CN109871821A - The pedestrian of adaptive network recognition methods, device, equipment and storage medium again - Google Patents
The pedestrian of adaptive network recognition methods, device, equipment and storage medium again Download PDFInfo
- Publication number
- CN109871821A CN109871821A CN201910159239.6A CN201910159239A CN109871821A CN 109871821 A CN109871821 A CN 109871821A CN 201910159239 A CN201910159239 A CN 201910159239A CN 109871821 A CN109871821 A CN 109871821A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- branch
- patch
- network structure
- fine granularity
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The present invention provides the pedestrian of adaptive network a kind of recognition methods, device, equipment and storage medium again, this method comprises: obtaining pedestrian's video, pedestrian detection algorithm is utilized to extract pedestrian image in pedestrian's video;The residual error network structure of redundancy is adaptively adjusted according to current scene to optimum network structure;Utilize the feature vector of the optimum network structure extraction pedestrian image to be measured based on residual error network;Identify that the cosine similarity of the feature vector between pedestrian to be measured and default pedestrian obtains pedestrian's weight recognition result.The present invention is from the precise and penetrating thick automatic patch branch divided in redundancy residual error network structure, until the optimum network structure being matched under current scene, it reduces and carries out repeatedly the huge workload that engineer is paid with experiment in order to obtain optimum network framework, simultaneously, erroneous judgement caused by avoiding because of human factor, the free switching being more suitable between multiple scenes extends convenient for subsequent scenario scene.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of adaptive network pedestrian again recognition methods,
Device, equipment and storage medium.
Background technique
Pedestrian identifies that (Person re-identification, RE-ID) is also referred to as pedestrian and identifies again again, that is, judgement is different
The camera of position takes whether pedestrian target is same people in different moments, can be used for video monitoring etc..Traditionally,
Realize that pedestrian identifies again by extracting the artificial design features of pedestrian image and the artificial design features of extraction being compared,
Since pedestrian's picture to be identified is shot from original picture in different cameras, the difference of equipment can bring error to image-forming condition;
Environment under different scenes is inconsistent, and the data of acquisition can also generate different deviations;And the change meeting of illumination is so that same
The performance of kind color is different;Importantly, attitudes vibration and occlusion issue of the pedestrian under camera, all make to same
Personal discrimination difficulty is quite big.
However, needing to combine global characteristics that could preferably express with local feature in existing pedestrian's weight identification process
Pedestrian image, if to reach certain discrimination, needs to design from the point of view of the pedestrian of current industry weight Study of recognition situation
The network structure of gloable branch and a series of fine granularity patch branch rationally to merge the global characteristics drawn game of pedestrian image
Portion's feature, since the environment under different scenes is inconsistent, the data of acquisition can also generate different deviations;And image recognition is calculated
Method is depended on contextual data unduly itself, causes the data of different scenes that may need to the division mode of fine granularity feature not
Together, artificial experiment is relied on to determine that fine granularity feature divides not only heavy workload, and being also difficult to look for can not find and current application
The most suitable network structure of scene matching.
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 of adaptive network weights
Recognition methods, device, equipment and storage medium, for solve pedestrian in the prior art identify again can not according to current scene from
It adapts to match suitable network structure problem.
In order to achieve the above objects and other related objects, the application's in a first aspect, present invention provides a kind of adaptive net
The pedestrian of network recognition methods again, comprising:
Pedestrian's video is obtained, extracts pedestrian image in pedestrian's video using pedestrian detection algorithm;
The residual error network structure of redundancy is adaptively adjusted according to current scene to optimum network structure;
Utilize the feature vector of the optimum network structure extraction pedestrian image to be measured based on residual error network;
Identify that the cosine similarity of the feature vector between pedestrian to be measured and default pedestrian obtains pedestrian's weight recognition result.
The second aspect of the application provides a kind of pedestrian's weight identification device of adaptive network, comprising:
Pedestrian detection module extracts pedestrian in pedestrian's video using pedestrian detection algorithm for obtaining pedestrian's video
Image;
Adaptive network structure, for adaptively adjusting the residual error network structure of redundancy according to current scene to optimum network
Structure;
Characteristic extracting module, for utilizing the optimum network structure extraction pedestrian to be measured based on residual error network
The feature vector of image;
Pedestrian's weight identification module, the cosine similarity of the feature vector between pedestrian to be measured and default pedestrian obtains for identification
To pedestrian's weight recognition result.
The third aspect of the application, the pedestrian for providing a kind of adaptive network identify electronic equipment again, comprising:
One or more processors;
Memory;And
One or more programs, wherein one or more of programs be stored in the memory and be configured as by
One or more of processors execute instruction, and execute instruction described in one or more of processors execution so that the electronics
Equipment execute the pedestrian of above-mentioned adaptive network again recognition methods the step of.
The fourth aspect of the application, provides a kind of computer readable storage medium, and the computer readable storage medium is deposited
Contain computer program, the computer program realizes above-mentioned adaptive network pedestrian recognition methods again when being executed by processor
Step.
As described above, the pedestrian of adaptive network of the invention recognition methods, device, equipment and storage medium again, have
Below the utility model has the advantages that
The present invention works as front court until being matched to from the precise and penetrating thick automatic patch branch divided in redundancy residual error network structure
Best (suitable) network structure under scape, reduction carries out repeatedly engineer in order to obtain optimum network framework and experiment is paid
Huge workload out, meanwhile, it is capable to the network structure adaptively more simplified under the premise of meeting index, avoid because
Erroneous judgement caused by human factor, the free switching being more suitable between multiple scenes extend convenient for subsequent scenario scene.
Detailed description of the invention
Fig. 1 is shown as a kind of pedestrian of adaptive network provided by the invention recognition methods flow chart again;
Fig. 2 be shown as the pedestrian of adaptive network provided by the invention a kind of again in recognition methods step S2 flow chart;
Fig. 3 be shown as the pedestrian of adaptive network provided by the invention a kind of again in recognition methods step S4 flow chart;
Fig. 4 is shown as a kind of pedestrian's weight identification device structural block diagram of adaptive network provided by the invention;
The pedestrian that Fig. 5 is shown as a kind of adaptive network provided by the invention identifies electronic devices structure block diagram again.
Component label instructions:
1 pedestrian detection module
2 adaptive network structures
3 characteristic extracting modules
4 pedestrian's identification modules
S1~S4 step 1 is to 4
Specific embodiment
Presently filed embodiment is illustrated by particular specific embodiment below, those skilled in the art can be by this explanation
Content disclosed by book understands other advantages and effect of the application easily.
In described below, with reference to attached drawing, attached drawing describes several embodiments of the application.It should be appreciated that also can be used
Other embodiments, and can be carried out without departing substantially from spirit and scope of the present disclosure mechanical composition, structure, electrically with
And the operational detailed description changed below should not be considered limiting, and the range of embodiments herein
Only the limited of claims of the patent by announcing term used herein is merely to describe specific embodiment, and be not
It is intended to limit the application.The term of space correlation, for example, "upper", "lower", "left", "right", " following ", " lower section ", " lower part ",
" top ", " top " etc. can be used in the text in order to an elements or features and another element or spy shown in explanatory diagram
The relationship of sign.
Although term first, second etc. are used to describe various elements herein in some instances, these elements
It should not be limited by these terms.These terms are only used to distinguish an element with another element.For example, first is pre-
If threshold value can be referred to as the second preset threshold, and similarly, the second preset threshold can be referred to as the first preset threshold, and
The range of various described embodiments is not departed from.First preset threshold and preset threshold are to describe a threshold value, still
Unless context otherwise explicitly points out, otherwise they are not the same preset thresholds.Similar situation further includes first
Volume and the second volume.
Furthermore as used in herein, singular " one ", "one" and "the" are intended to also include plural number shape
Formula, unless having opposite instruction in context it will be further understood that term "comprising", " comprising " show that there are the spies
Sign, step, operation, element, component, project, type, and/or group, but it is not excluded for one or more other features, step, behaviour
Work, element, component, project, the presence of type, and/or group, appearance or addition term "or" used herein and "and/or" quilt
Be construed to inclusive, or mean any one or any combination therefore, " A, B or C " or " A, B and/or C " mean " with
Descend any one: A;B;C;A and B;A and C;B and C;A, B and C " is only when element, function, step or the combination of operation are in certain modes
Under it is inherently mutually exclusive when, just will appear the exception of this definition.
Referring to Fig. 1, being a kind of pedestrian of adaptive network provided by the invention recognition methods flow chart again, comprising:
Step S1 obtains pedestrian's video, extracts pedestrian image in pedestrian's video using pedestrian detection algorithm;
Wherein, pedestrian's video refers to the continuous multiple image (original image) of the shooting of the camera with monitoring function.
Pedestrian image refer to include in original image pedestrian image.It is to be appreciated that may be without pedestrian (such as original in a frame image
Beginning image is all background), it is also possible to there are multiple pedestrians.I.e. detection image can be plural width, if detection image is plural width,
The technical solution of the present embodiment is executed to each width pedestrian image respectively.
In a specific embodiment, the acquisition of original image, that is, original image can be obtained from a video sequence
It arrives, client can also obtain original image from video sequence, then send server-side for original image, can also be direct
Server-side is sent by video sequence, server-side obtains original image from video sequence again, then uses pedestrian detection algorithm pair
Original image is detected, and pedestrian image to be measured is obtained.
Wherein, pedestrian detection algorithm refer to by detecting judge in image or video sequence with the presence or absence of pedestrian and to
Give pinpoint algorithm.Specifically, pedestrian detection algorithm can be the artificial pedestrian detection algorithm for extracting feature, can also be with base
In the pedestrian detection algorithm of neural network, it can also be the pedestrian detection algorithm based on deep learning, use base in the present embodiment
In the pedestrian detection algorithm of deep learning.
In a specific embodiment, after the step of being detected using pedestrian detection algorithm to original image, also
Pedestrian in original image is intercepted according to pixel size, the image that specific shape is obtained after interception is pedestrian image,
Such as figure is intercepted by pedestrian's actual size, by scaling interception image to fixed size, such as uniformly zoom to 256*
128。
In addition, Faster RCNN human testing model trained in advance can also be used directly from pedestrian video extraction pedestrian
Image, wherein Faster RCNN algorithm is to combine RPN and Fast-RCNN, extracts candidate frame by RPN, Fast R-CNN is negative
Duty detection detects the RPN candidate frame extracted and identifies the target in frame.
Step S2 adaptively adjusts the residual error network structure of redundancy according to current scene to optimum network structure;
Referring to Fig. 2, for the process of step S2 in a kind of pedestrian of adaptive network provided by the invention again recognition methods
Figure, comprising:
Step S201, the residual error network structure of the redundancy include a gloable branch and several fine granularities patch
Branch, wherein several fine granularity patch branches include two fine granularity patch branches, Yi Ji great by its division mode
In 2 multiple odd number fine granularity patch branches;
Step S202 adaptively adjusts fine granularity patch branch in the residual error network structure of redundancy according to current scene
Quantity, multiple odd number fine granularity patch branches from it is precise and penetrating slightly successively decrease division fine granularity patch branch quantity, by institute
State that each branch in residual error network structure connects softmax loss function respectively and triplet loss function is trained to convergence;
Wherein, loss function, learning training are used to estimate the predicted value of model and the departure degree of true value in the process.It
It is a non-negative real-valued function.It is to minimize loss function to the target that model optimizes in training process.
Step S203, judges whether the corresponding patch feature weight of each fine granularity branch is less than preset threshold, if institute
When stating patch feature weight less than preset threshold, then the fine granularity branch is deleted;If the patch feature weight is not less than
When preset threshold, then retain the fine granularity branch;
Specifically, the corresponding patch feature weight of each fine granularity branch is calculated, several division sides may be existed simultaneously
The feature weight of the corresponding fine granularity branch of formula is both less than preset threshold, that is, needs while deleting a variety of division mode fine granularities
Branch.
Step S204 is calculated and is pressed the respective feature weight of multiple odd number division mode corresponding fine granularity patch branch,
Delete the corresponding division mode of feature weight the smallest fine granularity patch branch;
Specifically, step S203 and two kinds of embodiments of step S204, wherein arbitrarily one step of selection executes,
In, step S204 only deletes a kind of fine granularity patch branch of division mode every time, relative to step S203 for subsequent identification
Rate accuracy is higher, and step S203 introduces preset threshold, obtains the more efficient of optimum network structure for subsequent.
Step S205 calculates the discrimination of the network structure after deleting fine granularity patch branch, until certain deletes certain
Fine granularity patch branch corresponding to a division mode causes the discrimination of current network structure to be less than application scenarios discrimination
When demand parameter, then it will retain optimum network knot of the network structure current corresponding to the division mode as current scene
Structure;Otherwise, the network structure re -training remaining fine granularity patch branch and gloable branch formed continues to convergence
Corresponding fine granularity patch branch is deleted, until being matched to the optimum network structure of current scene.
Specifically, if the discrimination for deleting the network structure after fine granularity patch branch is still greater than application scenarios knowledge
Not rate demand parameter when, then return step S202 execute;In above-mentioned steps, fine granularity patch of every corresponding deletion
The division mode of branch requires re -training to the identification for restraining and calculating corresponding network structure after the deletion division mode
Rate, and the discrimination is compared with the discrimination demand parameter of the scene.
In a specific embodiment, when the residual error network structure of redundancy includes a gloable branch and 5 particulates
When spending patch branch, wherein pedestrian image (characteristic pattern) is divided vertically into 2 respectively by 5 fine granularity patch branches, 3,5,7,
9 fine granularity characteristic patterns (if fine granularity patch numbers of branches is more, corresponding odd number number is also more), that is, every
Secondary pedestrian image must be divided vertically into two-part fine granularity characteristic pattern, for the finer situation of subsequent divided,
Global characteristics are expressed to a certain extent.In dividing patch branching process: by multiple odd numbers (3,5,7,9) from essence to thick, according to
Secondary division pedestrian image obtains corresponding fine granularity characteristic pattern;When the training convergence of all branches in the residual error network structure of redundancy
Discrimination afterwards is greater than application scenarios discrimination demand parameter, then start selection pedestrian image is successively divided vertically into 3,5,7,
9 fine granularity characteristic patterns calculate respective character pair weight size in four kinds of patch division modes, if the patch branch
When corresponding feature weight is less than preset threshold, then deletes the fine granularity branch and (be likely to be a patch branch, also having can
Can be several patch branches), illustrate that the fine granularity feature is not important for the data of the scene, attempts it from current
It is deleted in network structure;Alternatively, directly deleting the smallest patch branch of the corresponding feature weight of division mode, also can
Reach same effect;If the discrimination of all branches of current network structure after deleting is still greater than the need of current application scene
Index is sought, then continues to delete patch branch in a manner described;Until certain current net after deleting corresponding manner patch branch
Until network structure is unsatisfactory for the demand parameter of current application scene, then by network corresponding to not deleted corresponding patch branch
Structure is optimum network structure.For example, if being (3,5,7) three kinds of division mode feature weights by 7 vertical division pedestrian images
The smallest patch branch, deleting leads in Exist Network Structure that corresponding discrimination is little after branch's convergence after the patch branch
In application scenarios discrimination demand parameter, then by the network structure of first use thereon (2,3,5,7 fine granularity spy branches and
Gloable branch) optimum network structure as the scene.
In the present embodiment, the division principle of fine granularity patch branch should meet branch's training of whole network structure
Discrimination after convergence is greater than application scenarios discrimination demand parameter, meanwhile, under the premise of meeting discrimination, if deleted
The feature weight of patch branch is less than threshold value, but deletes the discontented toe of the corresponding discrimination of network structure after the patch branch
Mark, then cannot delete the patch branch, then the network structure for not deleting the patch branch is considered as optimum network structure.Pass through
Aforesaid way can quickly match optimum network structure according to current application scene, avoid traditional human factor
It influences, reduces artificial a large amount of artificial experiments to obtain the division mode of effect relatively good fine granularity patch branch, together
When, it also can satisfy the adaptability of the network structure between the biggish different scenes of scene difference, relatively general network structure
For, the convenience of the design to application scenarios network structure is improved, and can make under the premise of meeting precision index
Network structure is obtained to simplify as far as possible.
Step S3 utilizes the feature of the optimum network structure extraction pedestrian image to be measured based on residual error network
Vector;
Wherein, it is based on the resulting optimum network structure of above-mentioned residual error network, being both able to satisfy current application scene Recognition rate needs
Index is sought, and can network structure be simplified as far as possible, that is, simplifies network structure.
Specifically, the multiple regions feature of the pedestrian image to be measured is obtained using the network structure;And obtain to
The corresponding weight of each provincial characteristics for the pedestrian image surveyed;According to the provincial characteristics of the pedestrian image to be measured and
The feature vector of the corresponding Weight Acquisition of the provincial characteristics pedestrian image to be measured, wherein feature vector is to be measured described
The feature vector that pedestrian image is constituted according to specific combination mode, the classifier include various neural network, support vector machines,
Decision tree, nearest neighbor classifier, random forest, Boosting classifier etc., the embodiment of the present invention do not limit classifier specifically
System.
Step S4 identifies that the cosine similarity of the feature vector between pedestrian to be measured and default pedestrian obtains pedestrian and identifies again
As a result.
Referring to Fig. 3, for the process of step S4 in a kind of pedestrian of adaptive network provided by the invention again recognition methods
Figure, comprising:
Step S401 presets the cosine similarity of the feature vector between pedestrian;
Step S402 determines the two when the feature vector cosine similarity of pedestrian to be measured and default pedestrian reach preset value
For same a group traveling together;
Both when the feature vector cosine similarity of step S403, survey pedestrian and default pedestrian are not up to preset value, determine
It is not same a group traveling together.
In the present embodiment, cosine similarity, also known as cosine similarity, COS distance are by calculating two vectors
Included angle cosine value assess the similarity of vector.Cosine similarity for two multiplication of vectors and sentences multiplying for two vector moulds
Product.Cosine similarity is bigger, illustrates that the angle between two vectors is smaller, and two vectors are closer;Conversely, then remoter;It calculates
Cosine similarity between the feature vector of pedestrian image to be identified and the feature vector of default pedestrian, by cosine similarity
The corresponding pedestrian of maximum value determines pedestrian's weight recognition result of pedestrian image to be identified and exports, if its is corresponding remaining
When string similarity reaches preset value, determine that the two is same a group traveling together;If its corresponding cosine similarity is not up to preset value,
Determine that the two is not same a group traveling together.
In the above-described embodiments, the pedestrian detection algorithm detection original image based on deep learning gets pedestrian image,
It is specific as follows:
Pedestrian's convolution feature in original image is extracted using convolutional neural networks.
Wherein, convolutional neural networks are a kind of supervised learning neural networks being made of multiple convolutional layers and full articulamentum,
For extracting validity feature, to validity feature is used in different task, such as scene classification, target detection and image retrieval
Task.
The characteristic pattern number of convolutional layer is specified in netinit, and the size of the characteristic pattern of convolutional layer is by convolution
The size of core and upper one layer of input feature vector figure determine, it is assumed that upper one layer of characteristic pattern size be n*n, convolution kernel size be k*
K, then the characteristic pattern size of this layer is (n-k+1) * (n-k+1).
Specifically, by include in original image pedestrian image be input to convolutional neural networks after, pass through convolutional Neural
The convolutional calculation of network convolutional layer obtains the characteristic pattern of convolutional layer, i.e. pedestrian's convolution feature.
Pedestrian's convolution feature is calculated using the convolutional layer of convolutional neural networks, output convolutional neural networks convolutional layer connects entirely
The feature for connecing layer, obtains target feature vector.
Specifically, pedestrian's convolution feature includes shallow-layer characteristic information and further feature information, and shallow-layer characteristic information refers to volume
The characteristic pattern that the convolutional layer of the prime of product neural network obtains, further feature information refer to the convolution of the rear class of convolutional neural networks
The profile information that layer convolution obtains.
In convolutional neural networks, full articulamentum will be for that will pass through in the characteristics of image figure of multiple convolutional layers and pond layer
Feature is integrated, and is obtained the high-rise meaning that characteristics of image has, is used for image classification later.In a specific embodiment,
The pedestrian's convolution Feature Mapping for the characteristic pattern that convolutional layer generates (is defeated in present embodiment at a regular length by full articulamentum
Enter for the original image classification number in pedestrian's collective database, pedestrian and non-pedestrian (background), i.e., regular length is feature 2)
Vector.This feature vector contains the combined information of input all features of original image, which will contain most in image
The characteristics of image of feature, which is kept down, completes image classification task with this.
Specifically, there are many method that pondization calculates, and the most commonly used is maximum pond (Max Pooling) method and average ponds
Change (mean pooling) method.Wherein, after maximum pond method is using the maximum value of feature graph region as the pool area
Value, that is, take the greatest member value of target signature in each characteristic pattern as Chi Huajie by maximum pond method
Fruit.Average pond method be calculating feature graph region average value as the region pond as a result, for example, feature can be calculated
The average value of some special characteristic of figure.Global pool layer does not need not only to avoid over-fitting, and export to parameter optimization
Pond result relative to position have invariance, the output shared parameter of different location.
Understanding for image the deep information, using convolution feature in convolutional neural networks abstract image target, convolution mind
Subsequent convolutional layer on last stage can be calculated through network after completing feature extraction completion, the output of full articulamentum is directly made
For target feature vector.
Classified using support vector machines to target feature vector, obtains pedestrian image to be measured.
Specifically, data are subjected to the possible division of largest interval using the method for support vector machines, so that classifying quality
It achieves the desired results.It is to detect pedestrian image, therefore only need to determine picture material based on image classification task, meter
It calculates the specific generic numerical value of input picture (generic probability), classification is can be completed into most possible classification output and is appointed
Business, gets testing image.
In the present embodiment, pedestrian's convolution feature in original image is extracted by using convolutional neural networks, using depth
Convolutional neural networks carry out feature extraction, and deep learning can learn from the data of detection image automatically, therefore can be applicable in
A variety of environment improve and obtain detection image adaptability, are conducive to detection image subsequent processing.
Referring to Fig. 4, for a kind of pedestrian's weight identification device structural block diagram of adaptive network provided by the invention, comprising:
Pedestrian detection module 1 extracts pedestrian in pedestrian's video for obtaining pedestrian's video, using pedestrian detection algorithm
Image;
Adaptive network structure 2 is used to adaptively adjust the residual error network structure of redundancy according to current scene to optimum network
Structure;
Wherein, the residual error network structure of the redundancy includes a gloable branch and several fine granularities patch points
Branch, wherein several fine granularity patch branches include two fine granularity patch branches by its division mode and are greater than
2 multiple odd number fine granularity patch branches;
The quantity that fine granularity patch branch in the residual error network structure of redundancy is adaptively adjusted according to current scene, more
A odd number fine granularity patch branch from it is precise and penetrating slightly successively decrease division fine granularity patch branch quantity, by the residual error network
Each branch connects softmax loss function respectively in structure and triplet loss function is trained to convergence;
Judge whether the corresponding patch feature weight of each fine granularity branch is less than preset threshold, if the patch is special
When levying weight less than preset threshold, then the fine granularity branch is deleted;If the patch feature weight is not less than preset threshold
When, then retain the fine granularity branch;Or
It calculates and presses the respective feature weight of multiple odd number division mode corresponding fine granularity patch branch, delete feature
The corresponding division mode of weight the smallest fine granularity patch branch;
The discrimination for calculating the network structure after deleting fine granularity patch branch, until certain deletes some division side
Fine granularity patch branch corresponding to formula causes the discrimination of current network structure to be less than application scenarios discrimination demand parameter
When, then it will retain optimum network structure of the network structure current corresponding to the division mode as current scene;Otherwise, will
The network structure re -training that remaining fine granularity patch branch and gloable branch are formed continues to delete corresponding thin to restraining
Granularity patch branch, until being matched to the optimum network structure of current scene.
Characteristic extracting module 3 is used to utilize the optimum network structure extraction pedestrian to be measured based on residual error network
The feature vector of image;
Wherein, gloable spy is respectively obtained using gloable branch in optimum network structure and fine granularity patch branch
Sign and fine granularity patch feature, fusion gloable feature and fine granularity patch feature obtain the pedestrian image to be measured
Feature vector.
The cosine similarity of the feature vector between pedestrian to be measured and default pedestrian for identification of pedestrian's weight identification module 4 obtains
To pedestrian's weight recognition result.
Wherein, the cosine similarity of the feature vector between pedestrian to be measured and default pedestrian is identified;As pedestrian to be measured and in advance
If the feature vector cosine similarity of pedestrian reaches preset value, determine that the two is same a group traveling together;As pedestrian to be measured and default row
When the feature vector cosine similarity of people is not up to preset value, determine that the two is not same a group traveling together.
In the present embodiment, since the pedestrian of adaptive network again recognition methods and the pedestrian of adaptive network identify dress again
It is set to one-to-one relationship, please refers to above-mentioned side for technical characteristic, technological means and technical effect corresponding in the device
Method will not repeat them here in order to avoid repeating.
Referring to Fig. 5, identifying electronic devices structure block diagram again for the pedestrian that the present invention provides a kind of adaptive network, wrap
It includes:
One or more processors 51;
Memory 52;And
One or more programs, wherein one or more of programs are stored in the memory 52 and are configured as
It is executed instruction by one or more of processors 51, is executed instruction described in one or more of processors execution so that described
Electronic equipment execute as above-mentioned adaptive network pedestrian again recognition methods the step of.
The processor 51 is operationally coupled with memory and/or non-volatile memory device.More specifically, processor
51 can be performed the instruction stored in memory and/or non-volatile memory device to execute operation in calculating equipment, such as
It generates image data and/or image data is transferred to electronic console.In this way, processor may include one or more general micro-
Processor, one or more application specific processor (ASIC), one or more Field Programmable Logic Array (FPGA) or they
Any combination.
The application provides a kind of computer storage medium, and the storage medium includes the program of storage, wherein in the journey
Equipment where controlling the storage medium when sort run executes the pedestrian of adaptive network described in upper item, and recognition methods is deposited again
Storage media.
It, can be with if the function is realized in the form of SFU software functional unit and when sold or used as an independent product
It is stored in a computer readable storage medium.Based on this understanding, the technical solution of the application is substantially in other words
The part of the part that contributes to existing technology or the technical solution can be embodied in the form of software products, the meter
Calculation machine software product is stored in a storage medium, including some instructions are used so that a computer equipment (can be a
People's computer, server or network equipment etc.) execute each embodiment the method for the application all or part of the steps.
In embodiment provided by the present application, the computer-readable storage medium of writing may include read-only memory
(ROM, Read-OnlyMemory), random access memory (RAM, RandomAccessMemory), EEPROM, CD-ROM or
Other optical disk storage apparatus, disk storage device or other magnetic storage apparatus, flash memory, USB flash disk, mobile hard disk or it can be used in
Store any other Jie that there is the desired program code of instruction or data structure form and can be accessed by computer
Matter.In addition, any connection can be properly termed as computer-readable medium.For example, if instruction is using coaxial cable, light
The wireless technology of fine optical cable, twisted pair, digital subscriber line (DSL) or such as infrared ray, radio and microwave etc, from net
Stand, server or other remote sources send, then the coaxial cable, optical fiber cable, twisted pair, DSL or such as infrared ray,
The wireless technology of radio and microwave etc includes in the definition of the medium.
In conclusion the present invention is from the precise and penetrating thick automatic patch branch divided in redundancy residual error network structure, until matching
Best (suitable) network structure under to current scene, reduction carried out repeatedly to obtain optimum network framework engineer with
The huge workload paid is tested, meanwhile, it is capable to the network structure adaptively more simplified under the premise of meeting index,
Erroneous judgement caused by avoiding because of human factor, the free switching being more suitable between multiple scenes extend convenient for subsequent scenario scene.
So the present invention effectively overcomes various shortcoming in the prior art and has high industrial utilization value.
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 (10)
1. a kind of pedestrian of adaptive network recognition methods again, which is characterized in that the described method includes:
Pedestrian's video is obtained, extracts pedestrian image in pedestrian's video using pedestrian detection algorithm;
The residual error network structure of redundancy is adaptively adjusted according to current scene to optimum network structure;
Utilize the feature vector of the optimum network structure extraction pedestrian image to be measured based on residual error network;
Identify that the cosine similarity of the feature vector between pedestrian to be measured and default pedestrian obtains pedestrian's weight recognition result.
2. the pedestrian of adaptive network according to claim 1 recognition methods again, which is characterized in that the basis works as front court
Scape adaptively adjusts the step of residual error network structure to optimum network structure of redundancy, comprising:
The residual error network structure of the redundancy includes a gloable branch and several fine granularities patch branch, wherein if
Dry fine granularity patch branch includes two fine granularity patch branches and multiple surprises greater than 2 by its division mode
Several fine granularity patch branch;
The quantity that fine granularity patch branch in the residual error network structure of redundancy is adaptively adjusted according to current scene, in multiple surprises
Several fine granularity patch branch from it is precise and penetrating slightly successively decrease division fine granularity patch branch quantity, by the residual error network structure
Interior each branch connects softmax loss function respectively and triplet loss function is trained to convergence;
Judge whether the corresponding patch feature weight of each fine granularity branch is less than preset threshold, if the patch feature is weighed
When being less than preset threshold again, then the fine granularity branch is deleted;If the patch feature weight is not less than preset threshold,
Retain the fine granularity branch;Or
It calculates and presses the respective feature weight of multiple odd number division mode corresponding fine granularity patch branch, delete feature weight
The corresponding division mode of the smallest fine granularity patch branch;
The discrimination for calculating the network structure after deleting fine granularity patch branch, until certain deletes some division mode institute
When corresponding fine granularity patch branch causes the discrimination of current network structure to be less than application scenarios discrimination demand parameter,
Optimum network structure of the network structure current corresponding to the division mode as current scene will then be retained;Otherwise, it will remain
The network structure re -training that remaining fine granularity patch branch and gloable branch is formed continues to delete corresponding particulate to restraining
Patch branch is spent, until being matched to the optimum network structure of current scene.
3. the pedestrian of adaptive network according to claim 1 recognition methods again, which is characterized in that described using based on residual
The step of feature vector of the optimum network structure extraction pedestrian image to be measured of poor network, comprising:
Gloable feature and fine granularity are respectively obtained using gloable branch in optimum network structure and fine granularity patch branch
Patch feature, fusion gloable feature and fine granularity patch feature obtain the feature vector of the pedestrian image to be measured.
4. the pedestrian of adaptive network according to claim 1 recognition methods again, which is characterized in that the identification row to be measured
The cosine similarity of feature vector between people and default pedestrian obtains the step of pedestrian's weight recognition result, comprising:
Identify the cosine similarity of the feature vector between pedestrian to be measured and default pedestrian;As the spy of pedestrian to be measured and default pedestrian
When sign vector cosine similarity reaches preset value, determine that the two is same a group traveling together;When pedestrian to be measured and default pedestrian feature to
When amount cosine similarity is not up to preset value, determine that the two is not same a group traveling together.
5. a kind of pedestrian's weight identification device of adaptive network, which is characterized in that described device includes:
Pedestrian detection module extracts pedestrian image in pedestrian's video using pedestrian detection algorithm for obtaining pedestrian's video;
Adaptive network structure, for adaptively adjusting the residual error network structure of redundancy according to current scene to optimum network knot
Structure;
Characteristic extracting module, for utilizing the optimum network structure extraction pedestrian image to be measured based on residual error network
Feature vector;
Pedestrian's weight identification module, the cosine similarity of the feature vector between pedestrian to be measured and default pedestrian is gone for identification
People's weight recognition result.
6. pedestrian's weight identification device of adaptive network according to claim 1, which is characterized in that the adaptive network
Structure further comprises:
The residual error network structure of the redundancy includes a gloable branch and several fine granularities patch branch, wherein if
Dry fine granularity patch branch includes two fine granularity patch branches and multiple surprises greater than 2 by its division mode
Several fine granularity patch branch;
The quantity that fine granularity patch branch in the residual error network structure of redundancy is adaptively adjusted according to current scene, in multiple surprises
Several fine granularity patch branch from it is precise and penetrating slightly successively decrease division fine granularity patch branch quantity, by the residual error network structure
Interior each branch connects softmax loss function respectively and triplet loss function is trained to convergence;
Judge whether the corresponding patch feature weight of each fine granularity branch is less than preset threshold, if the patch feature is weighed
When being less than preset threshold again, then the fine granularity branch is deleted;If the patch feature weight is not less than preset threshold,
Retain the fine granularity branch;Or
It calculates and presses the respective feature weight of multiple odd number division mode corresponding fine granularity patch branch, delete feature weight
The corresponding division mode of the smallest fine granularity patch branch;
The discrimination for calculating the network structure after deleting fine granularity patch branch, until certain deletes some division mode institute
When corresponding fine granularity patch branch causes the discrimination of current network structure to be less than application scenarios discrimination demand parameter,
Optimum network structure of the network structure current corresponding to the division mode as current scene will then be retained;Otherwise, it will remain
The network structure re -training that remaining fine granularity patch branch and gloable branch is formed continues to delete corresponding particulate to restraining
Patch branch is spent, until being matched to the optimum network structure of current scene.
7. pedestrian's weight identification device of adaptive network according to claim 1, which is characterized in that the feature extraction mould
Block further comprises:
Gloable feature and fine granularity are respectively obtained using gloable branch in optimum network structure and fine granularity patch branch
Patch feature, fusion gloable feature and fine granularity patch feature obtain the feature vector of the pedestrian image to be measured.
8. pedestrian's weight identification device of adaptive network according to claim 1 again, which is characterized in that the pedestrian identifies
Module further comprises:
Identify the cosine similarity of the feature vector between pedestrian to be measured and default pedestrian;As the spy of pedestrian to be measured and default pedestrian
When sign vector cosine similarity reaches preset value, determine that the two is same a group traveling together;When pedestrian to be measured and default pedestrian feature to
When amount cosine similarity is not up to preset value, determine that the two is not same a group traveling together.
9. a kind of electronic equipment, which is characterized in that the electronic equipment includes:
One or more processors;
Memory;And
One or more programs, wherein one or more of programs are stored in the memory and are configured as by described
One or more processors execute instruction, and execute instruction described in one or more of processors execution so that the electronic equipment
Execute pedestrian's recognition methods again of the adaptive network as described in any one of Claims 1 to 4.
10. a kind of computer readable storage medium, the computer-readable recording medium storage has computer program, and feature exists
In pedestrian's weight of realization adaptive network as described in any one of Claims 1-4 when the computer program is executed by processor
The step of recognition methods.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910159239.6A CN109871821B (en) | 2019-03-04 | 2019-03-04 | Pedestrian re-identification method, device, equipment and storage medium of self-adaptive network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910159239.6A CN109871821B (en) | 2019-03-04 | 2019-03-04 | Pedestrian re-identification method, device, equipment and storage medium of self-adaptive network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109871821A true CN109871821A (en) | 2019-06-11 |
CN109871821B CN109871821B (en) | 2020-10-09 |
Family
ID=66919605
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910159239.6A Active CN109871821B (en) | 2019-03-04 | 2019-03-04 | Pedestrian re-identification method, device, equipment and storage medium of self-adaptive network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109871821B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110688897A (en) * | 2019-08-23 | 2020-01-14 | 深圳久凌软件技术有限公司 | Pedestrian re-identification method and device based on joint judgment and generation learning |
CN110807434A (en) * | 2019-11-06 | 2020-02-18 | 威海若维信息科技有限公司 | Pedestrian re-identification system and method based on combination of human body analysis and coarse and fine particle sizes |
CN111814857A (en) * | 2020-06-29 | 2020-10-23 | 浙江大华技术股份有限公司 | Target re-identification method, network training method thereof and related device |
CN111814705A (en) * | 2020-07-14 | 2020-10-23 | 广西师范大学 | Pedestrian re-identification method based on batch blocking shielding network |
CN112329835A (en) * | 2020-10-30 | 2021-02-05 | 天河超级计算淮海分中心 | Image processing method, electronic device, and storage medium |
CN112817725A (en) * | 2021-02-06 | 2021-05-18 | 成都飞机工业(集团)有限责任公司 | Microservice partitioning and optimizing method based on efficient global optimization algorithm |
CN115631509A (en) * | 2022-10-24 | 2023-01-20 | 智慧眼科技股份有限公司 | Pedestrian re-identification method and device, computer equipment and storage medium |
EP4136580A4 (en) * | 2020-06-29 | 2024-01-17 | Zhejiang Dahua Technology Co., Ltd | Target re-identification method, network training method thereof, and related device |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
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 |
CN109101913A (en) * | 2018-08-01 | 2018-12-28 | 北京飞搜科技有限公司 | Pedestrian recognition methods and device again |
-
2019
- 2019-03-04 CN CN201910159239.6A patent/CN109871821B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
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 |
CN109101913A (en) * | 2018-08-01 | 2018-12-28 | 北京飞搜科技有限公司 | Pedestrian recognition methods and device again |
Non-Patent Citations (2)
Title |
---|
COLDPLAYHU: "模型压缩总览", 《HTTPS://WWW.JIANSHU.COM/P/E73851F32C9F》 * |
GUANSHUO WANG等: "Learning Discriminative Features with Multiple Granularities for Person Re-Identification", 《PROCEEDINGS OF THE 26TH ACM INTERNATIONAL CONFERENCE ON MULTIMEDIA》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110688897A (en) * | 2019-08-23 | 2020-01-14 | 深圳久凌软件技术有限公司 | Pedestrian re-identification method and device based on joint judgment and generation learning |
CN110807434A (en) * | 2019-11-06 | 2020-02-18 | 威海若维信息科技有限公司 | Pedestrian re-identification system and method based on combination of human body analysis and coarse and fine particle sizes |
CN110807434B (en) * | 2019-11-06 | 2023-08-15 | 威海若维信息科技有限公司 | Pedestrian re-recognition system and method based on human body analysis coarse-fine granularity combination |
CN111814857A (en) * | 2020-06-29 | 2020-10-23 | 浙江大华技术股份有限公司 | Target re-identification method, network training method thereof and related device |
CN111814857B (en) * | 2020-06-29 | 2021-07-06 | 浙江大华技术股份有限公司 | Target re-identification method, network training method thereof and related device |
EP4136580A4 (en) * | 2020-06-29 | 2024-01-17 | Zhejiang Dahua Technology Co., Ltd | Target re-identification method, network training method thereof, and related device |
CN111814705A (en) * | 2020-07-14 | 2020-10-23 | 广西师范大学 | Pedestrian re-identification method based on batch blocking shielding network |
CN111814705B (en) * | 2020-07-14 | 2022-08-02 | 广西师范大学 | Pedestrian re-identification method based on batch blocking shielding network |
CN112329835A (en) * | 2020-10-30 | 2021-02-05 | 天河超级计算淮海分中心 | Image processing method, electronic device, and storage medium |
CN112817725A (en) * | 2021-02-06 | 2021-05-18 | 成都飞机工业(集团)有限责任公司 | Microservice partitioning and optimizing method based on efficient global optimization algorithm |
CN112817725B (en) * | 2021-02-06 | 2023-08-11 | 成都飞机工业(集团)有限责任公司 | Micro-service division and optimization method based on efficient global optimization algorithm |
CN115631509A (en) * | 2022-10-24 | 2023-01-20 | 智慧眼科技股份有限公司 | Pedestrian re-identification method and device, computer equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN109871821B (en) | 2020-10-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109871821A (en) | The pedestrian of adaptive network recognition methods, device, equipment and storage medium again | |
Yang et al. | Real-time face detection based on YOLO | |
CN106960195B (en) | Crowd counting method and device based on deep learning | |
Rachmadi et al. | Vehicle color recognition using convolutional neural network | |
WO2020164282A1 (en) | Yolo-based image target recognition method and apparatus, electronic device, and storage medium | |
CN104463117B (en) | A kind of recognition of face sample collection method and system based on video mode | |
CN109711366B (en) | Pedestrian re-identification method based on group information loss function | |
CN109447169A (en) | The training method of image processing method and its model, device and electronic system | |
CN109978918A (en) | A kind of trajectory track method, apparatus and storage medium | |
CN109918969A (en) | Method for detecting human face and device, computer installation and computer readable storage medium | |
CN105303150B (en) | Realize the method and system of image procossing | |
CN107679448A (en) | Eyeball action-analysing method, device and storage medium | |
CN109002755B (en) | Age estimation model construction method and estimation method based on face image | |
CN111178208A (en) | Pedestrian detection method, device and medium based on deep learning | |
CN109919246A (en) | Pedestrian's recognition methods again based on self-adaptive features cluster and multiple risks fusion | |
CN108921057A (en) | Prawn method for measuring shape of palaemon, medium, terminal device and device based on convolutional neural networks | |
Sapijaszko et al. | An overview of recent convolutional neural network algorithms for image recognition | |
CN109508675A (en) | A kind of pedestrian detection method for complex scene | |
CN109815814A (en) | A kind of method for detecting human face based on convolutional neural networks | |
CN108073940B (en) | Method for detecting 3D target example object in unstructured environment | |
CN109670517A (en) | Object detection method, device, electronic equipment and target detection model | |
CN110070106A (en) | Smog detection method, device and electronic equipment | |
CN113033468A (en) | Specific person re-identification method based on multi-source image information | |
Ballotta et al. | Fully convolutional network for head detection with depth images | |
CN116385430A (en) | Machine vision flaw detection method, device, medium and equipment |
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 |