CN110110689A - A kind of pedestrian's recognition methods again - Google Patents
A kind of pedestrian's recognition methods again Download PDFInfo
- Publication number
- CN110110689A CN110110689A CN201910403777.5A CN201910403777A CN110110689A CN 110110689 A CN110110689 A CN 110110689A CN 201910403777 A CN201910403777 A CN 201910403777A CN 110110689 A CN110110689 A CN 110110689A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- characteristic pattern
- channel
- pattern
- obtains
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- 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
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the present disclosure is related to a kind of pedestrian's recognition methods again comprising: it is extracted from multiple pictures and obtains pedestrian's CNN characteristic pattern；It is imitated by the way of confrontation erasing study and model training is carried out to the situation that the identification region of the pedestrian CNN characteristic pattern is blocked, obtain training pattern；Pedestrian is carried out using the training pattern combining target pedestrian image and pedestrian image to be identified to identify again, obtains pedestrian's weight recognition result.The method that the embodiment of the present disclosure provides provides a kind of feature rank data enhancing strategy, and the input feature vector figure of auxiliary classifier is partially wiped, and increases the variant of pedestrian's feature and resists the situation that pedestrian is blocked, improves the generalization ability of depth pedestrian weight identification model.
Description
Technical field
This disclosure relates to technical field of computer vision more particularly to a kind of pedestrian recognition methods again.
Background technique
It is that pedestrian's identity is matched and identified under non-overlapping multiple-camera monitoring system that pedestrian identifies again, in intelligence
Can video monitoring, crime prevention occurs, maintaining public order etc. plays an important role.However work as posture, gait, clothes
When the variation of the environmental factors such as human bodies attribute and illumination, background such as decorations, the appearance with a group traveling together is deposited under different monitor videos
In apparent difference, the appearance without same pedestrian under certain conditions can be more similar.
In recent years, the method for deep learning is widely used, and compared to traditional-handwork design method, deep learning can be taken
Obtain better performance.However, usually there are a large amount of network parameters in depth pedestrian weight identification model, it is but enterprising in limited data set
Row optimization, this just will increase the risk of over-fitting, reduce generalization ability.Therefore the generalization ability of model is improved for depth pedestrian
Identification is a significant and important problem again.
In order to improve the generalization ability of depth convolutional neural networks, the variant of training dataset can be increased and collected big
Amount is only able to achieve the data enhancing of image level comprising blocking the pedestrian image of situation, fail to provide image level it
Outer aspect carries out data enhancing, to improve the generalization ability of depth convolutional neural networks.
Drawbacks described above is that those skilled in the art's expectation overcomes.
Summary of the invention
(1) technical problems to be solved
In order to solve the above problem of the prior art, the disclosure provides a kind of pedestrian's recognition methods again, can be in feature
Grade further aspect carries out data enhancing, to improve the generalization ability of depth convolutional neural networks.
(2) technical solution
In order to achieve the above object, the main technical schemes of disclosure use include:
One embodiment of the disclosure provides a kind of pedestrian's recognition methods again comprising:
It is extracted from multiple pictures and obtains pedestrian's CNN characteristic pattern；
The situation being blocked to the identification region of the pedestrian CNN characteristic pattern is imitated by the way of confrontation erasing study
Model training is carried out, training pattern is obtained；
Pedestrian is carried out using the training pattern combining target pedestrian image and pedestrian image to be identified to identify again, is gone
People's weight recognition result.
In one embodiment of the disclosure, the extraction from multiple pictures obtains pedestrian's CNN characteristic pattern and includes:
It is concentrated from training data and randomly chooses the multiple picture；
Multiple and different semantic layers that the multiple picture is input to ResNet50 model are extracted, multiple channels are obtained
Characteristic pattern；
Notice that power module handles the characteristic pattern in the multiple channel using channel, obtains the feature handled through channel
Figure；
Using space transforms power module to it is described through channel processing characteristic pattern different location spatial context information
It is handled, obtains the pedestrian CNN characteristic pattern.
It is described to notice that power module carries out the characteristic pattern in the multiple channel using channel in one embodiment of the disclosure
Processing, obtain through channel handle characteristic pattern include:
According to the characteristic pattern in each channel in the characteristic pattern in the multiple channel, channel characteristics description is obtained；
To channel characteristics description by activation primitive operation, channel attention characteristic pattern is obtained；
The characteristic pattern that the channel attention characteristic pattern is polymerize with the characteristic pattern is multiplied, and obtains described handling through channel
Characteristic pattern.
In one embodiment of the disclosure, the Feature Descriptor includes the statistical value in the multiple channel, the feature
Description are as follows:
The statistical value in each channel are as follows:
Wherein N is the quantity in channel, and n is the number in channel, and A and B are respectively the length and width of the characteristic pattern；
The channel attention characteristic pattern are as follows:
E=σ (W2δ(W1(s)))
Wherein σ, δ respectively represent Sigmod activation primitive and ReLU activation primitive,It is the first full articulamentum Fc1
Weight,It is the weight of the second full articulamentum Fc2, r is the multiple of decaying.
It is described to be existed using space transforms power module to the characteristic pattern through channel processing in one embodiment of the disclosure
The spatial context information of different location is handled, and is obtained the pedestrian CNN characteristic pattern and is included:
To it is described through channel processing characteristic pattern carry out 1 × 1 convolution algorithm, obtain the first spatial information characteristic pattern T and
Second space information characteristics figure U；
The transposition of the first spatial information characteristic pattern T and the second space information characteristics figure U are subjected to matrix multiplication
Operation obtains spatial attention characteristic pattern；
The convolution algorithm that 1 × 1 is carried out to the characteristic pattern through channel processing, obtains third spatial information characteristic pattern V；
The transposition of the third spatial information characteristic pattern V and the spatial attention characteristic pattern is subjected to matrix multiplication fortune
It calculates, obtains spatially processed characteristic pattern；
The pedestrian CNN characteristic pattern is obtained according to channel processing and the spatial manipulation.
In one embodiment of the disclosure, it is described imitated by the way of confrontation erasing study it is special to the pedestrian CNN
The situation that the identification region of sign figure is blocked carries out model training, and obtaining training pattern includes:
The pedestrian CNN characteristic pattern is separately input to Main classification device and auxiliary classifier carries out classification based training, from the master
Classifier and the auxiliary classifier export the exclusive characteristic pattern of pedestrian's classification；
Part erasing, characteristic pattern after being wiped are carried out in the auxiliary classifier；
Described in the characteristic pattern exclusive to pedestrian's classification of Main classification device output and the auxiliary classifier output
Characteristic pattern after erasing, is calculated by loss function respectively, obtains penalty values；
Parameter update is carried out to the training pattern according to the penalty values.
In one embodiment of the disclosure, the Main classification device and the auxiliary classifier include identical quantity convolutional layer and
The average pond layer of the overall situation, and the number in the channel of the convolutional layer concentrates the number of pedestrian's classification identical with the training data,
Each channel of the exclusive characteristic pattern of pedestrian's classification represents the response temperature figure of body when pedestrian image belongs to a different category.
It is described to include: in the auxiliary classifier progress part erasing in one embodiment of the disclosure
The region that the body responds the confrontation erasing threshold value that temperature figure numerical value in temperature figure is higher than setting is determined as sentencing
Other property region；
The portion in the identification region is corresponded in the characteristic pattern exclusive to pedestrian's classification of the auxiliary classifier output
Divide and is wiped free of by the confrontation mode that response is substituted by 0.
It is described to utilize the training pattern combining target pedestrian image and pedestrian to be identified in one embodiment of the disclosure
Image carries out pedestrian and identifies again, obtains pedestrian's weight recognition result and includes:
It is input in the training pattern and is trained according to the target pedestrian image and the pedestrian image to be identified,
Respectively obtain corresponding depth characteristic；
Cosine is calculated according to the depth characteristic of the depth characteristic of the target pedestrian image and the pedestrian image to be identified
Distance；
It is determined according to the size of COS distance similar between the target pedestrian image and the pedestrian image to be identified
Degree, wherein the maximum pedestrian image to be identified of similarity is pedestrian weight recognition result.
In one embodiment of the disclosure, schemed according to the depth characteristic of the target pedestrian image and the pedestrian to be identified
The depth characteristic of picture calculates the calculation formula of COS distance are as follows:
Wherein feat1 is the depth characteristic of the target pedestrian image, and feat2 is the depth of the pedestrian image to be identified
Feature.
(3) beneficial effect
The beneficial effect of the disclosure is: pedestrian's recognition methods again that the embodiment of the present disclosure provides, by providing a kind of feature
Rank data enhancing strategy, the input feature vector figure of auxiliary classifier are partially wiped, and are increased the variant of pedestrian's feature and are resisted pedestrian
The situation being blocked improves the generalization ability of depth pedestrian weight identification model.
Detailed description of the invention
Fig. 1 is a kind of flow chart of pedestrian for providing of an embodiment of the present disclosure recognition methods again；
Fig. 2 is the schematic network structure that Fig. 1 the method is realized in an embodiment of the present disclosure；
Fig. 3 is the flow chart of step S110 in an embodiment of the present disclosure Fig. 1；
Fig. 4 is the flow chart of step S303 in an embodiment of the present disclosure Fig. 3；
Fig. 5 is channel attention schematic diagram in an embodiment of the present disclosure；
Fig. 6 is spatial attention schematic diagram in an embodiment of the present disclosure；
Fig. 7 is the flow chart of step S304 in an embodiment of the present disclosure Fig. 3；
Fig. 8 is confrontation erasing study schematic diagram in one embodiment of the disclosure；
Fig. 9 is the flow chart of step S120 in an embodiment of the present disclosure Fig. 1；
Figure 10 is the flow chart of step S130 in an embodiment of the present disclosure Fig. 1.
Specific embodiment
In order to preferably explain the disclosure, in order to understand, with reference to the accompanying drawing, by specific embodiment, to this public affairs
It opens and is described in detail.
All technical and scientific terms used herein and the those skilled in the art for belonging to the disclosure are usual
The meaning of understanding is identical.Description specific embodiment is intended merely in the term used in the description of the disclosure herein
Purpose, it is not intended that in the limitation disclosure.Term as used herein "and/or" includes one or more relevant listed items
Any and all combinations.
In the other embodiments of the disclosure, the variant for increasing training dataset is that raising depth convolutional neural networks are extensive
A kind of effective means of ability.However, different from target identification visual task, pedestrian identifies that needs collect the figure across camera again
As data, pedestrian's mark is also very difficult, identifies that one sufficiently large data set of building often needs high cost again for pedestrian
Investment, causes existing data set pedestrian mark amount small.In order to solve this problem, data enhancing can only be used only currently
Data set increases the variant of training set sample, without other costs.The research and utilization of nearest data enhancing is to antibiosis
Different human postures, camera style are generated at network (Generative Adversarial Networks, abbreviation GAN)
Pedestrian image, but this mode there are training time length, convergence difficulties, generate picture quality it is low the problems such as.In addition to explicitly
It generates outside new images, common method can also pass through shake pixel value, random cropping, overturning original image on training image
Operations are waited to enhance data.
In addition, blocking is also a key factor for influencing convolutional neural networks generalization ability.It collects largely comprising blocking
The pedestrian image of situation is that one kind can be in a manner of effectively solving occlusion issue, but this is also required to high cost input.It is another
The more reasonable method of kind is the situation accurately imitating pedestrian and being blocked.Example is said, random big using one on training image
Small, random site rectangle frame blocks training image, and replaces the pixel value of this rectangular area with random value, hides to imitate
It keeps off to increase the variant of data set.However above-mentioned occlusion area be it is randomly selected, optionally, training one pedestrian identify again divides
Then class model finds the region of image discriminating under the auxiliary of network visualization and multiple classifiers, and in original graph
Identification regional occlusion is generated into new samples on picture, new samples addition raw data set is finally carried out into re -training pedestrian and is identified again
Model.
Based on above-mentioned two methods, it is the variant for increasing sample by blocking in original pedestrian image, belongs to
The data enhancement methods of image level, and the application provides one kind.
Fig. 1 is a kind of flow chart of pedestrian for providing of an embodiment of the present disclosure recognition methods again, as shown in Figure 1, the party
Method the following steps are included:
As shown in Figure 1, in step s 110, being extracted from multiple pictures and obtaining pedestrian's CNN characteristic pattern；
As shown in Figure 1, in the step s 120, being imitated by the way of confrontation erasing study to the pedestrian CNN characteristic pattern
The situation that is blocked of identification region carry out model training, obtain training pattern；
As shown in Figure 1, in step s 130, being schemed using the training pattern combining target pedestrian image and pedestrian to be identified
It is identified again as carrying out pedestrian, obtains pedestrian's weight recognition result.
The specific implementation of each step of embodiment illustrated in fig. 1 is described in detail below:
Process as shown in connection with fig. 1, Fig. 2 are the network structure signal that Fig. 1 the method is realized in an embodiment of the present disclosure
Figure, as shown in Fig. 2, requiring to pay attention to using channel attention and the complementary of spatial attention during processing to each channel
Then power carries out confrontation erasing study and Softmax costing bio disturbance to obtained characteristic pattern.In addition, as shown in Fig. 2, therein
Three channels are divided into two channels for the semantic branch of middle rank and a high-level semantics branch.
In step s 110, it is extracted from multiple pictures and obtains pedestrian's CNN characteristic pattern.
Fig. 3 is the flow chart of step S110 in an embodiment of the present disclosure Fig. 1, specifically includes the following steps:
As shown in figure 3, being concentrated from training data in step S301 and randomly choosing the multiple picture.
As shown in figure 3, in step s 302, the multiple picture to be input to multiple and different semantemes of ResNet50 model
Layer extracts, and obtains the characteristic pattern in multiple channels.
In one embodiment of the present disclosure, firstly, input picture, randomly selects on training dataset and criticize in the step
The pictures for measuring number, that is, choose multiple pictures.Secondly, picture size is all adjusted to 384*128, and it is sent to backbone network
Network ResNet50 difference semantic layer (res_conv5a, res_conv5b, res_conv5c, as shown in Fig. 2, wherein res_
The semantic branch of the corresponding middle rank of conv5a, res_conv5b, res_conv5c correspond to high-level semantics branch) it is special to extract pedestrian CNN
Sign figure.
As shown in figure 3, in step S303, pay attention to power module to the characteristic pattern in the multiple channel using channel at
Reason obtains the characteristic pattern handled through channel.
Fig. 4 is the flow chart of step S303 in an embodiment of the present disclosure Fig. 3, specifically includes the following steps:
As shown in figure 4, in step S401, according to the characteristic pattern in each channel in the characteristic pattern in the multiple channel,
Obtain channel characteristics description.
As shown in figure 4, to channel characteristics description by activation primitive operation, obtaining channel in step S402
Attention characteristic pattern.
As shown in figure 4, in step S403, characteristic pattern that the channel attention characteristic pattern is polymerize with the characteristic pattern
It is multiplied, obtains the characteristic pattern handled through channel.
In one embodiment of the present disclosure, step S303 can notice that power module explores pedestrian CNN feature using channel
The connection of figure interchannel, the region of capture and description input picture identification.
Fig. 5 is channel attention schematic diagram in an embodiment of the present disclosure, as shown in figure 5, extracting for each channel
The characteristic pattern arrived, A and B are respectively the length and width of the characteristic pattern, and n is the number in channel, and N is the quantity in channel.
Firstly, using GAP operating polymerization characteristic patternThe spatial information in each channel generates the feature in channel
Description:As it can be seen that the Feature Descriptor includes the statistical value in the multiple channel, each channel
Statistical value are as follows:
Secondly, s is obtained channel attention characteristic pattern by a threshold mechanism module
E=σ (W2δ(W1(s))) formula (2)
Wherein σ, δ respectively represent Sigmod activation primitive and ReLU activation primitive,It is the first full articulamentum Fc1
Weight,It is the weight of the second full articulamentum Fc2, r is the multiple of decaying.
Finally, channel attention e is multiplied to obtain revised characteristic pattern with characteristic pattern S is originally inputted
Since attention characteristic pattern e coding in channel includes the dependence and relative importance between channel characteristics figure, nerve
Network will be updated e by dynamic and the characteristic pattern of study important kind is gone to ignore the characteristic pattern less reused.
As shown in figure 3, in step s 304, using space transforms power module to the characteristic pattern through channel processing not
Spatial context information with position is handled, and the pedestrian CNN characteristic pattern is obtained.
In one embodiment of the present disclosure, step S304 can use space transforms power module by characteristic pattern different location
Spatial context information incorporate pedestrian's local feature, enhance pedestrian's regional area spatial coherence.Fig. 6 is the disclosure one
Spatial attention schematic diagram in embodiment obtains as shown in fig. 6, carrying out convolution operation respectively to the characteristic pattern handled through channel
First spatial information characteristic pattern T, second space information characteristics figure U and third spatial information characteristic pattern V, T are multiplied after transposition with U
D is obtained, D is multiplied to obtain X with V, and X be added after a certain proportion of scaling with the characteristic pattern handled through channel, real
Now to the spatial manipulation of characteristic pattern, final pedestrian's CNN characteristic pattern is obtained.
Fig. 7 is the flow chart of step S304 in an embodiment of the present disclosure Fig. 3, specifically includes the following steps:
As shown in fig. 7, in step s 701,1 × 1 convolution algorithm being carried out to the characteristic pattern through channel processing, is obtained
To the first spatial information characteristic pattern T and second space information characteristics figure U.Attention corrected characteristic pattern in channel is (i.e. through channel
The characteristic pattern of processing)It is sent into 1 × 1 convolution fkeyWith fquery, two characteristic patterns T and U are obtained, wherein
As shown in fig. 7, in step S702, by the transposition of the first spatial information characteristic pattern T and the second space
Information characteristics figure U carries out matrix multiplication operation, obtains spatial attention characteristic pattern.T and U-shape are adjusted toWherein Z
=A × B represents the quantity of feature, and T is carried out transposition later and carries out matrix multiplication with U, applies one according to line direction
Softmax function obtains spatial attention characteristic pattern D ∈ RZ×Z, each element d of Dj,iIt can indicate are as follows:
Wherein, dj,iI-th of position is represented to the correlation of j-th of position feature, the feature representation of two positions gets over phase
Seemingly, correlation between the two is higher.
As shown in fig. 7, carrying out 1 × 1 convolution algorithm to the characteristic pattern through channel processing in step S703, obtaining
To third spatial information characteristic pattern V.
The characteristic pattern S' handled through channel is sent into 1 × 1 convolutional layer fvalue, obtain a new characteristic patternAnd it is by its Adjusting Shape
As shown in fig. 7, in step S704, by the third spatial information characteristic pattern V and the spatial attention feature
The transposition of figure carries out matrix multiplication operation, obtains spatially processed characteristic pattern.
The transposition of V and D is subjected to matrix multiplication first in the step, is by the Adjusting Shape of resultAnd by its
Pass through 1 × 1 convolution fup, obtain characteristic pattern
As shown in fig. 7, obtaining the pedestrian CNN according to channel processing and the spatial manipulation in step S705
Characteristic pattern, channel processing here is above-mentioned steps S401~S403, spatial manipulation, that is, above-mentioned steps S701~S704.
X is multiplied with zooming parameter α in the step, and is added to obtain according to element with the characteristic pattern S' handled through channel
Characteristic patternThat is:
Based on above-mentioned, the element of each position of characteristic pattern S " can be indicated are as follows:
Wherein, α be can learning parameter, 0 is set as when initial, can gradually be learnt since 0 to bigger weight.From
Formula (6) can see, the feature S " of each position of characteristic pattern S "jThe feature for being all positions and the spy through channel processing
Sign figure S'jWeighted sum, therefore it includes global receptive field, can be according to each element d in space transforms characteristic pattern Dj,i
Size, selectively polymerize relevant regional area ViSpatial context information, so as to enhance the different part of pedestrian
Connection between feature.
Based on abovementioned steps, channel attention and spatial attention are used in series amendment CNN characteristic pattern, allow neural network
Which which it is characterized in more effectively to focus on type feature and position automatically.Therefore, channel attention in the disclosure
Two modules of module and space transforms power module are used in combination, and the effect of the two is given full play to.As shown in figure 5, the disclosure is special
Sign figureIt first passes through channel and pays attention to power module, then by space transforms power module, that is, realize the amendment of complementary attention:
S'=Mc(S)
S "=Ms(S') formula (7)
In the step s 120, the identification area to the pedestrian CNN characteristic pattern is imitated by the way of confrontation erasing study
The situation that domain is blocked carries out model training, obtains training pattern.
Fig. 8 is confrontation erasing study schematic diagram in one embodiment of the disclosure, as shown in figure 8, for pedestrian's CNN characteristic pattern point
It Tong Guo not the processing such as Main classification device and auxiliary classifier progress convolution, GAP and Softmax loss function.
Fig. 9 is the flow chart of step S120 in an embodiment of the present disclosure Fig. 1, specifically includes the following steps:
As shown in figure 9, the pedestrian CNN characteristic pattern is separately input to Main classification device and auxiliary classification in step S901
Device carries out classification based training, the characteristic pattern exclusive from the Main classification device and auxiliary classifier output pedestrian's classification.
Wherein convolutional layer and global average pond layer of the Main classification device with the auxiliary classifier comprising identical quantity
(Global Average Pooling, abbreviation GAP), and the number in the channel of the convolutional layer and the training data are concentrated and are gone
The other number of the mankind is identical, when each channel of the exclusive characteristic pattern of pedestrian's classification represents pedestrian image and belongs to a different category
Body respond temperature figure.
In the step into, the full articulamentum of disaggregated model is changed to 1 × 1 convolutional layer, constitutes point based on full convolutional network
The convolutional layer that revised characteristic pattern (i.e. pedestrian CNN characteristic pattern) is sent into 1 × 1 can be directly obtained pedestrian's classification by class model
Exclusive characteristic pattern, since the number of active lanes of convolutional layer is pedestrian's class number in training set, each channel generation of characteristic pattern
Body when the table pedestrian image belongs to a different category responds temperature figure.In the training stage, pedestrian image classification mark can be obtained
The characteristic pattern in channel corresponding to class label is indexed out, obtains the exclusive characteristic pattern of pedestrian's classification, the i.e. pedestrian image by label
Body respond temperature figure.
As shown in figure 9, carrying out part erasing, characteristic pattern after being wiped in the auxiliary classifier in step S902.
Firstly, the region that the body responds the confrontation erasing threshold value that temperature figure numerical value in temperature figure is higher than setting is determined
For identification region；Secondly, corresponding to the differentiation in the characteristic pattern exclusive to pedestrian's classification of the auxiliary classifier output
The part in property region is wiped free of by the confrontation mode that response is substituted by 0.
The input feature vector figure for wiping auxiliary classifier in the step by part generates pedestrian by step S901 Main classification device
The exclusive characteristic pattern of classification further responds temperature figure numerical value according to body and is set to sentence higher than the position of confrontation erasing threshold value
Other property region, and corresponding region can be substituted by 0 confrontation mode by response in the input feature vector figure of auxiliary classifier
And it is wiped free of.The characteristic pattern of auxiliary classifier input is partially wiped, and can increase the variant of characteristic pattern in this way, while imitating row
The case where people is blocked.
As shown in figure 9, in step S903, the characteristic pattern exclusive to pedestrian's classification of Main classification device output with
Characteristic pattern after the erasing of the auxiliary classifier output, is calculated by loss function respectively, obtains penalty values.
As shown in figure 9, carrying out parameter update to the training pattern according to the penalty values in step S904.
In this step, for Main classification device and auxiliary classifier Liang Ge branch under the supervision of Softmax loss function
Carry out parameter update, loss function expression formula are as follows:
Wherein, P represents batch size, and M represents numbers of branches, and K represents the number of classifier in confrontation erasing study
Amount (being 2 in the present embodiment), C represents class number,When representing using full convolution sorter network, the m of p-th of sample
The l of k-th of classifier of a branchpThe nodal value of a Softmax input, wherein lpIt is the classification of p-th of sample.Each
First classifier of branch is all Main classification device, and second classifier is auxiliary classifier, parameter lambdakIt is allocated to the two points
The weight of class device loss, wherein parameter lambda1=1 it is corresponding be Main classification device, parameter lambda2=0.5 it is corresponding be auxiliary classifier.
In step s 130, pedestrian is carried out using the training pattern combining target pedestrian image and pedestrian image to be identified
It identifies again, obtains pedestrian's weight recognition result.
Figure 10 is the flow chart of step S130 in an embodiment of the present disclosure Fig. 1, specifically includes the following steps:
As shown in Figure 10, it in step S1001, is inputted according to the target pedestrian image and the pedestrian image to be identified
It is trained into the training pattern, respectively obtains corresponding depth characteristic.In the step by target pedestrian image with wait know
Other pedestrian image is sent into the trained CNN model extraction characteristics of image of step 2, specifically, by the spy of semantic levels different in Fig. 2
(res_conv5a, res_conv5b, res_conv5c) series connection is levied as final Feature Descriptor.
As shown in Figure 10, in step S1002, according to the depth characteristic of the target pedestrian image and the row to be identified
The depth characteristic of people's image calculates COS distance, calculation formula are as follows:
Wherein feat1 is the depth characteristic of the target pedestrian image, and feat2 is the depth of the pedestrian image to be identified
Feature.
As shown in Figure 10, in the step s 1003, the target pedestrian image and described is determined according to the size of COS distance
Similarity between pedestrian image to be identified, wherein the maximum pedestrian image to be identified of similarity is that the pedestrian identifies knot again
Fruit.
Similarity and feature cosine between the figure formed due to target pedestrian image and pedestrian image to be identified is opposite away from
From the relationship for being in negative linear correlation, therefore, feature COS distance is smaller, and it is higher to scheme opposite similarity.Based on above-mentioned, can obtain
Ascending order arrangement is carried out after to COS distance by size, i.e. image carries out descending sort to similarity size, and similarity is maximum
The result that pedestrian image to be identified identifies again as pedestrian.
In conclusion the pedestrian's recognition methods again provided using the embodiment of the present disclosure, on the one hand, by providing a kind of feature
Rank data enhancing strategy, the input feature vector figure of auxiliary classifier are partially wiped, and are increased the variant of pedestrian's feature and are resisted pedestrian
The situation being blocked improves the generalization ability of depth pedestrian weight identification model.On the other hand, the spatial attention mould in the disclosure
Spatial context information is incorporated pedestrian's local feature by type, enhances the spatial coherence of pedestrian's different location, while infusing with channel
Power model of anticipating constitutes complementary attention model, and the two is used in combination, and corrects characteristic pattern, Ke Yigeng from channel and space both direction
Identification region is captured well.The disclosure proposes a kind of disaggregated model based on full convolutional network, can be in the mistake of propagated forward
Body response temperature figure is directly obtained in journey, guidance erasing identification body region realizes the data enhancing of feature rank data.
It should be noted that although being referred to several modules or list for acting the equipment executed in the above detailed description
Member, but this division is not enforceable.In fact, according to embodiment of the present disclosure, it is above-described two or more
Module or the feature and function of unit can embody in a module or unit.Conversely, an above-described mould
The feature and function of block or unit can be to be embodied by multiple modules or unit with further division.
Through the above description of the embodiments, those skilled in the art is it can be readily appreciated that example described herein is implemented
Mode can also be realized by software realization in such a way that software is in conjunction with necessary hardware.Therefore, according to the disclosure
The technical solution of embodiment can be embodied in the form of software products, which can store non-volatile at one
Property storage medium (can be CD-ROM, USB flash disk, mobile hard disk etc.) in or network on, including some instructions are so that a calculating
Equipment (can be personal computer, server, touch control terminal or network equipment etc.) is executed according to disclosure embodiment
Method.
Those skilled in the art after considering the specification and implementing the invention disclosed here, will readily occur to its of the disclosure
Its embodiment.This application is intended to cover any variations, uses, or adaptations of the disclosure, these modifications, purposes or
Person's adaptive change follows the general principles of this disclosure and including the undocumented common knowledge in the art of the disclosure
Or conventional techniques.The description and examples are only to be considered as illustrative, and the true scope and spirit of the disclosure are by following
Claim is pointed out.
It should be understood that the present disclosure is not limited to the precise structures that have been described above and shown in the drawings, and
And various modifications and changes may be made without departing from the scope thereof.The scope of the present disclosure is only limited by the accompanying claims.
Claims (10)
1. a kind of pedestrian recognition methods again, characterized in that it comprises:
It is extracted from multiple pictures and obtains pedestrian's CNN characteristic pattern；
It is imitated by the way of confrontation erasing study and the situation that the identification region of the pedestrian CNN characteristic pattern is blocked is carried out
Model training obtains training pattern；
Pedestrian is carried out using the training pattern combining target pedestrian image and pedestrian image to be identified to identify again, obtains pedestrian's weight
Recognition result.
2. pedestrian as described in claim 1 recognition methods again, which is characterized in that described extract from multiple pictures obtains pedestrian
CNN characteristic pattern includes:
It is concentrated from training data and randomly chooses the multiple picture；
Multiple and different semantic layers that the multiple picture is input to ResNet50 model are extracted, the spy in multiple channels is obtained
Sign figure；
Notice that power module handles the characteristic pattern in the multiple channel using channel, obtains the characteristic pattern handled through channel；
The characteristic pattern through channel processing is carried out in the spatial context information of different location using space transforms power module
Processing, obtains the pedestrian CNN characteristic pattern.
3. pedestrian as claimed in claim 2 recognition methods again, which is characterized in that described to pay attention to power module to described using channel
The characteristic pattern in multiple channels is handled, obtain through channel handle characteristic pattern include:
According to the characteristic pattern in each channel in the characteristic pattern in the multiple channel, channel characteristics description is obtained；
To channel characteristics description by activation primitive operation, channel attention characteristic pattern is obtained；
The characteristic pattern that the channel attention characteristic pattern is polymerize with the characteristic pattern is multiplied, and obtains the spy handled through channel
Sign figure.
4. pedestrian as claimed in claim 3 recognition methods again, which is characterized in that the Feature Descriptor includes the multiple logical
The statistical value in road, the Feature Descriptor are as follows:
The statistical value in each channel are as follows:
Wherein N is the quantity in channel, and n is the number in channel, and A and B are respectively the length and width of the characteristic pattern；
The channel attention characteristic pattern are as follows:
E=σ (W2δ(W1(s)))
Wherein σ, δ respectively represent Sigmod activation primitive and ReLU activation primitive,It is the power of the first full articulamentum Fc1
Weight,It is the weight of the second full articulamentum Fc2, r is the multiple of decaying.
5. pedestrian as claimed in claim 2 recognition methods again, which is characterized in that described to utilize space transforms power module to described
The characteristic pattern handled through channel is handled in the spatial context information of different location, obtains the pedestrian CNN characteristic pattern packet
It includes:
The convolution algorithm that 1 × 1 is carried out to the characteristic pattern through channel processing, obtains the first spatial information characteristic pattern T and second
Spatial information characteristic pattern U；
The transposition of the first spatial information characteristic pattern T and the second space information characteristics figure U are subjected to matrix multiplication operation,
Obtain spatial attention characteristic pattern；
The convolution algorithm that 1 × 1 is carried out to the characteristic pattern through channel processing, obtains third spatial information characteristic pattern V；
The transposition of the third spatial information characteristic pattern V and the spatial attention characteristic pattern is subjected to matrix multiplication operation, is obtained
To spatially processed characteristic pattern；
The pedestrian CNN characteristic pattern is obtained according to channel processing and the spatial manipulation.
6. pedestrian as claimed in claim 2 recognition methods again, which is characterized in that the mould by the way of confrontation erasing study
The imitative situation being blocked to the identification region of the pedestrian CNN characteristic pattern carries out model training, and obtaining training pattern includes:
The pedestrian CNN characteristic pattern is separately input to Main classification device and auxiliary classifier carries out classification based training, from the Main classification
Device and the auxiliary classifier export the exclusive characteristic pattern of pedestrian's classification；
Part erasing, characteristic pattern after being wiped are carried out in the auxiliary classifier；
The erasing of the characteristic pattern and the auxiliary classifier output exclusive to pedestrian's classification of Main classification device output
Characteristic pattern afterwards is calculated by loss function respectively, obtains penalty values；
Parameter update is carried out to the training pattern according to the penalty values.
7. pedestrian as claimed in claim 6 recognition methods again, which is characterized in that the Main classification device and the auxiliary classifier packet
Convolutional layer and global average pond layer containing identical quantity, and the number in the channel of the convolutional layer and the training data are concentrated
The number of pedestrian's classification is identical, and each channel of the exclusive characteristic pattern of pedestrian's classification represents pedestrian image and belongs to a different category
When body respond temperature figure.
8. pedestrian as claimed in claim 7 recognition methods again, which is characterized in that described to carry out part wiping in the auxiliary classifier
Except including:
The region that the body responds the confrontation erasing threshold value that temperature figure numerical value in temperature figure is higher than setting is determined as identification
Region；
The part that the identification region is corresponded in the characteristic pattern exclusive to pedestrian's classification of the auxiliary classifier output is logical
It crosses response and is substituted by 0 confrontation mode and be wiped free of.
9. pedestrian as claimed in claim 3 recognition methods again, which is characterized in that described to utilize the training pattern combining target
Pedestrian image and pedestrian image to be identified carry out pedestrian and identify again, obtain pedestrian's weight recognition result and include:
It is input in the training pattern and is trained according to the target pedestrian image and the pedestrian image to be identified, respectively
Obtain corresponding depth characteristic；
COS distance is calculated according to the depth characteristic of the depth characteristic of the target pedestrian image and the pedestrian image to be identified；
The similarity between the target pedestrian image and the pedestrian image to be identified is determined according to the size of COS distance,
The middle maximum pedestrian image to be identified of similarity is pedestrian weight recognition result.
10. pedestrian as claimed in claim 9 recognition methods again, which is characterized in that according to the depth of the target pedestrian image
The depth characteristic of feature and the pedestrian image to be identified calculates the calculation formula of COS distance are as follows:
Wherein feat1 is the depth characteristic of the target pedestrian image, and feat2 is that the depth of the pedestrian image to be identified is special
Sign.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910403777.5A CN110110689B (en) | 2019-05-15 | 2019-05-15 | Pedestrian re-identification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910403777.5A CN110110689B (en) | 2019-05-15 | 2019-05-15 | Pedestrian re-identification method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110110689A true CN110110689A (en) | 2019-08-09 |
| CN110110689B CN110110689B (en) | 2023-05-26 |
Family
ID=67490255
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910403777.5A Active CN110110689B (en) | 2019-05-15 | 2019-05-15 | Pedestrian re-identification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110110689B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110516603A (en) * | 2019-08-28 | 2019-11-29 | 北京百度网讯科技有限公司 | Information processing method and device |
| CN111160096A (en) * | 2019-11-26 | 2020-05-15 | 北京海益同展信息科技有限公司 | Method, device and system for identifying poultry egg abnormality, storage medium and electronic device |
| CN111198964A (en) * | 2020-01-10 | 2020-05-26 | 中国科学院自动化研究所 | Image retrieval method and system |
| CN111461038A (en) * | 2020-04-07 | 2020-07-28 | 中北大学 | Pedestrian re-identification method based on layered multi-mode attention mechanism |
| CN111582587A (en) * | 2020-05-11 | 2020-08-25 | 深圳赋乐科技有限公司 | Prediction method and prediction system for video public sentiment |
| CN111814618A (en) * | 2020-06-28 | 2020-10-23 | 浙江大华技术股份有限公司 | Pedestrian re-identification method, gait identification network training method and related device |
| CN112131943A (en) * | 2020-08-20 | 2020-12-25 | 深圳大学 | Video behavior identification method and system based on dual attention model |
| WO2021057186A1 (en) * | 2019-09-24 | 2021-04-01 | 华为技术有限公司 | Neural network training method, data processing method, and related apparatuses |
| CN112784648A (en) * | 2019-11-07 | 2021-05-11 | 中国科学技术大学 | Method and device for optimizing feature extraction of pedestrian re-identification system of video |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| IT201600068348A1 (en) * | 2016-07-01 | 2018-01-01 | Octo Telematics Spa | Procedure for determining the status of a vehicle by detecting the vehicle's battery voltage. |
| CN107563313A (en) * | 2017-08-18 | 2018-01-09 | 北京航空航天大学 | Multiple target pedestrian detection and tracking based on deep learning |
| CN107679483A (en) * | 2017-09-27 | 2018-02-09 | 北京小米移动软件有限公司 | Number plate recognition methods and device |
| CN107992882A (en) * | 2017-11-20 | 2018-05-04 | 电子科技大学 | A kind of occupancy statistical method based on WiFi channel condition informations and support vector machines |
| WO2018153322A1 (en) * | 2017-02-23 | 2018-08-30 | 北京市商汤科技开发有限公司 | Key point detection method, neural network training method, apparatus and electronic device |
| CN109359559A (en) * | 2018-09-27 | 2019-02-19 | 天津师范大学 | A kind of recognition methods again of the pedestrian based on dynamic barriers sample |
| CN109583379A (en) * | 2018-11-30 | 2019-04-05 | 常州大学 | A kind of pedestrian's recognition methods again being aligned network based on selective erasing pedestrian |
| CN109583502A (en) * | 2018-11-30 | 2019-04-05 | 天津师范大学 | A kind of pedestrian's recognition methods again based on confrontation erasing attention mechanism |
-
2019
- 2019-05-15 CN CN201910403777.5A patent/CN110110689B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| IT201600068348A1 (en) * | 2016-07-01 | 2018-01-01 | Octo Telematics Spa | Procedure for determining the status of a vehicle by detecting the vehicle's battery voltage. |
| WO2018153322A1 (en) * | 2017-02-23 | 2018-08-30 | 北京市商汤科技开发有限公司 | Key point detection method, neural network training method, apparatus and electronic device |
| CN107563313A (en) * | 2017-08-18 | 2018-01-09 | 北京航空航天大学 | Multiple target pedestrian detection and tracking based on deep learning |
| CN107679483A (en) * | 2017-09-27 | 2018-02-09 | 北京小米移动软件有限公司 | Number plate recognition methods and device |
| CN107992882A (en) * | 2017-11-20 | 2018-05-04 | 电子科技大学 | A kind of occupancy statistical method based on WiFi channel condition informations and support vector machines |
| CN109359559A (en) * | 2018-09-27 | 2019-02-19 | 天津师范大学 | A kind of recognition methods again of the pedestrian based on dynamic barriers sample |
| CN109583379A (en) * | 2018-11-30 | 2019-04-05 | 常州大学 | A kind of pedestrian's recognition methods again being aligned network based on selective erasing pedestrian |
| CN109583502A (en) * | 2018-11-30 | 2019-04-05 | 天津师范大学 | A kind of pedestrian's recognition methods again based on confrontation erasing attention mechanism |
Non-Patent Citations (1)
| Title |
|---|
| 张文文 等: "基于稀疏学习的行人重识别算法", 《数据采集与处理》 * |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110516603A (en) * | 2019-08-28 | 2019-11-29 | 北京百度网讯科技有限公司 | Information processing method and device |
| CN110516603B (en) * | 2019-08-28 | 2022-03-18 | 北京百度网讯科技有限公司 | Information processing method and device |
| WO2021057186A1 (en) * | 2019-09-24 | 2021-04-01 | 华为技术有限公司 | Neural network training method, data processing method, and related apparatuses |
| CN112784648B (en) * | 2019-11-07 | 2022-09-06 | 中国科学技术大学 | Method and device for optimizing feature extraction of pedestrian re-identification system of video |
| CN112784648A (en) * | 2019-11-07 | 2021-05-11 | 中国科学技术大学 | Method and device for optimizing feature extraction of pedestrian re-identification system of video |
| WO2021104125A1 (en) * | 2019-11-26 | 2021-06-03 | 京东数科海益信息科技有限公司 | Abnormal egg identification method, device and system, storage medium, and electronic device |
| CN111160096A (en) * | 2019-11-26 | 2020-05-15 | 北京海益同展信息科技有限公司 | Method, device and system for identifying poultry egg abnormality, storage medium and electronic device |
| CN111198964A (en) * | 2020-01-10 | 2020-05-26 | 中国科学院自动化研究所 | Image retrieval method and system |
| CN111198964B (en) * | 2020-01-10 | 2023-04-25 | 中国科学院自动化研究所 | Image retrieval method and system |
| CN111461038A (en) * | 2020-04-07 | 2020-07-28 | 中北大学 | Pedestrian re-identification method based on layered multi-mode attention mechanism |
| CN111582587A (en) * | 2020-05-11 | 2020-08-25 | 深圳赋乐科技有限公司 | Prediction method and prediction system for video public sentiment |
| CN111582587B (en) * | 2020-05-11 | 2021-06-04 | 深圳赋乐科技有限公司 | Prediction method and prediction system for video public sentiment |
| CN111814618A (en) * | 2020-06-28 | 2020-10-23 | 浙江大华技术股份有限公司 | Pedestrian re-identification method, gait identification network training method and related device |
| CN111814618B (en) * | 2020-06-28 | 2023-09-01 | 浙江大华技术股份有限公司 | Pedestrian re-recognition method, gait recognition network training method and related devices |
| CN112131943A (en) * | 2020-08-20 | 2020-12-25 | 深圳大学 | Video behavior identification method and system based on dual attention model |
| CN112131943B (en) * | 2020-08-20 | 2023-07-11 | 深圳大学 | Dual-attention model-based video behavior recognition method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110110689B (en) | 2023-05-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110110689A (en) | A kind of pedestrian's recognition methods again | |
| CN109740419B (en) | Attention-LSTM network-based video behavior identification method | |
| CN110458844B (en) | Semantic segmentation method for low-illumination scene | |
| CN111898617B (en) | Target detection method and system based on attention mechanism and parallel void convolution network | |
| CN110008842A (en) | A kind of pedestrian's recognition methods again for more losing Fusion Model based on depth | |
| CN107408211A (en) | Method for distinguishing is known again for object | |
| CN109815867A (en) | A kind of crowd density estimation and people flow rate statistical method | |
| CN111310773A (en) | Efficient license plate positioning method of convolutional neural network | |
| CN109657715B (en) | Semantic segmentation method, device, equipment and medium | |
| CN108052884A (en) | A kind of gesture identification method based on improvement residual error neutral net | |
| CN108182447A (en) | A kind of adaptive particle filter method for tracking target based on deep learning | |
| CN109410168A (en) | For determining the modeling method of the convolutional neural networks model of the classification of the subgraph block in image | |
| CN106778852A (en) | A kind of picture material recognition methods for correcting erroneous judgement | |
| CN114782977B (en) | Pedestrian re-recognition guiding method based on topology information and affinity information | |
| CN113963032A (en) | Twin network structure target tracking method fusing target re-identification | |
| WO2022007867A1 (en) | Method and device for constructing neural network | |
| CN110889375A (en) | Hidden and double-flow cooperative learning network and method for behavior recognition | |
| CN108280436A (en) | A kind of action identification method based on the multi-grade remnant network for stacking recursive unit | |
| CN110175597A (en) | Video target detection method integrating feature propagation and aggregation | |
| CN113763417B (en) | Target tracking method based on twin network and residual error structure | |
| CN113052254A (en) | Multi-attention ghost residual fusion classification model and classification method thereof | |
| CN114419732A (en) | HRNet human body posture identification method based on attention mechanism optimization | |
| CN110532959A (en) | Real-time act of violence detection system based on binary channels Three dimensional convolution neural network | |
| CN116434010A (en) | Multi-view pedestrian attribute identification method | |
| CN114492634A (en) | Fine-grained equipment image classification and identification 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |