CN109886242A - A kind of method and system that pedestrian identifies again - Google Patents
A kind of method and system that pedestrian identifies again Download PDFInfo
- Publication number
- CN109886242A CN109886242A CN201910156179.2A CN201910156179A CN109886242A CN 109886242 A CN109886242 A CN 109886242A CN 201910156179 A CN201910156179 A CN 201910156179A CN 109886242 A CN109886242 A CN 109886242A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- feature
- parameter
- image
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 31
- 238000000605 extraction Methods 0.000 claims abstract description 38
- 238000011282 treatment Methods 0.000 claims abstract description 20
- 230000004913 activation Effects 0.000 claims description 22
- 239000013598 vector Substances 0.000 claims description 19
- 238000012360 testing method Methods 0.000 claims description 9
- 238000012549 training Methods 0.000 claims description 7
- 238000001514 detection method Methods 0.000 claims description 6
- 230000014509 gene expression Effects 0.000 claims description 5
- 238000004590 computer program Methods 0.000 claims description 2
- 238000005192 partition Methods 0.000 abstract description 6
- 239000000284 extract Substances 0.000 abstract description 5
- 238000004364 calculation method Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 6
- 238000012545 processing Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 210000001015 abdomen Anatomy 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 241001269238 Data Species 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000003313 weakening effect Effects 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000003825 pressing Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Abstract
The present invention provides a kind of pedestrian recognition methods and system again, and wherein recognition methods includes obtaining multiple pedestrian images to pedestrian again, constructs sample data set: being trained to the sample data set, constructs pedestrian's feature extraction network；Feature extraction pretreatment is carried out to pedestrian image using pedestrian's feature extraction network, the fisrt feature figure with multiple features channel is obtained, multidomain treat-ment is carried out to the fisrt feature figure, obtains second feature figure, and feature extraction is carried out to the second feature figure, obtain standard feature parameter；It obtains pedestrian image to be measured: extracting the feature of the pedestrian image to be measured, obtain the characteristic parameter of the pedestrian image to be measured；The similarity of the characteristic parameter and the standard feature parameter that calculate the pedestrian image to be measured obtains similarity parameter；Pedestrian is completed according to the similarity parameter to identify again；In this way, the feature that the partition characteristics extracting mode being weighted extracts more has identification, to improve the discriminating power of depth model.
Description
Technical field
The present invention relates to one mode identification technology fields, more particularly to a kind of method and system that pedestrian identifies again.
Background technique
In recent years, as public safety of the people to society is increasingly paid close attention to, video monitoring system is largely popularized.Such as machine
The public places such as field, railway station, campus and office block are all needed to monitor, be escort for security protection.Monitoring in face of magnanimity regards
Frequency evidence, a large amount of manpower demand put into the monitoring and retrieval of video information, and the efficiency of this mode is not only low, also make
At extra resource waste.If computer vision analysis technology can be utilized, automatically-monitored and analysis video information is inevitable
The construction of " safe city " can greatly be accelerated.
It is task crucial in the research of computer vision that pedestrian identifies again.In general, one about pedestrian is given
Perhaps one section of video pedestrian identifies exactly in the picture or video that the camera of not overlapping region takes picture again,
The process that the same person is identified.Although relevant research is increasingly taken seriously, the accuracy rate that pedestrian identifies again is also
It is improved much, but still needs to solve there are many difficulty.Since pedestrian's picture to be identified is shot from original picture in different
The difference of camera, equipment can bring error to image-forming condition；Environment under different scenes is inconsistent, and the data of acquisition can also produce
Raw different deviation；And the change meeting of illumination is so that the performance of same color is different；Importantly, pedestrian is under camera
Attitudes vibration and occlusion issue, all make the discrimination difficulty to the same person quite big.
In recent years, due to the tide of deep learning, convolutional neural networks are widely used in pedestrian and identify field again, pass through
Depth network extracts characteristics of image, and is carried out using deep learning or conventional method apart from degree on corresponding feature space
Amount, substantially increases the accuracy rate that pedestrian identifies again.The progress of these work all has benefited from depth convolutional network model in feature
Ability in extraction, but given feature space is confined in the exploration of discriminating power, hence limit depth model
The raising of discriminating power.
Summary of the invention
In view of the foregoing deficiencies of prior art, the purpose of the present invention is to provide a kind of pedestrian again recognition methods and it is
System, for solving the problems, such as that accuracy of identification is not high enough in the prior art.
In order to achieve the above objects and other related objects, the present invention provides a kind pedestrian recognition methods again, comprising: obtains
Multiple pedestrian images construct sample data set；The sample data set is trained, pedestrian's feature extraction network is constructed；It adopts
Feature extraction pretreatment is carried out to the pedestrian image with pedestrian's feature extraction network, obtains with multiple features channel
One characteristic pattern carries out multidomain treat-ment to the fisrt feature figure, obtains second feature figure, and carries out to the second feature figure special
Sign is extracted, and standard feature parameter is obtained；Obtain pedestrian image to be measured；The feature for extracting the pedestrian image to be measured obtains described
The characteristic parameter of pedestrian image to be measured；The characteristic parameter for calculating the pedestrian image to be measured is similar to the standard feature parameter
Degree obtains similarity parameter；Pedestrian is completed according to the similarity parameter to identify again.
Optionally, the multiple pedestrian images of acquisition include: acquisition original image；The original image is detected,
It positions according to testing result and cuts out to obtain pedestrian area image；The pedestrian area image scaling is adjusted, is obtained fixed big
Small pedestrian image.
It is optionally, described that feature extraction pretreatment is carried out to the pedestrian image using pedestrian's feature extraction network,
The fisrt feature figure with multiple features channel is obtained, multidomain treat-ment is carried out to the fisrt feature figure, obtains second feature figure, and
Feature extraction is carried out to the second feature figure, obtaining standard feature parameter includes: to the fisrt feature figure according to fine granularity
Multidomain treat-ment is carried out, second feature figure is obtained；Connect the channel data in each area of the second feature figure；To each area
Channel data compressed, obtain multiple channel characteristics vectors；Full connection is carried out to the channel characteristics vector to calculate, and is obtained
First passage data；The weighted value of the first passage data is calculated by activation primitive；By the power of the first passage data
Weight values weight one by one with the channel data to be multiplied, and obtains initial characteristics parameter；Convolution sum is carried out to the initial characteristics parameter
Full connection calculates, and obtains standard feature parameter.
It is appreciated that for example above-mentioned fisrt feature figure has C feature channel, to the fisrt feature figure according to fine granularity into
Row multidomain treat-ment obtains second feature figure, and 3 areas can be divided into fisrt feature figure, is such as segmented into head, abdomen chest, leg
Portion, each granularity do independent guidance, so that model learns more information to each granularity as far as possible.The division mode of subregion is pressed,
Each channel data of fisrt feature figure with C channel data is subjected to subregion division, generate subregion has C port number
According to second feature figure, it is subsequent to be handled by pond layer, i.e., the channel data in each area is compressed, is obtained
To multiple channel characteristics vectors；Full connection is carried out to the channel characteristics vector by full articulamentum again to calculate, and it is logical to obtain first
Track data；The weighted value of the first passage data is calculated by activation primitive, in general, the activation primitive for calculating weight can be with
It is sigmoid；The weighted value of the first passage data is weighted one by one with the channel data of fisrt feature figure and is multiplied, is obtained
To initial characteristics parameter, then convolution sum is carried out to the initial characteristics parameter and connects calculating entirely, obtains standard feature parameter.Such as
This, can be obtained the corresponding weighted value in each area in each channel data, to enhance more aobvious in the fisrt feature figure of input
More faint provincial characteristics, the partition characteristics being weighted mention in the provincial characteristics of work and the fisrt feature figure of weakening input
The feature for taking mode to extract more has identification, improves the accuracy rate that pedestrian identifies again.In addition, to the area of entire subregion
Domain calculates a weighted value, reduces the quantity of parameter, improves the efficiency of operation.
Optionally, pedestrian recognition methods again, further includes: be labeled as having together with the pedestrian image of a group traveling together
One kind, and a number expression is assigned, assigning different numbers to the inhomogeneous pedestrian image indicates.
In general, the pedestrian image detection processing or the side by manually marking can be carried out by pedestrian detection algorithm
Formula carries out pedestrian area mark.It should be understood that pedestrian image refers at least one pedestrian on image, when having in pedestrian image
When multiple pedestrians, which can be cut out, to guarantee to obtain a fairly large number of reliably sample, in general, by cutting
The size for cutting out obtained pedestrian area image has difference, and to be handled convenient for subsequent using unified algorithm, needing will be each
A pedestrian area Image Adjusting is identical size, and specific size can be subject to the demand of practical operation, is not done herein
It limits.In addition, the pedestrian area image to fixed size carries out classification processing, by the pedestrian with the fixed size with a group traveling together
Area image is labeled as same class, and a given number, the pedestrian area image of the fixed size with different pedestrians are divided into
Different classes, and mark into different numbers, in this way, in order to subsequent judgement operation.
Optionally, image classification and number operation can be carried out by artificial mode.
Optionally, the similarity of the characteristic parameter for calculating the pedestrian image to be measured and the standard feature parameter obtains
It include: characteristic parameter and the standard feature by cosine similarity function to the pedestrian image to be measured to similarity parameter
Parameter is calculated, and similarity parameter is obtained.
Optionally, complete pedestrian identifies to include: to judge whether the similarity parameter surpasses again according to the similarity parameter
Cross default value；If it is not, determining that pedestrian image to be measured and the pedestrian image are same class.
The present invention also provides a kind of pedestrian weight identifying systems, comprising: the first image collection module, for obtaining multiple pedestrians
Image constructs sample data set；
Training module constructs pedestrian's feature extraction network for being trained to the sample data set；
Characteristic extracting module, it is pre- for carrying out feature extraction to the pedestrian image using pedestrian's feature extraction network
Processing, obtains the fisrt feature figure with multiple features channel, carries out multidomain treat-ment to the fisrt feature figure, obtains second feature
Figure, and feature extraction is carried out to the second feature figure, obtain standard feature parameter；
Second image collection module obtains pedestrian image to be measured；
The characteristic extracting module is also used to extract the feature of the pedestrian image to be measured, obtains the pedestrian image to be measured
Characteristic parameter；
Judgment module, for calculating the characteristic parameter of the pedestrian image to be measured and the similarity of the standard feature parameter
Similarity parameter is obtained, and pedestrian is completed according to the similarity parameter and is identified again.
Optionally, it includes: acquisition unit that the first image, which obtains module, for obtaining original image；Detection unit is used
It is detected in the original image, position according to testing result and cuts out to obtain pedestrian area image；Adjustment unit is used for
The pedestrian area image scaling is adjusted, the pedestrian image of fixed size is obtained.
Optionally, the characteristic extracting module is also used to: multidomain treat-ment is carried out according to fine granularity to the fisrt feature figure,
Obtain second feature figure；Connect the channel data in each area of the second feature figure；To the channel data in each area into
Row compression, obtains multiple channel characteristics vectors；Full connection is carried out to the channel characteristics vector to calculate, and obtains first passage number
According to；The weighted value of the first passage data is calculated by activation primitive, in general, calculating the activation primitive of weight can be
Sigmoid activation primitive；The weighted value of the first passage data is weighted one by one with the channel data and is multiplied, is obtained initial
Characteristic parameter；Convolution sum is carried out to the initial characteristics parameter and connects calculating entirely, obtains standard feature parameter.
Optionally, the training module is also used to that will there is the pedestrian image of same a group traveling together to be labeled as same class, and
Assigning a number indicates, assigning different numbers to the inhomogeneous pedestrian image indicates.
Optionally, the judgment module includes: comparing unit, for passing through cosine similarity function to the pedestrian to be measured
The characteristic parameter of image is calculated with the standard feature parameter, obtains similarity parameter.
Optionally, the judgment module further include: decision package, for judging whether the similarity parameter is more than default
Numerical value；If it is not, the decision package determines pedestrian image to be measured and the pedestrian image is same class.
The present invention also provides a kind of computer readable storage mediums, are stored thereon with computer program, and the program is processed
Device realizes above-mentioned pedestrian recognition methods again when executing.
As described above, a kind of pedestrian provided by the invention recognition methods and system again, are carrying out feature extraction unit to pedestrian
Point, area's processing is carried out to characteristic image, the channel data in each area is compressed respectively, obtains the multiple of each subregion
Channel characteristics vector；Full connection is carried out to the channel characteristics vector to calculate, and obtains first passage data；Pass through activation primitive meter
The weighted value of the first passage data is calculated, in general, the activation primitive for calculating weight can be sigmoid activation primitive；It will
The weighted value of the first passage data weights one by one with the channel data to be multiplied, and each area in each channel data can be obtained
Corresponding weighted value, to enhance more significant provincial characteristics in the pedestrian image of input and weaken the pedestrian's figure inputted
More faint provincial characteristics, the feature that the partition characteristics extracting mode being weighted extracts more have identification as in,
To improve the discriminating power of depth model.
Detailed description of the invention
Fig. 1 is shown as pedestrian of the invention recognition methods flow chart again.
Fig. 2 is shown as the structural block diagram of pedestrian's weight identifying system of the invention.
Fig. 3 is shown as the structural block diagram of pedestrian's weight identifying system of the invention.
Fig. 4 is shown as the flow diagram of pedestrian of the invention recognition methods again.
Fig. 5 is shown as the test result comparison diagram of pedestrian of the invention recognition methods again.
Component label instructions
10 first image collection modules
11 acquisition units
12 detection units
13 adjustment units
20 training modules
30 characteristic extracting modules
40 second image collection modules
50 judgment modules
51 comparison units
52 decision packages
S10~S70 step
Specific embodiment
Illustrate embodiments of the present invention below by way of specific specific example, those skilled in the art can be by this specification
Other advantages and efficacy of the present invention can be easily understood for disclosed content.The present invention can also pass through in addition different specific realities
The mode of applying is embodied or practiced, the various details in this specification can also based on different viewpoints and application, without departing from
Various modifications or alterations are carried out under spirit of the invention.It should be noted that in the absence of conflict, following embodiment and implementation
Feature in example can be combined with each other.
It should be noted that illustrating the basic structure that only the invention is illustrated in a schematic way provided in following embodiment
Think, only shown in schema then with related component in the present invention rather than component count, shape and size when according to actual implementation
Draw, when actual implementation kenel, quantity and the ratio of each component can arbitrarily change for one kind, and its assembly layout kenel
It is likely more complexity.
Referring to Fig. 1, the present invention provides a kind of pedestrian's recognition methods again, comprising:
S10: obtaining multiple pedestrian images, constructs sample data set；
S20: being trained the sample data set, constructs pedestrian's feature extraction network；
S30: feature extraction pretreatment is carried out to the pedestrian image using pedestrian's feature extraction network, is had
The fisrt feature figure in multiple features channel carries out multidomain treat-ment to the fisrt feature figure, obtains second feature figure, and to described the
Two characteristic patterns carry out feature extraction, obtain standard feature parameter；
S40: pedestrian image to be measured is obtained；
S50: the feature of the pedestrian image to be measured is extracted, the characteristic parameter of the pedestrian image to be measured is obtained；
S60: the characteristic parameter of the calculating pedestrian image to be measured obtains similar to the similarity of the standard feature parameter
Spend parameter；
S70: pedestrian is completed according to the similarity parameter and is identified again；
In some embodiments, step S10 includes: acquisition original image；The original image is detected, according to
Testing result positions and cuts out to obtain pedestrian area image；The pedestrian area image scaling is adjusted, fixed size is obtained
Pedestrian image
It should be understood that original image can refer at least one pedestrian on the image, it is multiple when having on original image
When pedestrian, which can be cut out, to guarantee to obtain a fairly large number of reliably sample, in general, by cutting out
To the size of pedestrian area image have difference, to be handled convenient for subsequent using unified algorithm, need each row
People's area image is adjusted to identical size, obtains pedestrian image.Specific size can be subject to the demand of practical operation,
It is not limited here.
It is appreciated that can will will have the pedestrian image of same a group traveling together to be labeled as together after obtaining multiple pedestrian images
One kind, and a number expression is assigned, assigning different numbers to the inhomogeneous pedestrian image indicates, in this way, sentencing so as to subsequent
Disconnected operation.
In some embodiments, image classification and number operation can be carried out by artificial mode.
It is appreciated that using trained feature extraction network carries out the operation of step S50 in step S20, to mention
The feature for taking the pedestrian image to be measured obtains the characteristic parameter of the pedestrian image to be measured.
It is understood that same class will be classified as with a group traveling together after pedestrian image is classified and numbered, and assign a volume
Number, after the subsequent judgment step for completing step S60 and S70, such as meet whether relevant preset condition i.e. similarity parameter surpasses
Default value is crossed, the image data to match with testing data can be focused to find out in data, pedestrian image to be measured can match
Multiple image datas of same class pedestrian image are found, user can call same class pedestrian image according to practical application request
Multiple image datas or an image data.
In some embodiments, step S30 is the following steps are included: divide the fisrt feature figure according to fine granularity
Area's processing, obtains second feature figure；Connect the channel data in each area of the second feature figure；To the channel in each area
Data are compressed, and multiple channel characteristics vectors are obtained；Full connection is carried out to the channel characteristics vector to calculate, and it is logical to obtain first
Track data；The weighted value of the first passage data is calculated by activation primitive, in general, the activation primitive for calculating weight can be with
It is sigmoid；The weighted value of the first passage data is weighted one by one with the channel data of the fisrt feature figure and is multiplied, is obtained
To initial characteristics parameter；Convolution sum is carried out to the initial characteristics parameter and connects calculating entirely, obtains standard feature parameter.
In some embodiments, the characteristic parameter for calculating the pedestrian image to be measured and the standard feature parameter
Similarity to obtain similarity parameter include: characteristic parameter and the institute by cosine similarity function to the pedestrian image to be measured
It states standard feature parameter to be calculated, obtains similarity parameter.Optionally, can also with other distance metric functions, as geneva away from
From etc., COS distance and Euclidean distance are not limited to calculate the characteristic parameter of pedestrian image to be measured and the standard feature and join
Several similarities obtains similarity parameter, and the calculation about similarity is herein with no restrictions.
In some embodiments, complete pedestrian identifies to include: to judge the similarity again according to the similarity parameter
Whether parameter is more than default value；If it is not, determining that pedestrian image to be measured and the pedestrian image are same class.It should be understood that
Preset value herein can be user according to the allowable range of error of actual operation setting, such as can be 0.100, herein not
Specific numerical value is defined.
Referring to Fig. 4, Fig. 4 is shown as the flow diagram of pedestrian of the invention recognition methods again, after C representative is conversion
Characteristic pattern port number, H represents the height of characteristic pattern, and W represents the width of characteristic pattern.It is appreciated that for example above-mentioned first is special
Sign figure has C feature channel, carries out multidomain treat-ment according to fine granularity to the fisrt feature figure, obtains second feature figure, can be with
3 areas are divided into fisrt feature figure, are such as segmented into head, abdomen chest, leg, each granularity does independent guidance, so that mould
Type learns more information to each granularity as far as possible.The division mode for pressing subregion, by the fisrt feature figure with C channel data
Each channel data carries out subregion division, generates the second feature figure for having C channel data of subregion, subsequent to pass through pond
Layer is handled, i.e., compresses to the channel data in each area, obtain multiple channel characteristics vectors；Again by connecting entirely
Layer carries out full connection to the channel characteristics vector and calculates, and obtains first passage data；Described first is calculated by activation primitive
The weighted value of channel data, in general, the activation primitive for calculating weight can be sigmoid；By the first passage data
Weighted value weights one by one with the channel data of the fisrt feature figure to be multiplied, and obtains initial characteristics parameter；To the initial characteristics
Parameter carries out convolution sum and connects calculating entirely, obtains standard feature parameter.In this way, it is corresponding that each area in each channel data can be obtained
Weighted value, to enhance more significant provincial characteristics in the fisrt feature figure of input and weaken the fisrt feature of input
More faint provincial characteristics in figure, the feature that the partition characteristics extracting mode being weighted extracts more have identification,
Improve the accuracy rate that pedestrian identifies again.In addition, the region to entire subregion calculates a weighted value, reduce the number of parameter
Amount, improves the efficiency of operation.
In some embodiments, multidomain treat-ment is carried out according to fine granularity to the fisrt feature figure, obtains second feature
Figure, i.e. second figure from left to right in Fig. 4, after being operated later using global average pondization to each of which second feature figure, that is, subregion
Characteristic pattern compressed, so that the characteristic pattern in C channel is eventually become the real number ordered series of numbers of 1*1*C*3.General convolutional neural networks
In the filter that learns of each channel the feature channel of local experiences is operated, therefore each characteristic pattern can not
It is all very little that using the contextual information of other feature figure, and on the lower level of network, it, which experiences feature channel size,
Therefore simplest global average pondization operation is selected to compress characteristic image, so that global receptive field is made it have,
So that Network Low-layer can also utilize global information.Characteristic pattern herein is the characteristic pattern after second feature figure, that is, subregion.
It should be understood that step S70: calculating the weighted value of the first passage data by activation primitive；Activation operation
Similar to the mechanism of door in Recognition with Recurrent Neural Network, it is made of two full articulamentums, dimension-reduction treatment can be carried out, such as by feature dimensions
Degree is reduced to the 1/16 of input, then after activation primitive ReLu activation, original dimension is raised to by full articulamentum, finally
Normalized weight between 0~1 is obtained by the door of excitation function Sigmoid, expression formula is as follows: s=F (t, W)=σ (g (t,
W in))=σ (W δ (Wt)) formula, δ is ReLu activation primitive, and σ is Sigmoid function, and parameter W is used to explicitly Modelling feature channel
Between correlation.Later the weighted value of the first passage data is weighted again one by one with the channel data and be multiplied, is marked
Quasi- characteristic parameter realizes the recalibration to each subregion and each channel characteristics.
It should be understood that the fisrt feature figure and second feature figure in present invention statement are all characteristic pattern, wherein second
Characteristic pattern refers in particular to the characteristic pattern after subregion.By carrying out subsequent processing again after carrying out subregion to characteristic pattern, that is, fisrt feature figure, point
Qu Hou, then handled to obtain characteristic parameter more than the parameter than directly carrying out weight calculation to each pixel to specified region
The characteristic parameter quantity obtained less is less, and the characteristic pattern after subregion is more targeted.Such as subregion is segmented into head, abdomen
Chest, the granularity of leg three, each granularity do independent guidance, such as subsequent acquisition pedestrian image to be measured this three parts feature more
Add significantly, feature extraction can be carried out for head, abdomen chest, the leg for surveying pedestrian image, obtain corresponding parameter, reduce work
While amount target also definitely, the subsequent characteristic parameter for calculating the pedestrian image to be measured and the standard feature parameter
Similarity obtains similarity parameter；Also more having purpose is judged to the similarity parameter.Therefore using weighting
Partition characteristics extracting mode has carried out recalibration to the different subregions in different characteristic channel, and the feature extracted, which more has, to be sentenced
Other property improves the accuracy rate that pedestrian identifies again.In addition, multidomain treat-ment has been carried out in view of to characteristic pattern, it is subsequent to multichannel spy
Sign carries out subregion weight calculation, without carrying out weight calculation to characteristic pattern subregion with the feature channel of outer portion, in this way, reducing
The calculation amount of parameter.
Referring to Fig. 2, the present invention provides a kind of pedestrian's weight identifying system, comprising: the first image collection module 10, for obtaining
Multiple pedestrian images are taken, sample data set is constructed；Training module 20, for being trained to the sample data set, building row
People's feature extraction network；Characteristic extracting module 30, for being carried out using pedestrian's feature extraction network to the pedestrian image
Feature extraction pretreatment, obtains the fisrt feature figure with multiple features channel, carries out multidomain treat-ment to the fisrt feature figure, obtains
Feature extraction is carried out to second feature figure, and to the second feature figure, obtains standard feature parameter；Second image collection module
40, obtain pedestrian image to be measured；The characteristic extracting module 30 is also used to extract the feature of the pedestrian image to be measured, obtains institute
State the characteristic parameter of pedestrian image to be measured；Judgment module 50, for calculate the characteristic parameter of the pedestrian image to be measured with it is described
The similarity of standard feature parameter obtains similarity parameter, and completes pedestrian according to the similarity parameter and identify again.
Referring to Fig. 3, it includes: acquisition unit 11 that the first image, which obtains module 10, in some embodiments, it is used for
Obtain original image；Detection unit 12, for being detected to the original image, positioning according to testing result and cutting out to obtain
Pedestrian area image；Adjustment unit 13 obtains pedestrian's figure of fixed size for adjusting to the pedestrian area image scaling
Picture.
In some embodiments, characteristic extracting module 30 is also used to: being carried out to the fisrt feature figure according to fine granularity
Multidomain treat-ment obtains second feature figure；Connect the channel data in each area of the second feature figure；Each area is led to
Track data is compressed, and multiple channel characteristics vectors are obtained；Full connection is carried out to the channel characteristics vector to calculate, and obtains first
Channel data；The weighted value of the first passage data is calculated by activation primitive；By the weighted value of the first passage data
It weights and is multiplied one by one with the channel data of fisrt feature figure, obtain initial characteristics parameter；The initial characteristics parameter is rolled up
Long-pending and full connection calculates, and obtains standard feature parameter.
In some embodiments, the training module 20 is also used to have the pedestrian image of same a group traveling together to mark
It for same class, and assigns a number and indicates, different numbers expressions are assigned to the inhomogeneous pedestrian image.
In some embodiments, the judgment module 50 further includes comparing unit 51, for passing through cosine similarity letter
Several characteristic parameters to the pedestrian image to be measured are calculated with the standard feature parameter, obtain similarity parameter.
In some embodiments, the judgment module 50 further includes decision package 52, for judging the similarity ginseng
Whether number is more than default value；If it is not, the decision package determines pedestrian image to be measured and the pedestrian image is same class.This
The specific embodiment of recognition methods system and associated beneficial effect are identical as pedestrian's again recognition methods again by the pedestrian of invention, herein
It repeats no more.
Referring to Fig. 5, Rank1 is the first hit rate, the figure exactly to make number one is either with or without hit himself, map
Thoroughly evaluating pedestrian weight identification technology index mean accuracy mean value, be respectively adopted existing recognition methods (baseline),
Do not carry out the knowledge of recognition methods (se), the present inventor of subregion weight calculation to the multi-channel feature of characteristic pattern compared with the present invention
Other method (se_patch) respectively surveys three data sets Market1501, DukeMTMC-reID and cuhk03-np
Examination, obtained test result show equal using pedestrian of the invention recognition methods obtains again the first hit rate and mean accuracy
The value of value is all relatively high.
As described above, a kind of pedestrian provided by the invention recognition methods and system again, are carrying out feature extraction unit to pedestrian
Point, to characteristic image carry out multidomain treat-ment, the channel data of each subregion is compressed, obtain multiple channel characteristics to
Amount；Full connection is carried out to the channel characteristics vector to calculate, and obtains first passage data；Described first is calculated by activation primitive
The weighted value of channel data；The weighted value of the first passage data is weighted one by one with the channel data and is multiplied, can be obtained
The corresponding weighted value in each area in each channel data, so that it is special to enhance more significant region in the fisrt feature figure of input
More faint provincial characteristics, the partition characteristics extracting mode being weighted extract in sign and the fisrt feature figure of weakening input
To feature more there is identification, to improve the discriminating power of depth model.
The above-described embodiments merely illustrate the principles and effects of the present invention, and is not intended to limit the present invention.It is any ripe
The personage for knowing this technology all without departing from the spirit and scope of the present invention, carries out modifications and changes to above-described embodiment.Cause
This, institute is complete without departing from the spirit and technical ideas disclosed in the present invention by those of ordinary skill in the art such as
At all equivalent modifications or change, should be covered by the claims of the present invention.
Claims (13)
1. a kind of pedestrian recognition methods again characterized by comprising
Multiple pedestrian images are obtained, sample data set is constructed；
The sample data set is trained, pedestrian's feature extraction network is constructed；
Feature extraction pretreatment is carried out to the pedestrian image using pedestrian's feature extraction network, is obtained logical with multiple features
The fisrt feature figure in road carries out multidomain treat-ment to the fisrt feature figure, obtains second feature figure, and to the second feature figure
Feature extraction is carried out, standard feature parameter is obtained；
Obtain pedestrian image to be measured；
The feature for extracting the pedestrian image to be measured obtains the characteristic parameter of the pedestrian image to be measured；
The similarity of the characteristic parameter and the standard feature parameter that calculate the pedestrian image to be measured obtains similarity parameter；
Pedestrian is completed according to the similarity parameter to identify again.
2. pedestrian according to claim 1 recognition methods again, which is characterized in that described to obtain multiple pedestrian images and include:
Obtain original image；
The original image is detected, position according to testing result and cuts out to obtain pedestrian area image；
The pedestrian area image scaling is adjusted, the pedestrian image of fixed size is obtained.
3. pedestrian according to claim 1 recognition methods again, which is characterized in that described to use pedestrian's feature extraction net
Network carries out feature extraction pretreatment to the pedestrian image, the fisrt feature figure with multiple features channel is obtained, to described first
Characteristic pattern carries out multidomain treat-ment, obtains second feature figure, and carry out feature extraction to the second feature figure, obtains standard feature
Parameter includes:
Multidomain treat-ment is carried out according to fine granularity to the fisrt feature figure, obtains second feature figure；
Connect the channel data in each area of the second feature figure；
The channel data in each area is compressed, multiple channel characteristics vectors are obtained；
Full connection is carried out to the channel characteristics vector to calculate, and obtains first passage data；
The weighted value of the first passage data is calculated by activation primitive；
The weighted value of the first passage data is weighted one by one with the channel data and is multiplied, initial characteristics parameter is obtained；
Convolution sum is carried out to the initial characteristics parameter and connects calculating entirely, obtains standard feature parameter.
4. pedestrian according to claim 1 recognition methods again, which is characterized in that further include:
To have and be labeled as same class with the pedestrian image of a group traveling together, and assign a number to indicate, to inhomogeneous described
Pedestrian image is assigned different numbers and is indicated.
5. pedestrian according to claim 1 recognition methods again, which is characterized in that the calculating pedestrian image to be measured
The similarity of characteristic parameter and the standard feature parameter obtains similarity parameter and includes:
It is calculated by characteristic parameter of the cosine similarity function to the pedestrian image to be measured with the standard feature parameter,
Obtain similarity parameter.
6. pedestrian according to claim 4 recognition methods again, which is characterized in that complete pedestrian according to the similarity parameter
It identifies again and includes:
Judge whether the similarity parameter is more than default value；
If it is not, determining that the pedestrian image to be measured and the pedestrian image are same class.
7. a kind of pedestrian's weight identifying system characterized by comprising
First image collection module constructs sample data set for obtaining multiple pedestrian images；
Training module constructs pedestrian's feature extraction network for being trained to the sample data set；
Characteristic extracting module is located in advance for carrying out feature extraction to the pedestrian image using pedestrian's feature extraction network
Reason, obtains the fisrt feature figure with multiple features channel, carries out multidomain treat-ment to the fisrt feature figure, obtains second feature
Figure, and feature extraction is carried out to the second feature figure, obtain standard feature parameter；
Second image collection module obtains pedestrian image to be measured；
The characteristic extracting module is also used to extract the feature of the pedestrian image to be measured, obtains the spy of the pedestrian image to be measured
Levy parameter；
Judgment module, characteristic parameter and the similarity of the standard feature parameter for calculating the pedestrian image to be measured obtain
Similarity parameter, and pedestrian is completed according to the similarity parameter and is identified again.
8. pedestrian's weight identifying system according to claim 7, which is characterized in that the first image obtains module and includes:
Acquisition unit, for obtaining original image；
Detection unit, for being detected to the original image, positioning according to testing result and cutting out to obtain pedestrian area figure
Picture；
Adjustment unit obtains the pedestrian image of fixed size for adjusting to the pedestrian area image scaling.
9. pedestrian's weight identifying system according to claim 7, which is characterized in that the characteristic extracting module is also used to:
Multidomain treat-ment is carried out according to fine granularity to the fisrt feature figure, obtains second feature figure；
Connect the channel data in each area of the second feature figure；
The channel data in each area is compressed, multiple channel characteristics vectors are obtained；
Full connection is carried out to the channel characteristics vector to calculate, and obtains first passage data；
The weighted value of the first passage data is calculated by activation primitive；
The weighted value of the first passage data is weighted one by one with the channel data and is multiplied, initial characteristics parameter is obtained；
Convolution sum is carried out to the initial characteristics parameter and connects calculating entirely, obtains standard feature parameter.
10. pedestrian's weight identifying system according to claim 7, which is characterized in that the training module is also used to have
The pedestrian image with a group traveling together is labeled as same class, and assigns a number expression, assigns to the inhomogeneous pedestrian image
Difference number indicates.
11. pedestrian according to claim 7 weight identifying system, which is characterized in that the judgment module includes:
Comparing unit, for passing through characteristic parameter and the standard feature of the cosine similarity function to the pedestrian image to be measured
Parameter is calculated, and similarity parameter is obtained.
12. pedestrian's weight identifying system according to claim 10, which is characterized in that the judgment module further include:
Decision package, for judging whether the similarity parameter is more than default value；
If it is not, the decision package determines pedestrian image to be measured and the pedestrian image is same class.
13. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is by processor
The recognition methods again of pedestrian described in any one of claims 1 to 6 is realized when execution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910156179.2A CN109886242A (en) | 2019-03-01 | 2019-03-01 | A kind of method and system that pedestrian identifies again |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910156179.2A CN109886242A (en) | 2019-03-01 | 2019-03-01 | A kind of method and system that pedestrian identifies again |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109886242A true CN109886242A (en) | 2019-06-14 |
Family
ID=66930310
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910156179.2A Pending CN109886242A (en) | 2019-03-01 | 2019-03-01 | A kind of method and system that pedestrian identifies again |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109886242A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110532884A (en) * | 2019-07-30 | 2019-12-03 | 平安科技（深圳）有限公司 | Pedestrian recognition methods, device and computer readable storage medium again |
| CN110929628A (en) * | 2019-11-18 | 2020-03-27 | 北京三快在线科技有限公司 | Human body identification method and device |
| CN111080626A (en) * | 2019-12-19 | 2020-04-28 | 联想(北京)有限公司 | Detection method and electronic equipment |
| CN111611846A (en) * | 2020-03-31 | 2020-09-01 | 北京迈格威科技有限公司 | Pedestrian re-identification method and device, electronic equipment and storage medium |
| CN112215840A (en) * | 2020-10-30 | 2021-01-12 | 上海商汤临港智能科技有限公司 | Image detection method, image detection device, driving control method, driving control device, electronic equipment and storage medium |
| CN113627352A (en) * | 2021-08-12 | 2021-11-09 | 塔里木大学 | Pedestrian re-identification method and system |
| CN115858846A (en) * | 2023-02-16 | 2023-03-28 | 云南派动科技有限公司 | Deep learning-based skier image retrieval method and system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107133575A (en) * | 2017-04-13 | 2017-09-05 | 中原智慧城市设计研究院有限公司 | A kind of monitor video pedestrian recognition methods again based on space-time characteristic |
| CN108090203A (en) * | 2017-12-25 | 2018-05-29 | 上海七牛信息技术有限公司 | Video classification methods, device, storage medium and electronic equipment |
| CN108427920A (en) * | 2018-02-26 | 2018-08-21 | 杭州电子科技大学 | A kind of land and sea border defense object detection method based on deep learning |
| CN108764308A (en) * | 2018-05-16 | 2018-11-06 | 中国人民解放军陆军工程大学 | A kind of recognition methods again of the pedestrian based on convolution loop network |
| CN109271870A (en) * | 2018-08-21 | 2019-01-25 | 平安科技（深圳）有限公司 | Pedestrian recognition methods, device, computer equipment and storage medium again |
-
2019
- 2019-03-01 CN CN201910156179.2A patent/CN109886242A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107133575A (en) * | 2017-04-13 | 2017-09-05 | 中原智慧城市设计研究院有限公司 | A kind of monitor video pedestrian recognition methods again based on space-time characteristic |
| CN108090203A (en) * | 2017-12-25 | 2018-05-29 | 上海七牛信息技术有限公司 | Video classification methods, device, storage medium and electronic equipment |
| CN108427920A (en) * | 2018-02-26 | 2018-08-21 | 杭州电子科技大学 | A kind of land and sea border defense object detection method based on deep learning |
| CN108764308A (en) * | 2018-05-16 | 2018-11-06 | 中国人民解放军陆军工程大学 | A kind of recognition methods again of the pedestrian based on convolution loop network |
| CN109271870A (en) * | 2018-08-21 | 2019-01-25 | 平安科技（深圳）有限公司 | Pedestrian recognition methods, device, computer equipment and storage medium again |
Non-Patent Citations (1)
| Title |
|---|
| LAVI_QQ_2910138025: "SEnet详解", 《HTTPS://BLOG.CSDN.NET/LIUWEIYUXIANG/ARTICLE/DETAILS/84075343》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110532884A (en) * | 2019-07-30 | 2019-12-03 | 平安科技（深圳）有限公司 | Pedestrian recognition methods, device and computer readable storage medium again |
| CN110532884B (en) * | 2019-07-30 | 2024-04-09 | 平安科技（深圳）有限公司 | Pedestrian re-recognition method, device and computer readable storage medium |
| CN110929628A (en) * | 2019-11-18 | 2020-03-27 | 北京三快在线科技有限公司 | Human body identification method and device |
| CN111080626A (en) * | 2019-12-19 | 2020-04-28 | 联想(北京)有限公司 | Detection method and electronic equipment |
| CN111611846A (en) * | 2020-03-31 | 2020-09-01 | 北京迈格威科技有限公司 | Pedestrian re-identification method and device, electronic equipment and storage medium |
| WO2021196547A1 (en) * | 2020-03-31 | 2021-10-07 | 北京迈格威科技有限公司 | Person re-identification method and apparatus, electronic device and storage medium |
| CN112215840A (en) * | 2020-10-30 | 2021-01-12 | 上海商汤临港智能科技有限公司 | Image detection method, image detection device, driving control method, driving control device, electronic equipment and storage medium |
| CN113627352A (en) * | 2021-08-12 | 2021-11-09 | 塔里木大学 | Pedestrian re-identification method and system |
| CN115858846A (en) * | 2023-02-16 | 2023-03-28 | 云南派动科技有限公司 | Deep learning-based skier image retrieval method and system |
| CN115858846B (en) * | 2023-02-16 | 2023-04-21 | 云南派动科技有限公司 | Skier image retrieval method and system based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109886242A (en) | A kind of method and system that pedestrian identifies again | |
| CN104166841B (en) | The quick detection recognition methods of pedestrian or vehicle is specified in a kind of video surveillance network | |
| CN107016370B (en) | A kind of partial occlusion face identification method based on data enhancing | |
| CN104063719B (en) | Pedestrian detection method and device based on depth convolutional network | |
| CN106960210B (en) | The method and apparatus of target detection | |
| CN103996195B (en) | Image saliency detection method | |
| CN109819208A (en) | A kind of dense population security monitoring management method based on artificial intelligence dynamic monitoring | |
| CN107085696A (en) | A kind of vehicle location and type identifier method based on bayonet socket image | |
| CN109446925A (en) | A kind of electric device maintenance algorithm based on convolutional neural networks | |
| CN108470138A (en) | Method for target detection and device | |
| CN107133569A (en) | The many granularity mask methods of monitor video based on extensive Multi-label learning | |
| CN109815874A (en) | A kind of personnel identity recognition methods, device, equipment and readable storage medium storing program for executing | |
| CN103971106B (en) | Various visual angles facial image gender identification method and device | |
| CN110263656A (en) | A kind of cancer cell identification methods, devices and systems | |
| CN110119726A (en) | A kind of vehicle brand multi-angle recognition methods based on YOLOv3 model | |
| CN104504395A (en) | Method and system for achieving classification of pedestrians and vehicles based on neural network | |
| CN111027378A (en) | Pedestrian re-identification method, device, terminal and storage medium | |
| CN108596211A (en) | It is a kind of that pedestrian's recognition methods again is blocked based on focusing study and depth e-learning | |
| CN106557740B (en) | The recognition methods of oil depot target in a kind of remote sensing images | |
| CN102629320A (en) | Ordinal measurement statistical description face recognition method based on feature level | |
| CN108647695A (en) | Soft image conspicuousness detection method based on covariance convolutional neural networks | |
| CN108960142B (en) | Pedestrian re-identification method based on global feature loss function | |
| CN105184229A (en) | Online learning based real-time pedestrian detection method in dynamic scene | |
| CN112183438B (en) | Image identification method for illegal behaviors based on small sample learning neural network | |
| CN109410184A (en) | Live streaming pornographic image detection method based on dense confrontation network semi-supervised learning |
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 |
Application publication date: 20190614 |
|
| RJ01 | Rejection of invention patent application after publication |