustment.Model is to encirclement
Frame generates new encirclement frame bbox ' after being adjusted, according to the attitude detection of bbox ' progress such as step 4, and obtain new precision
acc1.It is defined according to the following formulas the reward value R of intensified learning:
R=sign (acc1-acc0)
Wherein, if acc1Greater than acc0When, then reward value is 1, otherwise is 0；
The movement A and state st of model, are controlled by function Q (st, A), which can be estimated by Q learning function
Meter.Model can select that the movement of reward can be obtained by function.Q learning function uses the following continuous iteration of Bellman equation
Update model parameter:
Q (st, A)=R+ γ maxa'Q(st',A')
Wherein st and A is current corresponding movement and state, and R is current reward, maxa'Q (st', A') is indicated not
The reward come, γ indicate discount factor.
6. utilizing step 1,2,3,4,5 can train to obtain more people's attitude detection algorithms based on intensified learning optimization, input
Original more people's picture ImgMP, multiple human body targets are positioned by object detector OD, target refined model OR adjusts target and surrounds
Frame bbox, gesture detector PE detect targeted attitude respectively, achieve the purpose that more people's attitude detections.
The present invention is by the target refined model OR based on intensified learning to the obtained encirclement frame of object detector OD
Bbox is adjusted, it is made more to be bonded human body, to achieve the purpose that promote more people's attitude detection precision.
Pass through the above-mentioned target refined model OR based on intensified learning encirclement frame bbox obtained to object detector OD
It is adjusted, it is made more to be bonded human body, to achieve the purpose that promote more people's attitude detection precision.
The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention.The technology of the industry
Personnel are it should be appreciated that the present invention is not limited to the above embodiments, and the above embodiments and description only describe this
The principle of invention, without departing from the spirit and scope of the present invention, various changes and improvements may be made to the invention, these changes
Change and improvement all fall within the protetion scope of the claimed invention.The claimed scope of the invention by appended claims and its
Equivalent thereof.Specific embodiment described herein is only an example for the spirit of the invention.Skill belonging to the present invention
The technical staff in art field can make various modifications or additions to the described embodiments or using similar side
Formula substitution, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.
Claims (10)
1. a kind of more people's attitude detecting methods based on intensified learning optimization, which comprises the following steps:
S1. multiple more people's pictures are acquired and are handled, more personal data collection and single data set are generated, by one and more people
Data set is split as training set and test set by preset ratio respectively, obtains more people's training sets, more people's test sets, single training
Collection, single test set；
S2. establish object detector for positioning target, for the target refined model of adjustment package peripheral frame, for detecting human body
The gesture detector of posture, three form more people's attitude detection algorithm structures；
S3. using the characteristic extraction part in single training set training objective refined model, the pre- instruction as target refined model
Practice parameter, and is tested using single test set to prevent model over-fitting；
S4. more people's training set training objective detectors are utilized, using single training set collection training gesture detector, and using respective
Test set tested respectively to prevent model over-fitting；It is generated using the training precision of object detector and gesture detector
The single image data of square, which is assembled for training, practices target refined model, fine using pre-training parameters on target described in S3 in training
Model is initialized；
S5. input needs the more people's pictures detected, is positioned using object detector more in the more people's pictures for needing to detect
A human body target is adjusted using encirclement frame of the target refined model to multiple human body targets, and utilizes gesture detector point
The posture of multiple human body targets is not detected.
2. a kind of more people's attitude detecting methods based on intensified learning optimization according to claim 1, which is characterized in that institute
Stating more personal data collection in step S1 includes that more people's pictures and human body surround frame coordinate label.
3. a kind of more people's attitude detecting methods based on intensified learning optimization according to claim 1, which is characterized in that institute
The single data set in step S1 is stated including the use of the initial artis for surrounding the single picture after frame is cut and each human body
Coordinate.
4. a kind of more people's attitude detecting methods based on intensified learning optimization according to claim 1 or 2 or 3, feature
It is, the step S1 specifically includes following procedure:
S11. more people's pictures are acquired, multiple human body targets in more people's pictures are positioned using frame is surrounded, and saves multiple
The encirclement frame coordinate of human body target surrounds frame coordinate and is made of upper left angle point and bottom right angle point, forms more personal data collection；
S12. the initial body joint point coordinate of each of each of more people's pictures body target is positioned, with each human body target
Encirclement frame correspond and save each initial body joint point coordinate, the initial body joint point coordinate is by single coordinate points structure
At；
S13. human body target is cut to obtain single picture according to the encirclement frame that more personal datas are concentrated, one after cutting
Picture mends into the single picture of square for the length that side length is single humanoid figure piece long side by way of zero padding around；By more people
Body joint point coordinate in picture maps in the single picture of square, saves body joint point coordinate data, forms single data
Collection；
S14. randomly selecting for total quantity 10% is carried out to more personal data collection and single data set, as more people's test sets and one
Test set, remaining picture is as more people's training sets and single training set.
5. a kind of more people's attitude detecting methods based on intensified learning optimization according to claim 4, which is characterized in that institute
State zero padding around refer to picture be non-square when, around picture supplement RGB zero luminance pixel, become pros
Shape picture.
6. a kind of more people's attitude detecting methods based on intensified learning optimization according to claim 1 or 2 or 3, feature
It is, the step S4 specifically includes following procedure:
S41. more people's training set training objective detectors are utilized, and object detector is carried out with more people's test sets after training
Test is to prevent its over-fitting；Human body target detection is carried out to more people's pictures in more people's training sets and test set, it is more to every
Human body target in people's picture positions and saves all encirclements frame coordinates using surrounding frame, the encirclement frame coordinate and every
More people's pictures correspond；
S42. the more people's training sets obtained through object detector in S41 and test set human body target is utilized to surround frame to more people
More people's pictures in training set and test set are cut to obtain single picture, use surrounding zero padding according to the long side of single picture
Mode mend into the single picture of square, the multiple single picture composition of square cut after square single picture training
Collection and test set, image credit are corresponding with more people's training sets and test set；
S43. using single training set training gesture detector, and after training to single test set to object detector into
Row test is to prevent its over-fitting；The single picture training set of square and test set are carried out using trained gesture detector
Attitude detection simultaneously saves the detection body joint point coordinate that detection obtains, and the detection body joint point coordinate and the single picture of square are one by one
It is corresponding；
S44. the detection body joint point coordinate of the single image data collection (including training set and test set) of square in S43 is calculated one by one
Precision corresponding with the human body encirclement frame coordinate label of all single pictures of square gone out from single more people's croppings, calculates
Method is identical with the initial body joint point coordinate number of detection artis in one single picture of square of statistics, divided by it is described just
Artis number in rectangular single picture；Choosing the highest label of precision is humanoid target corresponding to more people's pictures, and
Save as original precision；
S45. start to train after being initialized using the characteristic extraction part of pre-training parameters on target refined model described in S3,
Target refined model reads the encirclement frame coordinate of the object detector and is adjusted to it, forms human body adjusted and surrounds
Frame；
S46. using gesture detector to the human body target progress attitude detection in the human body encirclement frame adjusted, and with
Original precision compares, and obtains the reward that intensified learning intelligent body needs.
7. a kind of more people's attitude detecting methods based on intensified learning optimization according to claim 6, which is characterized in that institute
Stating step S5 includes following procedure:
S51. the more people's pictures for needing to detect are read using object detector, carry out identification and confine a human body target with surrounding；
S52. the magnitude range for surrounding frame is adjusted using target refined model；
S53. attitude detection is carried out to the human body target in encirclement frame adjusted using gesture detector；
S54. the result of attitude detection is mapped back into more people's pictures.
8. a kind of more people's attitude detecting methods based on intensified learning optimization according to claim 1 or 2 or 3, feature
It is,
The object detector, including the extraction character network and coordinate Recurrent networks for positioning human body target；
Or, the target refined model, including character network and Q network are extracted for adjustment package peripheral frame；
Or, the gesture detector, including the extraction character network and coordinate Recurrent networks for attitude detection.
9. a kind of more people's attitude detecting methods based on intensified learning optimization according to claim 1 or 2 or 3, feature
It is,
It include convolutional layer in the structure of the object detector, BN layers, pond layer and full articulamentum；
Or, include convolutional layer in the structure of the target refined model, BN layers, pond layer and full articulamentum；Or, the posture
It include convolutional layer in the structure of detector, BN layers, pond layer and full articulamentum；
Or, the adjustment package peripheral frame, refers to that carrying out upper and lower, left and right four direction to the coordinate for surrounding the frame upper left corner and the lower right corner is total to
Eight kinds of adjustment.
10. a kind of more people's attitude detecting methods based on intensified learning optimization according to claim 9, which is characterized in that
The target refined model further includes terminating the termination movement of adjustment to the adjustment movement for surrounding frame.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910080912.7A CN109858430A (en) | 2019-01-28 | 2019-01-28 | A kind of more people's attitude detecting methods based on intensified learning optimization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910080912.7A CN109858430A (en) | 2019-01-28 | 2019-01-28 | A kind of more people's attitude detecting methods based on intensified learning optimization |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109858430A true CN109858430A (en) | 2019-06-07 |
Family
ID=66896472
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910080912.7A Pending CN109858430A (en) | 2019-01-28 | 2019-01-28 | A kind of more people's attitude detecting methods based on intensified learning optimization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109858430A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079851A (en) * | 2019-12-27 | 2020-04-28 | 常熟理工学院 | Vehicle type identification method based on reinforcement learning and bilinear convolution network |
| CN111415389A (en) * | 2020-03-18 | 2020-07-14 | 清华大学 | Label-free six-dimensional object posture prediction method and device based on reinforcement learning |
| CN113205043A (en) * | 2021-04-30 | 2021-08-03 | 武汉大学 | Video sequence two-dimensional attitude estimation method based on reinforcement learning |
| CN114092556A (en) * | 2021-11-22 | 2022-02-25 | 北京百度网讯科技有限公司 | Method, apparatus, electronic device, medium for determining human body posture |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105637540A (en) * | 2013-10-08 | 2016-06-01 | 谷歌公司 | Methods and apparatus for reinforcement learning |
| CN106778740A (en) * | 2016-12-06 | 2017-05-31 | 北京航空航天大学 | A kind of TFDS non-faulting image detecting methods based on deep learning |
| CN108229445A (en) * | 2018-02-09 | 2018-06-29 | 深圳市唯特视科技有限公司 | A kind of more people's Attitude estimation methods based on cascade pyramid network |
| CN108805268A (en) * | 2018-06-08 | 2018-11-13 | 中国科学技术大学 | Deeply learning strategy network training method based on evolution algorithm |
-
2019
- 2019-01-28 CN CN201910080912.7A patent/CN109858430A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105637540A (en) * | 2013-10-08 | 2016-06-01 | 谷歌公司 | Methods and apparatus for reinforcement learning |
| CN106778740A (en) * | 2016-12-06 | 2017-05-31 | 北京航空航天大学 | A kind of TFDS non-faulting image detecting methods based on deep learning |
| CN108229445A (en) * | 2018-02-09 | 2018-06-29 | 深圳市唯特视科技有限公司 | A kind of more people's Attitude estimation methods based on cascade pyramid network |
| CN108805268A (en) * | 2018-06-08 | 2018-11-13 | 中国科学技术大学 | Deeply learning strategy network training method based on evolution algorithm |
Non-Patent Citations (6)
| Title |
|---|
| ALEJANDRO NEWELL,AT EL.: ""Stacked Hourglass Networks for Human Pose Estimation"", 《ARXIV》 * |
| ELDAR INSAFUTDINOV,AT EL.: ""DeeperCut: A Deeper, Stronger, and Faster Multi-Person Pose Estimation Model"", 《ARXIV》 * |
| HAO-SHU FANG,AT EL.: ""RMPE: Regional Multi-Person Pose Estimation"", 《ARXIV》 * |
| JUAN C. CAICEDO,AT EL.: ""Active Object Localization with Deep Reinforcement Learning"", 《ARXIV》 * |
| VOLODYMYR MNIH,AT EL.: ""Human-level control through deep reinforcement learning"", 《NATURE》 * |
| YILUN CHEN,AT EL.: ""Cascaded Pyramid Network for Multi-Person Pose Estimation"", 《ARXIV》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079851A (en) * | 2019-12-27 | 2020-04-28 | 常熟理工学院 | Vehicle type identification method based on reinforcement learning and bilinear convolution network |
| CN111079851B (en) * | 2019-12-27 | 2020-09-18 | 常熟理工学院 | Vehicle type identification method based on reinforcement learning and bilinear convolution network |
| CN111415389A (en) * | 2020-03-18 | 2020-07-14 | 清华大学 | Label-free six-dimensional object posture prediction method and device based on reinforcement learning |
| CN111415389B (en) * | 2020-03-18 | 2023-08-29 | 清华大学 | Label-free six-dimensional object posture prediction method and device based on reinforcement learning |
| CN113205043A (en) * | 2021-04-30 | 2021-08-03 | 武汉大学 | Video sequence two-dimensional attitude estimation method based on reinforcement learning |
| CN113205043B (en) * | 2021-04-30 | 2022-06-07 | 武汉大学 | Video sequence two-dimensional attitude estimation method based on reinforcement learning |
| CN114092556A (en) * | 2021-11-22 | 2022-02-25 | 北京百度网讯科技有限公司 | Method, apparatus, electronic device, medium for determining human body posture |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107145908B (en) | A kind of small target detecting method based on R-FCN | |
| CN109858430A (en) | A kind of more people's attitude detecting methods based on intensified learning optimization | |
| CN108229268A (en) | Expression Recognition and convolutional neural networks model training method, device and electronic equipment | |
| CN114220035A (en) | Rapid pest detection method based on improved YOLO V4 | |
| CN109800628A (en) | A kind of network structure and detection method for reinforcing SSD Small object pedestrian detection performance | |
| CN109635875A (en) | A kind of end-to-end network interface detection method based on deep learning | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN107862694A (en) | A kind of hand-foot-and-mouth disease detecting system based on deep learning | |
| CN107169435A (en) | A kind of convolutional neural networks human action sorting technique based on radar simulation image | |
| CN108615046A (en) | A kind of stored-grain pests detection recognition methods and device | |
| CN110211173A (en) | A kind of paleontological fossil positioning and recognition methods based on deep learning | |
| CN113435282B (en) | Unmanned aerial vehicle image ear recognition method based on deep learning | |
| CN109242829A (en) | Liquid crystal display defect inspection method, system and device based on small sample deep learning | |
| CN108648211A (en) | A kind of small target detecting method, device, equipment and medium based on deep learning | |
| CN108334878A (en) | Video images detection method and apparatus | |
| CN109508661A (en) | A kind of person's of raising one's hand detection method based on object detection and Attitude estimation | |
| CN116778391A (en) | Multi-mode crop disease phenotype collaborative analysis model and device | |
| CN114882301B (en) | Self-supervision learning medical image identification method and device based on region of interest | |
| CN108053418A (en) | A kind of animal background modeling method and device | |
| CN115019386A (en) | Exercise assistant training method based on deep learning | |
| CN107633527A (en) | Target tracking method and device based on full convolutional neural networks | |
| CN113536926A (en) | Human body action recognition method based on distance vector and multi-angle self-adaptive network | |
| CN114549516B (en) | Intelligent analysis system applied to multi-type high-density tiny insect body behaviourology | |
| CN116311521A (en) | Multitasking-oriented rat robot behavior analysis method | |
| CN115188051A (en) | Object behavior-based online course recommendation method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190607 |