CN107832672A - A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information - Google Patents
A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information Download PDFInfo
- Publication number
- CN107832672A CN107832672A CN201710946443.3A CN201710946443A CN107832672A CN 107832672 A CN107832672 A CN 107832672A CN 201710946443 A CN201710946443 A CN 201710946443A CN 107832672 A CN107832672 A CN 107832672A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- mrow
- msub
- information
- loss
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information, this method can effectively solve that monitor video one skilled in the art is blocked frequently, video light differential is big and non-rigid pedestrian's posture it is changeable caused by it is difficult, there is extensive use in fields such as safety monitorings.This method is broadly divided into two stages, is off-line phase and on-line stage respectively.Wherein off-line phase is responsible for the deep learning network model of training study high accuracy, the stage includes pretreatment, artis information extraction, extraction local feature and carries out Fusion Features with the global characteristics of core network framework extraction, and the characteristic use five-tuple loss function of fusion finally is completed into training.On-line stage then carries out feature extraction using the deep learning network model trained, so as to again the pedestrian for realizing target to be analyzed by Similarity Measure and having stored between Target Photo storehouse identifies.
Description
Technical field
The invention belongs to technical field of computer vision, more specifically, is related to a kind of designed using attitude information and damages more
Pedestrian's recognition methods again of function is lost, is that one kind can resist that pedestrian is blocked, posture is changeable and can be applied to intelligent monitoring point
The accurate pedestrian recognition methods again of analysis system.
Background technology
Pedestrian's weight identification technology refers to retrieve in multiple cameras to setting the goal, and retrieval result is associated into matching.
The technology is field of video monitoring, such as pedestrian retrieval, and across camera tracking, man-machine interaction etc., which is applied, provides infrastructural support.For sea
The personage of the video data of amount searches task, and pedestrian identifies again can greatly liberate manpower.Yet with camera shooting visual angle
It is different, illumination condition is complicated, the reason such as changeable of blocking frequent, non-rigid pedestrian's posture is so that pedestrian identifies that problem extremely has again
Challenge.To overcome these difficult, researchers proposed many different solutions in past 20 years.According to algorithm
Principle can substantially be divided into design expression characteristic and optimization distance measures two class algorithms.
Design expression characteristic refers to find the feature for changing robust to picture appearance.The method concern of feature based expression is such as
What design with to pedestrian with identification and to image change with stability feature description.Including color histogram, line
Manage the low-level visual features such as feature, local feature region and the middle level features with semantic attribute.
In order to effective utilization space information, existing method generally divides an image into different regions, and such as 2006
Year pedestrian image is all once divided into multiple horizontal strips from top to bottom to Zheng Wei poems in 2013 et al..Farenzena in 2010 etc.
Using image symmetrical and asymmetric priori theoretical, pedestrian image is divided into human head, trunk and leg three parts, with extraction
Combinations of features between different zones.Have benefited from big data quantity pedestrian weight identification data collection Marker1501, MARS appearance, grind
The persons of studying carefully begin to use the method based on deep learning to represent characteristics of image.Cheng in 2016 et al., which is proposed, a kind of is based on office
The deep neural network framework of the multichannel of portion's block, they using partial strip band and the artwork of horizontal division extract simultaneously it is global with
Local feature.Yet with the change of different cameral visual angle and pedestrian's posture, horizontal segmentation can produce to align by mistake, on the contrary can shadow
Ring the accuracy rate of model.Based on considerations above, the present invention is detected to obtain accurate local location, reached using pedestrian's artis
To the alignment based on semanteme, key condition is provided for the fusion of global and local feature complementary.
Optimization distance measurement refers to learn a kind of metric space so as to belong to characteristic distance between the image of same people close,
Characteristic distance is remote between belonging to the image of different people.Weinberger in 2009 et al. proposes large-spacing arest neighbors classification (large
Margin nearest neighbour, LMNN), employing triple constraint makes in new metric space, k of each sample
Nearest neighbours belong to same class.2012, Kostinger et al. proposed to keep simple directly (Keep is Simple
And Straight, KISS) learning distance metric algorithm.Hereafter scholar gradually is combined distance metric with deep learning
Come, establish checking model progress pedestrian and identify again.The class model with image to for network inputs, after characteristics of image is extracted simultaneously
Calculate the distance between feature, the similarity between final output image.Extraction feature and similarity measurement are integrated in one
It is the main advantage of the class model in framework.But using only checking model, the otherness that can only be extracted between picture pair is special
Sign.The significant characteristics having per pictures itself are often ignored.Therefore the present invention considers joint classification model and checking mould
Type is trained, while calculates Classification Loss and checking loss, and both are weighted to reach model complementary.
As deep learning is commonly used on the multiple subproblems of computer vision field, what Wei et al. was proposed is answering
Accurately the method for extraction artis information identifies that accurate local message obtains for pedestrian and provides possibility again in miscellaneous scene.Consider
Certain rule is presented in pedestrian's data attitudes vibration in monitor video, abnormal posture seldom occurs, based on deep learning method
The algorithm for automatically extracting artis information may apply to pedestrian and identify again in problem.Therefore, the present invention is obtained using this method
Joint information calculate human body local location and speculate pedestrian's posture direction, wherein local location information can be used for extraction office
Portion's feature merges with global characteristics, and for pedestrian's posture towards can be used for design five-tuple loss function, these information can be multiple
The degree of accuracy that pedestrian identifies again is improved under miscellaneous monitors environment.
The content of the invention
The purpose of the present invention is：Overcome the deficiencies in the prior art, there is provided one kind designs more loss functions using attitude information
Pedestrian's recognition methods again, cope with monitor video one skilled in the art block frequently, light differential is big, non-rigid pedestrian's posture is changeable
Situations such as, and it is desirably integrated into the fundamental analysis realized in any intelligent monitor system to pedestrian.
The technical solution adopted by the present invention is：It is a kind of to design the pedestrians of the more loss functions side of identification again using attitude information
Method, including：Offline extraction character network model training, online pedestrian identify two parts main contents again；
Step (1), offline extraction character network model training stage：
(m1) all pictures are pre-processed, original image rIiI is used after processingiRepresent；
(m2) for each pictures detection artis information, there is P in 18 obtained artis informationIi={ x1,
y1... ..., x18, y18In, and there is corresponding Boolean type array label to indicate whether to detect different artis, labeli
=(True or False)；
(m3) according to the artis information extracted in step (m2), thus it is speculated that the height high of each pedestriani, respectively calculate
The local region information of head-trunk-leg
(m4) the artis information extracted according to step (m2), thus it is speculated that the posture direction of pedestrian target, be designated as diri=
(1or2or 3), wherein representing positive sample when equal to 1,2 represent lateral samples, and 3 represent backwards to sample；
(m5) global characteristics are extracted according to the core network of design, the regional area positional information extracted according to step (m3)
And branching networks structure extraction local feature, and by the global characteristics and Local Feature Fusion of every pictures, it is collectively forming statement
Property characteristic vector；
(m6) more Classification Loss functions and the first weight triple loss letter of the present invention are calculated according to data true tag
Number, while the second heavy five-tuple loss function towards design five-tuple and is calculated according to pedestrian's posture that step (m4) speculates；
(m7) a variety of loss function errors training current signature extraction network that joint step (m6) calculates, and analyze difference
Influence of the loss function weight to network, select optimal weight λ1And λ2To complete joint training；
Step (2), online pedestrian weight cognitive phase：
(s1) all picture I in picture library are pre-processedgallery, and the network for training to obtain using step (1) off-line phase
Model carries out feature extraction, and identification information stores to form feature database F one by one according to corresponding to picturegallery；
(s2) picture I to be analyzed is pre-processedquery, using step (1) off-line phase train obtained network model carry out it is special
Sign extraction, final characteristic vector fqueryUnique effective information as subsequent step (s3) similarity measurement；
(s3) f extracted in calculation procedure (s2)queryWith feature database FgalleryBetween characteristic distance, and carry out normalizing
Change, sorting operation, therefrom select similarity more than 0.7 and the retrieval result that is identified again as pedestrian in preceding M picture of ranking, its
Middle M numerical value is selected according to Number dynamics in photo current storehouse；
(s4) picture library and its corresponding feature database are regularly updated, emphasis is supported static library and detected by dynamic video storehouse
The dynamic base both modalities which collected.
Further, described step (m3) comprises the following steps：
(m3.1) the artis information P extracted according to step (m2)Ii, remove labeliFalse is all, that is, detects artis
The sample of failure；Removal represents the sample that torso portion artis is largely False；
(m3.2) step (m2) extraction artis information meets the sample participation training of application claims, according to existing joint
Point information PIiSpeculate pedestrian's height highi；
(m3.3) according to artis such as left-right ear, noses, head zone information is calculated
(m3.4) according to right and left shoulders, left and right hip associated joint dot position information, torso area information is calculated
(m3.5) according to information such as waist location, ankle, heights, leg area information is calculatedDue to
Detect obtained encirclement frame and often do not include foot, therefore according to the proportional downward scaling of height；
(m3.6) the regional area positional information being calculated according to step (m3.3) to (m3.5) produces interest region, profit
With improved region of interest feature extraction layer, local shape factor is carried out into branching networks.
Further, described step (m4) comprises the following steps：
(m4.1) after being screened according to step (m3.1), it is determined that its posture direction of the sample analysis of participation training, without left shoulder
Or the sample of right shoulder is determined as lateral diri=2；
(m4.2) the existing sample of the right shoulder of left shoulder, then right and left shoulders vector is calculated
(m4.3) the right and left shoulders vector obtained according to step (m4.2)Calculate the angle dir_angle with vertical curveIi；
(m4.4) the angle dir_angle that judgment step (m4.3) is calculatedIiAffiliated angular range, as angle dir_
angleIiThen bearing mark is positive dir in scope [260 °, 280 °]i=1, if not whether positive then judge angle in scope
In [80 °, 100 °], this is labeled as backwards to dir within the rangei=3, if not in the range of above-mentioned two, mark the sample
This is lateral diri=2.
Further, described step (m5) comprises the following steps：
(m5.1) the core network extraction global characteristics in network frame proposed by the present invention, labeled as fglobal
(Ii)；
(m5.2) three local feature flocal (I of Connection Step (m3.6) extractioni), it is respectively labeled as fh (Ii)、
ft(Ii)、fl(Ii)；
(m5.3) global characteristics and the part of step (m5.2) extraction of step (m5.1) extraction are realized using full articulamentum
Feature f (Ii)。
Further, described step (m6) comprises the following steps：
(m6.1) more Classification Loss function errors are calculated；
(m6.2) the first weight triple constraint of the present invention is calculated：
Did(Ii a,Ii p,Ii n)=d (f (Ii a)-f(Ii p))-d(f(Ii a)-f(Ii n)) ＜ α
Wherein, IaIt is any one benchmark pedestrian image, I in data seti pRepresent to represent the another of same people with benchmark pedestrian
One image, i.e. positive sample, Ii nFor other people image, i.e. negative sample, triple input obtains each after network calculations
From characteristic vector { f (Ii a),f(Ii p),f(Ii n), d (f (Ii a)-f(Ii p)) on the basis of figure with positive sample to the distance between,
d(f(Ii a)-f(Ii n)) on the basis of figure with negative sample to the distance between, α be triple constraint threshold value；
(m6.3) the second weight five-tuple constraint of the present invention is calculated：
Dpose(Ii a,Ii ps,Ii pd)=d (f (Ii a)-f(Ii ps))-d(f(Ii a)-f(Ii pd)) ＜ β
Wherein,Represent withPosture identical positive sample,Represent withThe different positive sample of posture, β are five-tuple
The threshold value of double constraint.
Further, described step (m7) comprises the following steps：
(m7.1) the joint error value of the more loss function error calculation backpropagations obtained according to step (m6)：
Loss3(I, w)=λ1Loss1(I,w)+λ2Loss2(I,w)
Wherein, Loss1Represent more Classification Loss functions, Loss2Represent five-tuple loss function, Loss3Represent associated losses
Function, λ1And λ2To balance the weighted value of associated losses function, λ is balance triple and the weighted value of five-tuple constraint, and w is net
Network parameter,Represent the probability of prediction, piIt is destination probability, N is pedestrian's number of species, and n is five-tuple quantity.
(m7.2) Error weight parameter lambda in analytical procedure (m7.1)1And λ2Selection, determine that off-line phase uses optimal
Loss function distributes weight.
Further, described step (s3) comprises the following steps：
(s3.1) M numerical value is selected according to Number dynamics in photo current storehouse；
(s3.2) f extracted successively in calculation procedure (s2)queryWith feature database FgalleryBetween characteristic distance；
(s3.3) all characteristic distances that step (s3.2) is calculated are normalized, sorting operation, therefrom selected
Similarity is more than 0.7 and the retrieval result that is identified again as pedestrian in preceding M picture of ranking；
(s3.4) pedestrian that visualization step (s3.3) obtains identifies retrieval result again, is then shown for static map valut
IqueryAnd the I after sequenceresults, for dynamic video storehouse then according to IresultsIn the camera ID of database purchase, pedestrian ID, bag
Peripheral frame positional information and frame number time etc. recover the truth of result in video at that time.
Further, described step (s4) comprises the following steps：
(s4.1) the time t regularly updated is set；
(s4.2) in the range of time t, inquiry picture I is then constantly added for static map valutqueryInformation and spy
Sign；Reach t and change or update picture library as requested, extract the picture feature of change again, set up new feature database；
(s4.3) in the range of time t, the fresh target detected is then continuously added for dynamic video storehouse, and in database
The world informations such as middle storage camera ID, pedestrian ID, encirclement frame positional information, frame number, time, place；Reach then basis after t
Time, half pedestrian's data message in current storehouse is emptied, then adds new testing result frame by frame, while extract its feature
A primary attribute as deposit database.
The principle of the present invention is：
The present invention, which proposes, a kind of to be calculated a variety of loss functions using human body attitude information and knows again come the pedestrian of learning characteristic
Other method.It is pedestrian that the design of the invention comes from monitoring camera, the increase of bayonet socket quantity and the enhancing of storage capacity first
Big data provides resource guarantee.Different magnitude of pedestrian's data provide good for pedestrian's weight identification technology based on deep learning
Good data basis.Secondly, the present invention considers the rule of presentation of the pedestrian in monitor video, and every is schemed in pretreatment stage
Piece carries out the ratio of width to height adjustment, can possess good spatial information spy during feature to be extracted in successive depths network frame
Sign.In addition, the present invention introduces local feature and come more to handle the situation that background in monitor video is noisy, frequently blocks
Mend the deficiency of global characteristics.Introduce artis information and calculate the position of pedestrian's regional area corresponding thereto in camera towards appearance
State.Further according to the local feature of local location information extraction human body, merged with global characteristics.It is last present invention further contemplates that from
The expressive faculty of deep learning network model is improved on Training strategy, five-tuple loss function is designed using orientation information, with
Cross entropy loss function joint completes training, so as to obtain efficient, robust Feature Selection Model in off-line phase.
In face of the huge data volume of monitor video, being accomplished manually pedestrian and identifying again has become unrealistic.At automation
Pedestrian's weight identification technology of reason will promote the development of many applications such as video analysis, security protection.Effect is identified again in view of artificial pedestrian
The main reason for rate is low is that destination number to be analyzed is more, in human brain can not the target largely watched of short time memory storage it is special
Sign.It is that this is of the invention after Feature Selection Model is obtained, devises the online technology path completed pedestrian and identified again.This
During, first have to timing and existing picture library is updated according to monitoring content, and its correlated characteristic is prefetched, to shorten retrieval time.
Then after target to be analyzed is obtained, Rapid matching is carried out to it, necessary pedestrian is completed and identifies again.
Specifically, a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information of the present invention, this method point
For offline and online two stages.In off-line phase, present invention firstly provides a kind of feature extraction for keeping pedestrian's the ratio of width to height
Depth network frame；Secondly introduce artis information and calculate the position of pedestrian's regional area corresponding thereto in camera towards appearance
State；Afterwards according to the local feature of local location information extraction human body, merged with global characteristics；Finally utilize orientation information
Design five-tuple loss function, and with cross entropy loss function joint training.In on-line stage, the present invention is first with offline
The Feature Selection Model that stage-training obtains carries out feature extraction to pretreated picture library, and stores；Secondly to be analyzed
Target Photo carries out the ratio of width to height adjustment, and feature is extracted after adjustment；For the target signature to be analyzed of extraction, in the storage of picture library
Measuring similarity is carried out in feature, sequence is normalized in the similitude of calculating, and therefrom selection meets similarity condition and ranking
Picture is as retrieval result in forward storehouse；The information such as the camera for matching these retrieval results, ID are finally integrated, to visualize mould
Formula is exported, while is stored in inquiry storehouse, and input is provided for other application analysis.In addition, for monitor video or existing adopt
For pedestrian's data of collection, picture library of the present invention and its feature need to be regularly updated, most accurate to ensure to obtain
Pedestrian weight recognition result.
In off-line phase, comprise the following steps that：
First, the present invention is pre-processed all pictures (to be analyzed picture, picture library), and its ratio of width to height is adjusted to 1:
2,107*219 is resized to before training is entered, to ensure that the feature extraction phases in next step can retain effectively
Spatial information, while also play the effect for reducing network parameter.
Secondly, network structure proposed by the present invention forms for several under：The artis detection network of 1 preset parameter, 1
Individual core network, 3 localized branches networks, the articulamentum and two loss layers of 3 fusion features.Artis detection network carries
For pedestrian's artis information.Core network, branching networks and articulamentum are responsible for extracting global characteristics and local feature, and by two
Person is merged.Loss layer is responsible for combining two kinds of loss functions and carries out metric learning.
Wherein, artis detection network mainly extracts 18 artis of human body, including neck, nose, and right and left shoulders,
Elbow, wrist, knee, ankle, hip, shoulder, eyes, ear, it is allowed to which partial joint point is lost.After obtaining each body joint point coordinate, the present invention utilizes
It estimates pedestrian's height of pedestrian, and using height as auxiliary, area is calculated by the max min of regional joint coordinates
Domain border, extract local feature for subsequent network and positional information is provided.
At the same time, the present invention also speculates pedestrian's posture direction using obtained artis information.For artis information
Failure is obtained, or the pedestrian target of no torso portion artis is abandoned, and is not instructed using these inferior samples
Practice, in order to avoid its contamination characteristics extraction model.Left shoulder/or right shoulder joint are then preferentially judged whether for the sample for participating in training
Point, first lock the lateral sample of standard.Then left shoulder is calculated to the direction vector of right shoulder, the angle between the vector and vertical curve
As the chief argument towards differentiation.Mark of the angle in the range of [80 °, 100 °] is sample, angle [260 °,
280 °] in mark be sample.
Again, core network, branching networks and the local region information extraction pedestrian designed using the present invention is identified again
The global characteristics and local feature needed.Wherein core network structure, uses for reference inception_v3 thought, but is different from
Expenditure is that the structure of the present invention is interior and includes 5 kinds of convolution modules, and each module has multiple branches, and each branch is by more
The convolutional layer and pond layer of kind yardstick, which stack, to be formed.Such structure can increase the width of network and reduce network ginseng
Number, while can also strengthen the adaptability to yardstick.The Web vector graphic ReLU of the present invention introduces non-thread sexuality, and each
All accelerate to restrain and slow down the influence that parameter distribution changes using batch regularization before ReLU, in last full connection
There is provided 50% Dropout for layer to prevent over-fitting.Branching networks structure shares the ginseng before conv5_x with core network
Number.It is responsible for the local feature in the respective region of extraction after the layer coal addition position information of pond.Branching networks and core network structure
It is similar, the difference is that the output number of last pond layer and full articulamentum is less than core network, play the work of regulation weight
With.Local feature and global characteristics are incorporated as the characteristic vector of pedestrian using full articulamentum by network backend.
Finally, the present invention observes to obtain pedestrian's similarity rule according to long-term experiment, i.e., the feature between different pedestrians away from
From more than the characteristic distance between identical pedestrian, it is identical that the characteristic distance between same person difference posture is also more than same person
Characteristic distance between posture.According to the rule, the present invention devises the five-tuple loss function of double constraint, and proposes to make
With the loss function and the strategy of more Classification Loss function joint trainings.New loss function can correct network and think and posture
The more like cognition mistake of the identical negative sample positive sample outward appearance more different than posture, fundamentally allow the network to study to gram
Take the expressing feature of attitudes vibration.And the strategy of joint training, then it can increase network in the case where not changing network structure
Expressive faculty, so as to get network model possess more preferable migration, identify spy again so as to obtain the required pedestrian of the present invention
Sign extraction network model.
After off-line phase obtains the feature extraction network model that above-mentioned steps obtain, on-line stage carries out pedestrian and identified again,
Comprise the following steps that：
First, whole pictures of picture library are pre-processed, adjust to meet Feature Selection Model input unified size, then
Feature extraction is carried out to pretreated picture library using the Feature Selection Model that off-line phase training obtains, and it is crucial according to it
Information slitting stores characteristic vector, forms feature database；
Secondly, Target Photo to be analyzed is pre-processed and extracts the characteristic vector with expressive faculty；
Again, for the target feature vector to be analyzed of extraction, measuring similarity is carried out in the characteristic vector storehouse of storage.
And sequence is normalized in the similitude to being calculated, therefrom selection meets picture in similarity condition and storehouse in the top
As retrieval result；
Finally, the information such as the camera for matching these retrieval results, ID are integrated, are exported with visualization formulation, while deposit is looked into
Storehouse is ask, input is provided for other application analysis.In addition, for for monitor video or pedestrian's data of existing collection,
Picture library of the present invention and its characteristic vector storehouse need to be regularly updated, and to ensure to obtain, most accurately pedestrian identifies again
As a result.
The advantage of the present invention compared with prior art is：
1st, the present invention proposes a deep neural network framework being made up of master network and three sub-networks, and master network is used
In extraction global characteristics, three sub-networks utilize the local feature of artis information extraction human body head, trunk and leg.Most
Afterwards by global and local Fusion Features to improve the accuracy rate of retrieval, and it is effective against frequently blocking in monitor video.
2nd, using the pedestrian's orientation information design five-tuple constraint deduced, metric learning ability is strengthened, and use
Joint classification loses and the strategy for carrying out training network model is lost in checking.Ensure to belong to feature between the image of same person away from
From less than the characteristic distance between the image for belonging to different people, the characteristic distance of image is less than between the identical posture of same people
Characteristic distance between different pose presentations.Changeable identified again to pedestrian of non-rigid pedestrian target posture is fundamentally overcome to bring
Difficulty.
3rd, the present invention is separated from each other with conventional analysis modules such as detection, tracking.It can be used as standalone module is integrated to take office
Anticipate in intelligent monitor system, accurate input, easy to use, model robust are provided for upper layer analysis.
Brief description of the drawings
Fig. 1 is that the pedestrian proposed by the present invention using the more loss functions of attitude information design again illustrate by the overall of recognition methods
Figure；
Fig. 2 is that the present invention speculates that pedestrian's direction, calculating local region information extract head, trunk, leg using attitude information
Portion's local fine feature schematic diagram；
Fig. 3 is present invention extraction artis and regional area positional information and the contrast of conventional strip belt regional area division
Exemplary plot, wherein first group is artwork, second group is extraction effect of the present invention, and the 3rd group is that strip-type divides effect, contrast the
Two groups and the 3rd group of picture can be found that the method for the present invention can more effectively align pedestrian target regional area, and can
Exclusive segment ambient interferences；
Fig. 4 is to speculate pedestrian target posture towards flow chart；
Fig. 5 speculates that pedestrian's posture towards exemplary plot, is generally divided into method for the present invention using artis information
Three kinds of side, front, reverse side directions；
Fig. 6 is the principle schematic of present invention design five-tuple loss function.
Embodiment
Specific steps of the present invention are described in detail with reference to the accompanying drawings and examples.
The present invention proposes a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information, combines figure first
Pedestrian's weight identification processing procedure of the present invention is discussed in detail in 1 general illustration.The present invention includes offline and online two stages, its
Middle off-line phase includes pretreatment, rough features extraction, fine-feature extraction, Fusion Features, five-tuple similarity measurement, multiclass
The parts such as loss function calculates and network parameter learns；On-line stage then includes feature extraction, similarity measurement, picture library more
New and four parts of result visualization.
Stage (1) off-line phase：The network model of training study extraction feature.
A. the step of data prediction is：Notice in actual video, pedestrian's encirclement frame is mostly rectangle, and the ratio of width to height is about
0.5.And existing most of network inputs that recognition methods uses again of the pedestrian based on deep learning are mostly square, it is unfavorable for protecting
Hold the space characteristics of pedestrian.Therefore, the input size of network is changed to 107x219 by the present invention, with pedestrian image actual aspect ratio
It is consistent, is advantageous to validity feature extraction, while also play the effect for reducing network parameter.For image list LI1,I2,…,
InIn any one pictures IiCarry out above-mentioned pretreatment.
B. rough features extract the step of be：Mainly by core network, branching networks in the network architecture that the present invention designs
And more several pith compositions of loss function layer.The part of core network and branching networks before Conv5_x uses altogether
The network parameter enjoyed, the part are mainly responsible for the rough features of extraction picture, the information such as semanteme are included in these features, close to
Global characteristics, therefore be also the foundation characteristic of global characteristics extraction.With picture IiExemplified by, the characteristic pattern of the extracting section is labeled asIt is the input of next part fine-feature extraction.
C. fine-feature extract the step of be：After rough features are obtained, core network further extracts fine-feature, point
Then local shape factor schematic diagram according to Fig. 2 completes the extraction of local fine feature to branch network.Idiographic flow is as follows：
1) pedestrian's artis is detected：As shown in Fig. 2 carry out artis detection for pretreated picture.The present invention is main
Extract 18 artis (allowing to lose) of human body, including neck, nose, and right and left shoulders, elbow, wrist, knee, ankle, hip, shoulder,
Eyes, ear.Artis detection failure, the sample that there is no trunk information are not involved in training.Qualified sample is obtaining
To after each body joint point coordinate, the positional information of right and left shoulders can be predicted as 2) pedestrian's posture towards all Main Basiss speculated
The auxiliary that height can then be extracted as 3) local region information.
2) pedestrian's posture direction is speculated：As shown in figure 4, remove detection artis failure and without the sample of trunk after,
Observation right and left shoulders whether there is, so that it is determined that the lateral sample of standard.Shoulder vector is carried out for the existing sample of right and left shoulders
Calculate, obtained vector carries out angle calcu-lation with vertical curve, and according to angular range, it is specially positive, backwards to still to judge sample
Laterally.Pedestrian's posture that the inventive method speculates is towards exemplary plot, as shown in Figure 5.
3) pedestrian's local region information is calculated：The regional area division of conventional strip belt can not only exclude complex background
Interference, when in face of example as Fig. 3, it can not realize that local characteristic region aligns, this error can drive network mould
The feature of type study mistake.Therefore the present invention first estimates the height of pedestrian, supplemented by height after each body joint point coordinate is obtained
Help, by the max min zoning border of regional joint coordinates, extract local feature for subsequent network and provide
Positional information.Equally with picture IiExemplified by explain that the formula in the present invention is expressed, three local region informations are expressed as respectivelyEach regional area positional information is by four-tuple
(xi,yi,wi,hi) composition, wherein x, y, w, h represent the length and height of region top left co-ordinate (x, y) and the region respectively.
4) local fine feature is extracted：After 3) local region information is obtained, the branched network in network structure is used
Network extracts local fine feature.The element branches network parameter is not shared.
D. Fusion Features mode is as follows：Experimental analysis of the present invention various features amalgamation modes, focusing on comparative
ElementWise, Concat two ways.As a result show：Meet design principle of the present invention, global characteristics are complementary with local feature
Concat modes can obtain maximally efficient characteristic vector.
E. the design of five-tuple is as follows with forming step：Five-tuple loss letter is to add posture about by triple loss function
Beam improves.Conventional triple loss function is mathematically represented as a triple { Ii a,Ii p,Ii n, wherein IaIt is data set
In any one benchmark pedestrian image, Ii pExpression represents another image of same people, i.e. positive sample, I with benchmark pedestriani nFor
Other people image, i.e. negative sample.Triple input obtains respective characteristic vector { f (I after network calculationsi a),f
(Ii p),f(Ii n), then there is triple constraint：
Did(Ii a,Ii p,Ii n)=d (f (Ii a)-f(Ii p))-d(f(Ii a)-f(Ii n)) ＜ α
Wherein, d (f (Ii a)-f(Ii p)) on the basis of figure with positive sample to the distance between, d (f (Ii a)-f(Ii n)) on the basis of
Figure with negative sample to the distance between, α be triple constraint threshold value.The meaning of the inequality is a kind of measurement of study, at this
In metric space, the characteristic distance between same person is necessarily less than the characteristic distance between different people.The figure of i.e. same people
As the characteristics of image between aspect ratio different people is more like.On this basis, present invention introduces posture to carry out double constraint.According to
C2) for pedestrian's posture of middle acquisition knowable to, sample has been divided into forward direction, laterally and backwards to three classes by the present invention.It is newly-designed
The double constraint formula of five-tuple, which is expressed, is then：
Dpose(Ii a,Ii ps,Ii pd)=d (f (Ii a)-f(Ii ps))-d(f(Ii a)-f(Ii pd)) ＜ β
Wherein,Represent withPosture identical positive sample,Represent withThe different positive sample of posture, β are five yuan
The threshold value of the double constraint of group.The purpose of the loss function acquires a measurement, in the metric space, the identical posture of same person
Under the distance of characteristics of image be less than the distance between same person difference pose presentation feature.Such constraint ensures posture
The distance between identical positive sample is smaller, can mitigate the influence that attitudes vibration is brought.
Original triple is constrained as first and weighs about beam by the present invention, is designed improved posture restraint and is weighed about as second
Beam, both are combined as five-tuple structure, calculate its loss and confirmatory model training can be achieved.Loss function calculation is such as
Under：
F. multiclass loss function calculates as follows with network association training embodiment：Two kinds of losses have been used in combination in the present invention
Function, one kind are softmax loss functions, lay particular emphasis on to do image and classify.Another kind is the five-tuple for adding posture restraint
Loss function, lays particular emphasis on whether two images of checking are same people.Fig. 6 is that the principle of present invention design five-tuple loss function is shown
It is intended to.It is training that disaggregated model, which is generally used for the softmax layers using an output for k, wherein k after network total characteristic layer,
The classification number of collection.The training of sorter network be minimize intersect entropy loss, that is, Classification Loss and above-mentioned steps E) in retouch
The five-tuple loss function stated together, can be with joint training network model.Loss function calculation after joint is：
Loss3(I, w)=λ1Loss1(I,w)+λ2Loss2(I,w)
Stage (2) on-line stage：Carry out specifying pedestrian to identify again in personal data of being expert at storehouse.
A. declarative feature extraction：The feature extraction network extraction figure to be analyzed that the present invention trains to obtain using off-line phase
The declarative feature of piece and existing pedestrian's picture library.Simultaneously before update next time, store one by one special corresponding to photo current storehouse
Sign vector.
B. with pedestrian's picture library and the similarity measurement of feature database：After comparative analysis Euclidean distance, COS distance, more than selection
Chordal distance is as gauge mode.It is analysed to the characteristic vector of picture and the eigen vector of picture library carries out similitude successively
Measurement.Obtained similitude is normalized, sequence processing.Similitude is more than more than 0.7, and M picture is used as inspection before ranking
Hitch fruit.Wherein M is set according to photo current sum dynamic.
C. pedestrian picture library regularly updates mode：For static map valut, then picture to be analyzed every time is constantly added,
And store its characteristic vector.For pedestrian's data caused by dynamic video, then once check and obtained every renewal in 30 minutes
Pedestrian information, and before next update, the pedestrian target for being determined as new person is continuously added, feature is extracted later in system, so as to
Online completion pedestrian identifies again after obtaining target to be analyzed.
D. pedestrian's weight recognition result visualization scheme：The present invention except record inquire about every time result, deposit database with
Outside.Pedestrian's weight recognition result is also shown with both modalities which, for static map valut, then shows target to be analyzed and not more than M
It is determined as the picture of same a group traveling together；For dynamic video, then the picture of retrieval result is locked first, and according to it in database
The camera ID of storage, pedestrian ID, frame number, in video the information visualization such as position in corresponding video pictures.These storages
Items for information can be not only used for visualizing, and can be also used for the upper layer applications such as camera topological analysis, video content analysis.
Claims (8)
- A kind of 1. pedestrian's recognition methods again that more loss functions are designed using attitude information, it is characterised in that：Including offline extraction Character network model training, online pedestrian identify two parts main contents again；Step (1), offline extraction character network model training stage：(m1) all pictures are pre-processed, original image rIiI is used after processingiRepresent；(m2) for each pictures detection artis information, there is P in 18 obtained artis informationIi={ x1, y1... ..., x18, y18In, and there is corresponding Boolean type array label to indicate whether to detect different artis, labeli=(True or False)；(m3) according to the artis information extracted in step (m2), thus it is speculated that the height high of each pedestriani, respectively calculate head- The local region information of trunk-leg(m4) the artis information extracted according to step (m2), thus it is speculated that the posture direction of pedestrian target, be designated as diri=(1 or 2 Or 3), wherein representing positive sample when equal to 1,2 represent lateral samples, and 3 represent backwards to sample；(m5) global characteristics are extracted according to the core network of design, according to the regional area positional information of step (m3) extraction and divided Branch network structure extracts local feature, and by the global characteristics and Local Feature Fusion of every pictures, is collectively forming declarative spy Sign vector；(m6) the first heavy triple of more Classification Loss functions and the present invention are calculated according to data true tag, while according to step Suddenly pedestrian's posture that (m4) speculates is towards design five-tuple and calculates the second heavy five-tuple loss function；(m7) a variety of loss function errors training current signature extraction network that joint step (m6) calculates, and analyze different losses Influence of the function weight to network, select optimal weight λ1And λ2To complete joint training；Step (2), online pedestrian weight cognitive phase：(s1) all picture I in picture library are pre-processedgallery, and the network model for training to obtain using step (1) off-line phase enters Row feature extraction, identification information stores to form feature database F one by one according to corresponding to picturegallery；(s2) picture I to be analyzed is pre-processedquery, train obtained network model using step (1) off-line phase and carry out feature to carry Take, final characteristic vector fqueryUnique effective information as subsequent step (s3) similarity measurement；(s3) f extracted in calculation procedure (s2)queryWith feature database FgalleryBetween characteristic distance, and be normalized, arrange Sequence operates, and therefrom selects similarity more than 0.7 and the retrieval result that is identified again as pedestrian in preceding M picture of ranking, wherein M's Numerical value is selected according to Number dynamics in photo current storehouse；(s4) picture library and its corresponding feature database are regularly updated, emphasis supports static library and detected by dynamic video storehouse to gather The dynamic base both modalities which arrived.
- 2. a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information according to claim 1, it is special Sign is：Described step (m3) comprises the following steps：(m3.1) the artis information P extracted according to step (m2)Ii, remove labeliFalse is all, that is, detects artis failure Sample；Removal represents the sample that torso portion artis is largely False；(m3.2) step (m2) extraction artis information meets the sample participation training of application claims, is believed according to existing artis Cease PIiSpeculate pedestrian's height highi；(m3.3) according to artis such as left-right ear, noses, head zone information is calculated(m3.4) according to right and left shoulders, left and right hip associated joint dot position information, torso area information is calculated(m3.5) according to information such as waist location, ankle, heights, leg area information is calculatedDue to detection Obtained encirclement frame does not often include foot, therefore according to the proportional downward scaling of height；(m3.6) the regional area positional information being calculated according to step (m3.3) to (m3.5) produces interest region, using changing The region of interest feature extraction layer entered, local shape factor is carried out into branching networks.
- 3. a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information according to claim 1, it is special Sign is：Described step (m4) comprises the following steps：(m4.1) after being screened according to step (m3.1), it is determined that participate in sample analysis its posture direction of training, without left shoulder or The sample of right shoulder is determined as lateral diri=2；(m4.2) the existing sample of the right shoulder of left shoulder, then right and left shoulders vector is calculated(m4.3) the right and left shoulders vector obtained according to step (m4.2)Calculate the angle dir_angle with vertical curveIi；(m4.4) the angle dir_angle that judgment step (m4.3) is calculatedIiAffiliated angular range, as angle dir_ angleIiThen bearing mark is positive dir in scope [260 °, 280 °]i=1, if not whether positive then judge angle in scope In [80 °, 100 °], this is labeled as backwards to dir within the rangei=3, if not in the range of above-mentioned two, mark the sample This is lateral diri=2.
- 4. a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information according to claim 1, it is special Sign is：Described step (m5) comprises the following steps：(m5.1) the core network extraction global characteristics in network frame proposed by the present invention, labeled as fglobal (Ii)；(m5.2) three local feature flocal (I of Connection Step (m3.6) extractioni), it is respectively labeled as fh (Ii)、ft (Ii)、fl(Ii)；(m5.3) global characteristics and the local feature f of step (m5.2) extraction of step (m5.1) extraction are realized using full articulamentum (Ii)。
- 5. a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information according to claim 1, it is special Sign is：Described step (m6) comprises the following steps：(m6.1) more Classification Loss function errors are calculated；(m6.2) the first weight triple constraint of the present invention is calculated：Did(Ii a,Ii p,Ii n)=d (f (Ii a)-f(Ii p))-d(f(Ii a)-f(Ii n)) ＜ αWherein, IaIt is any one benchmark pedestrian image, I in data seti pExpression represents another of same people with benchmark pedestrian Image, i.e. positive sample, Ii nFor other people image, i.e. negative sample, triple input obtains respective after network calculations Characteristic vector { f (Ii a),f(Ii p),f(Ii n), d (f (Ii a)-f(Ii p)) on the basis of figure with positive sample to the distance between, d (f (Ii a)-f(Ii n)) on the basis of figure with negative sample to the distance between, α be triple constraint threshold value；(m6.3) the second weight five-tuple constraint of the present invention is calculated：Dpose(Ii a,Ii ps,Ii pd)=d (f (Ii a)-f(Ii ps))-d(f(Ii a)-f(Ii pd)) ＜ βWherein,Represent withPosture identical positive sample,Represent withThe different positive sample of posture, β are that five-tuple is double The threshold value of constraint.
- 6. a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information according to claim 1, it is special Sign is that described step (m7) comprises the following steps：(m7.1) the joint error value of the more loss function error calculation backpropagations obtained according to step (m6)：<mrow> <msub> <mi>Loss</mi> <mn>1</mn> </msub> <mrow> <mo>(</mo> <mi>I</mi> <mo>,</mo> <mi>w</mi> <mo>)</mo> </mrow> <mo>=</mo> <msubsup> <mo>&Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </msubsup> <mo>-</mo> <msub> <mi>p</mi> <mi>i</mi> </msub> <mi>log</mi> <mrow> <mo>(</mo> <msub> <mover> <mi>p</mi> <mo>^</mo> </mover> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>;</mo> </mrow><mfenced open = "" close = ""> <mtable> <mtr> <mtd> <mrow> <msub> <mi>Loss</mi> <mn>2</mn> </msub> <mrow> <mo>(</mo> <mi>I</mi> <mo>,</mo> <mi>w</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mi>n</mi> </mfrac> <msubsup> <mo>&Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </msubsup> <mrow> <mo>(</mo> <mi>max</mi> <mo>{</mo> <msub> <mi>D</mi> <mrow> <mi>i</mi> <mi>d</mi> </mrow> </msub> <mo>(</mo> <mrow> <msup> <msub> <mi>I</mi> <mi>i</mi> </msub> <mi>a</mi> </msup> <mo>,</mo> <msup> <msub> <mi>I</mi> <mi>i</mi> </msub> <mi>b</mi> </msup> <mo>,</mo> <msup> <msub> <mi>I</mi> <mi>i</mi> </msub> <mi>n</mi> </msup> </mrow> <mo>)</mo> <mo>,</mo> <mi>&alpha;</mi> <mo>}</mo> <mo>+</mo> <mi>&lambda;</mi> <mi>max</mi> <mo>{</mo> <msub> <mi>D</mi> <mrow> <mi>p</mi> <mi>o</mi> <mi>s</mi> <mi>e</mi> </mrow> </msub> <mo>(</mo> <mrow> <msup> <msub> <mi>I</mi> <mi>i</mi> </msub> <mi>a</mi> </msup> <mo>,</mo> <msup> <msub> <mi>I</mi> <mi>i</mi> </msub> <mrow> <mi>p</mi> <mi>s</mi> </mrow> </msup> <mo>,</mo> <msup> <msub> <mi>I</mi> <mi>i</mi> </msub> <mrow> <mi>p</mi> <mi>d</mi> </mrow> </msup> </mrow> <mo>)</mo> <mo>,</mo> <mi>&beta;</mi> <mo>}</mo> <mo>)</mo> </mrow> <mo>;</mo> </mrow> </mtd> </mtr> </mtable> </mfenced>Loss3(I, w)=λ1Loss1(I,w)+λ2Loss2(I,w)；Wherein, Loss1Represent more Classification Loss functions, Loss2Represent five-tuple loss function, Loss3Represent associated losses letter Number, λ1And λ2To balance the weighted value of associated losses function, the weighted value that λ is balance triple and five-tuple constrains, w networks ginseng Number,Represent the probability of prediction, piIt is destination probability, N is pedestrian's species number, and n is five-tuple quantity；(m7.2) Error weight parameter lambda in analytical procedure (m7.1)1And λ2Selection, determines the optimal loss letter that off-line phase uses Number distribution weight.
- 7. a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information according to claim 1, it is special Sign is：Described step (s3) comprises the following steps：(s3.1) M numerical value is selected according to Number dynamics in photo current storehouse；(s3.2) f extracted successively in calculation procedure (s2)queryWith feature database FgalleryBetween characteristic distance；(s3.3) all characteristic distances that step (s3.2) is calculated are normalized, sorting operation, therefrom selected similar Spend the retrieval result that more than 0.7 and ranking identifies again in preceding M picture as pedestrian；(s3.4) pedestrian that visualization step (s3.3) obtains identifies retrieval result again, and I is then shown for static map valutqueryAnd I after sequenceresults, for dynamic video storehouse then according to IresultsIn the camera ID, pedestrian ID, encirclement frame position of database purchase Confidence breath and frame number time etc. recover the truth of result in video at that time.
- 8. a kind of pedestrian's recognition methods again that more loss functions are designed using attitude information according to claim 1, it is special Sign is：Described step (s4) comprises the following steps：(s4.1) the time t regularly updated is set；(s4.2) in the range of time t, inquiry picture I is then constantly added for static map valutqueryInformation and feature；Arrive Changed as requested up to t or update picture library, extract the picture feature of change again, set up new feature database；(s4.3) in the range of time t, the fresh target detected is then continuously added for dynamic video storehouse, and deposited in database Store up the world informations such as camera ID, pedestrian ID, encirclement frame positional information, frame number, time, place；Reach t after then according to when Between, half pedestrian's data message in current storehouse is emptied, then adds new testing result frame by frame, while extracts its feature work To be stored in a primary attribute of database.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710946443.3A CN107832672B (en) | 2017-10-12 | 2017-10-12 | Pedestrian re-identification method for designing multi-loss function by utilizing attitude information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710946443.3A CN107832672B (en) | 2017-10-12 | 2017-10-12 | Pedestrian re-identification method for designing multi-loss function by utilizing attitude information |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107832672A true CN107832672A (en) | 2018-03-23 |
| CN107832672B CN107832672B (en) | 2020-07-07 |
Family
ID=61647742
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710946443.3A Active CN107832672B (en) | 2017-10-12 | 2017-10-12 | Pedestrian re-identification method for designing multi-loss function by utilizing attitude information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107832672B (en) |
Cited By (83)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108537181A (en) * | 2018-04-13 | 2018-09-14 | 盐城师范学院 | A kind of gait recognition method based on the study of big spacing depth measure |
| CN108596211A (en) * | 2018-03-29 | 2018-09-28 | 中山大学 | It is a kind of that pedestrian's recognition methods again is blocked based on focusing study and depth e-learning |
| CN108764065A (en) * | 2018-05-04 | 2018-11-06 | 华中科技大学 | A kind of method of pedestrian's weight identification feature fusion assisted learning |
| CN108960140A (en) * | 2018-07-04 | 2018-12-07 | 国家新闻出版广电总局广播科学研究院 | The pedestrian's recognition methods again extracted and merged based on multi-region feature |
| CN109063607A (en) * | 2018-07-17 | 2018-12-21 | 北京迈格威科技有限公司 | The method and device that loss function for identifying again determines |
| CN109165589A (en) * | 2018-08-14 | 2019-01-08 | 北京颂泽科技有限公司 | Vehicle based on deep learning recognition methods and device again |
| CN109190646A (en) * | 2018-06-25 | 2019-01-11 | 北京达佳互联信息技术有限公司 | A kind of data predication method neural network based, device and nerve network system |
| CN109190446A (en) * | 2018-07-06 | 2019-01-11 | 西北工业大学 | Pedestrian's recognition methods again based on triple focused lost function |
| CN109214271A (en) * | 2018-07-17 | 2019-01-15 | 北京迈格威科技有限公司 | The method and device that loss function for identifying again determines |
| CN109299707A (en) * | 2018-10-30 | 2019-02-01 | 天津师范大学 | A kind of unsupervised pedestrian recognition methods again based on fuzzy depth cluster |
| CN109446898A (en) * | 2018-09-20 | 2019-03-08 | 暨南大学 | A kind of recognition methods again of the pedestrian based on transfer learning and Fusion Features |
| CN109472248A (en) * | 2018-11-22 | 2019-03-15 | 广东工业大学 | A kind of pedestrian recognition methods, system and electronic equipment and storage medium again |
| CN109492583A (en) * | 2018-11-09 | 2019-03-19 | 安徽大学 | A kind of recognition methods again of the vehicle based on deep learning |
| CN109508663A (en) * | 2018-10-31 | 2019-03-22 | 上海交通大学 | A kind of pedestrian's recognition methods again based on multi-level supervision network |
| CN109522850A (en) * | 2018-11-22 | 2019-03-26 | 中山大学 | A kind of movement similarity estimating method based on small-sample learning |
| CN109583502A (en) * | 2018-11-30 | 2019-04-05 | 天津师范大学 | A kind of pedestrian's recognition methods again based on confrontation erasing attention mechanism |
| CN109583315A (en) * | 2018-11-02 | 2019-04-05 | 北京工商大学 | A kind of quick human posture recognition method of multichannel towards intelligent video monitoring |
| CN109614853A (en) * | 2018-10-30 | 2019-04-12 | 国家新闻出版广电总局广播科学研究院 | It is a kind of based on body structure divide bilinearity pedestrian identify network establishing method again |
| CN109711386A (en) * | 2019-01-10 | 2019-05-03 | 北京达佳互联信息技术有限公司 | Obtain method, apparatus, electronic equipment and the storage medium of identification model |
| CN109711366A (en) * | 2018-12-29 | 2019-05-03 | 浙江大学 | A kind of recognition methods again of the pedestrian based on group information loss function |
| CN109800794A (en) * | 2018-12-27 | 2019-05-24 | 上海交通大学 | A kind of appearance similar purpose identifies fusion method and system across camera again |
| CN109886141A (en) * | 2019-01-28 | 2019-06-14 | 同济大学 | A kind of pedestrian based on uncertainty optimization discrimination method again |
| CN109902573A (en) * | 2019-01-24 | 2019-06-18 | 中国矿业大学 | Multiple-camera towards video monitoring under mine is without mark pedestrian's recognition methods again |
| CN109919320A (en) * | 2019-01-23 | 2019-06-21 | 西北工业大学 | Triplet online learning methods based on Semantic hierarchy |
| CN109919141A (en) * | 2019-04-09 | 2019-06-21 | 广东省智能制造研究所 | A kind of recognition methods again of the pedestrian based on skeleton pose |
| CN109934197A (en) * | 2019-03-21 | 2019-06-25 | 深圳力维智联技术有限公司 | Training method, device and the computer readable storage medium of human face recognition model |
| CN109993116A (en) * | 2019-03-29 | 2019-07-09 | 上海工程技术大学 | A kind of pedestrian mutually learnt based on skeleton recognition methods again |
| CN110046553A (en) * | 2019-03-21 | 2019-07-23 | 华中科技大学 | A kind of pedestrian weight identification model, method and system merging attributive character |
| CN110110755A (en) * | 2019-04-04 | 2019-08-09 | 长沙千视通智能科技有限公司 | Based on the pedestrian of PTGAN Regional disparity and multiple branches weight recognition detection algorithm and device |
| CN110163110A (en) * | 2019-04-23 | 2019-08-23 | 中电科大数据研究院有限公司 | A kind of pedestrian's recognition methods again merged based on transfer learning and depth characteristic |
| CN110232330A (en) * | 2019-05-23 | 2019-09-13 | 复钧智能科技(苏州)有限公司 | A kind of recognition methods again of the pedestrian based on video detection |
| CN110288677A (en) * | 2019-05-21 | 2019-09-27 | 北京大学 | It is a kind of based on can deformation structure pedestrian image generation method and device |
| CN110309701A (en) * | 2019-04-17 | 2019-10-08 | 武汉大学 | It is a kind of based on the pedestrian across visual angle the same area again recognition methods |
| CN110321862A (en) * | 2019-07-09 | 2019-10-11 | 天津师范大学 | A kind of pedestrian's recognition methods again based on the loss of compact ternary |
| CN110321813A (en) * | 2019-06-18 | 2019-10-11 | 南京信息工程大学 | Cross-domain pedestrian recognition methods again based on pedestrian's segmentation |
| CN110334738A (en) * | 2019-06-05 | 2019-10-15 | 大连理工大学 | The method of more sorter networks for image recognition |
| CN110334675A (en) * | 2019-07-11 | 2019-10-15 | 山东大学 | A kind of pedestrian's recognition methods again based on skeleton key point segmentation and column convolution |
| CN110458004A (en) * | 2019-07-02 | 2019-11-15 | 浙江吉利控股集团有限公司 | A kind of recongnition of objects method, apparatus, equipment and storage medium |
| CN110490901A (en) * | 2019-07-15 | 2019-11-22 | 武汉大学 | The pedestrian detection tracking of anti-attitudes vibration |
| CN110543817A (en) * | 2019-07-25 | 2019-12-06 | 北京大学 | Pedestrian re-identification method based on posture guidance feature learning |
| CN110619271A (en) * | 2019-08-12 | 2019-12-27 | 浙江浩腾电子科技股份有限公司 | Pedestrian re-identification method based on depth region feature connection |
| CN110688888A (en) * | 2019-08-02 | 2020-01-14 | 浙江省北大信息技术高等研究院 | Pedestrian attribute identification method and system based on deep learning |
| CN110858295A (en) * | 2018-08-24 | 2020-03-03 | 广州汽车集团股份有限公司 | Traffic police gesture recognition method and device, vehicle control unit and storage medium |
| CN110968734A (en) * | 2019-11-21 | 2020-04-07 | 华东师范大学 | Pedestrian re-identification method and device based on depth measurement learning |
| CN111091020A (en) * | 2018-10-22 | 2020-05-01 | 百度在线网络技术（北京）有限公司 | Automatic driving state distinguishing method and device |
| CN111126198A (en) * | 2019-12-11 | 2020-05-08 | 中山大学 | Pedestrian re-identification method based on deep representation learning and dynamic matching |
| CN111274958A (en) * | 2020-01-20 | 2020-06-12 | 福州大学 | Pedestrian re-identification method and system with network parameter self-correction function |
| CN111310518A (en) * | 2018-12-11 | 2020-06-19 | 北京嘀嘀无限科技发展有限公司 | Image feature extraction method, target re-identification method and device and electronic equipment |
| CN111401113A (en) * | 2019-01-02 | 2020-07-10 | 南京大学 | Pedestrian re-identification method based on human body posture estimation |
| CN111582154A (en) * | 2020-05-07 | 2020-08-25 | 浙江工商大学 | Pedestrian re-identification method based on multitask skeleton posture division component |
| CN111597876A (en) * | 2020-04-01 | 2020-08-28 | 浙江工业大学 | Cross-modal pedestrian re-identification method based on difficult quintuple |
| CN111598037A (en) * | 2020-05-22 | 2020-08-28 | 北京字节跳动网络技术有限公司 | Method and device for obtaining human body posture predicted value, server and storage medium |
| CN111657926A (en) * | 2020-07-08 | 2020-09-15 | 中国科学技术大学 | Arrhythmia classification method based on multi-lead information fusion |
| CN111738039A (en) * | 2019-05-10 | 2020-10-02 | 北京京东尚科信息技术有限公司 | Pedestrian re-identification method, terminal and storage medium |
| CN111783506A (en) * | 2019-05-17 | 2020-10-16 | 北京京东尚科信息技术有限公司 | Method and device for determining target characteristics and computer-readable storage medium |
| CN111797813A (en) * | 2020-07-21 | 2020-10-20 | 天津理工大学 | Partial pedestrian re-identification method based on visible perception texture semantic alignment |
| CN111832348A (en) * | 2019-04-17 | 2020-10-27 | 中国科学院宁波材料技术与工程研究所 | Pedestrian re-identification method based on pixel and channel attention mechanism |
| CN111861335A (en) * | 2020-07-23 | 2020-10-30 | 广元量知汇科技有限公司 | Industrial interconnection material management system |
| CN112084917A (en) * | 2020-08-31 | 2020-12-15 | 腾讯科技（深圳）有限公司 | Living body detection method and device |
| CN112101300A (en) * | 2020-11-02 | 2020-12-18 | 北京妙医佳健康科技集团有限公司 | Medicinal material identification method and device and electronic equipment |
| CN112307979A (en) * | 2020-10-31 | 2021-02-02 | 成都新潮传媒集团有限公司 | Personnel attribute identification method and device and computer equipment |
| CN112381859A (en) * | 2020-11-20 | 2021-02-19 | 公安部第三研究所 | System, method, device, processor and storage medium for realizing intelligent analysis, identification and processing for video image data |
| CN112382068A (en) * | 2020-11-02 | 2021-02-19 | 陈松山 | Station waiting line crossing detection system based on BIM and DNN |
| CN112417932A (en) * | 2019-08-23 | 2021-02-26 | 中移雄安信息通信科技有限公司 | Method, device and equipment for identifying target object in video |
| CN112488071A (en) * | 2020-12-21 | 2021-03-12 | 重庆紫光华山智安科技有限公司 | Method, device, electronic equipment and storage medium for extracting pedestrian features |
| CN112597944A (en) * | 2020-12-29 | 2021-04-02 | 北京市商汤科技开发有限公司 | Key point detection method and device, electronic equipment and storage medium |
| CN112733707A (en) * | 2021-01-07 | 2021-04-30 | 浙江大学 | Pedestrian re-identification method based on deep learning |
| CN112733921A (en) * | 2020-12-31 | 2021-04-30 | 深圳辰视智能科技有限公司 | Neural network loss function calculation method and system for predicting rigid body 6D posture |
| CN112733594A (en) * | 2020-12-01 | 2021-04-30 | 贵州电网有限责任公司 | Machine room figure re-identification method based on deformable convolutional network |
| WO2021082078A1 (en) * | 2019-10-30 | 2021-05-06 | 平安科技（深圳）有限公司 | Pedestrian re-recognition method and apparatus, computer device, and readable storage medium |
| CN112784772A (en) * | 2021-01-27 | 2021-05-11 | 浙江大学 | In-camera supervised cross-camera pedestrian re-identification method based on contrast learning |
| CN112990120A (en) * | 2021-04-25 | 2021-06-18 | 昆明理工大学 | Cross-domain pedestrian re-identification method using camera style separation domain information |
| CN112989911A (en) * | 2020-12-10 | 2021-06-18 | 奥比中光科技集团股份有限公司 | Pedestrian re-identification method and system |
| CN113255598A (en) * | 2021-06-29 | 2021-08-13 | 南京视察者智能科技有限公司 | Pedestrian re-identification method based on Transformer |
| CN113408351A (en) * | 2021-05-18 | 2021-09-17 | 河南大学 | Pedestrian re-recognition method for generating confrontation network based on attitude guidance |
| CN113963206A (en) * | 2021-10-20 | 2022-01-21 | 中国石油大学(华东) | Posture guidance-based target detection method for fast skating athletes |
| US11238274B2 (en) * | 2017-07-04 | 2022-02-01 | Peking University | Image feature extraction method for person re-identification |
| CN114067356A (en) * | 2021-10-21 | 2022-02-18 | 电子科技大学 | Pedestrian re-identification method based on joint local guidance and attribute clustering |
| CN114120665A (en) * | 2022-01-29 | 2022-03-01 | 山东科技大学 | Intelligent phase control method and system based on pedestrian number |
| CN114550220A (en) * | 2022-04-21 | 2022-05-27 | 中国科学技术大学 | Training method of pedestrian re-recognition model and pedestrian re-recognition method |
| CN115631464A (en) * | 2022-11-17 | 2023-01-20 | 北京航空航天大学 | Pedestrian three-dimensional representation method oriented to large space-time target association |
| CN115762172A (en) * | 2022-11-02 | 2023-03-07 | 济南博观智能科技有限公司 | Method, device, equipment and medium for identifying vehicles entering and exiting parking places |
| CN116192997A (en) * | 2023-02-21 | 2023-05-30 | 上海兴容信息技术有限公司 | Event detection method and system based on network flow |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070211938A1 (en) * | 2006-03-09 | 2007-09-13 | General Electric Company | Method and system for performing image re-identification |
| CN105138998A (en) * | 2015-09-07 | 2015-12-09 | 上海交通大学 | Method and system for re-identifying pedestrian based on view angle self-adaptive subspace learning algorithm |
| CN105518744A (en) * | 2015-06-29 | 2016-04-20 | 北京旷视科技有限公司 | Pedestrian re-identification method and equipment |
| CN106778527A (en) * | 2016-11-28 | 2017-05-31 | 中通服公众信息产业股份有限公司 | A kind of improved neutral net pedestrian recognition methods again based on triple losses |
| CN107145852A (en) * | 2017-04-28 | 2017-09-08 | 深圳市唯特视科技有限公司 | A kind of character recognition method based on homologous cosine losses function |
-
2017
- 2017-10-12 CN CN201710946443.3A patent/CN107832672B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070211938A1 (en) * | 2006-03-09 | 2007-09-13 | General Electric Company | Method and system for performing image re-identification |
| CN105518744A (en) * | 2015-06-29 | 2016-04-20 | 北京旷视科技有限公司 | Pedestrian re-identification method and equipment |
| CN105138998A (en) * | 2015-09-07 | 2015-12-09 | 上海交通大学 | Method and system for re-identifying pedestrian based on view angle self-adaptive subspace learning algorithm |
| CN106778527A (en) * | 2016-11-28 | 2017-05-31 | 中通服公众信息产业股份有限公司 | A kind of improved neutral net pedestrian recognition methods again based on triple losses |
| CN107145852A (en) * | 2017-04-28 | 2017-09-08 | 深圳市唯特视科技有限公司 | A kind of character recognition method based on homologous cosine losses function |
Non-Patent Citations (3)
| Title |
|---|
| DE CHENG ETAL.: "Person Re-Identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| WEI LI ETAL.: "DeepReID: Deep Filter Pairing Neural Network for Person Re-Identification", 《2014 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| 王枫等: "基于双卷积神经网络的行人精细化识别", 《中国科技论文》 * |
Cited By (127)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11238274B2 (en) * | 2017-07-04 | 2022-02-01 | Peking University | Image feature extraction method for person re-identification |
| CN108596211A (en) * | 2018-03-29 | 2018-09-28 | 中山大学 | It is a kind of that pedestrian's recognition methods again is blocked based on focusing study and depth e-learning |
| CN108596211B (en) * | 2018-03-29 | 2020-08-28 | 中山大学 | Shielded pedestrian re-identification method based on centralized learning and deep network learning |
| CN108537181A (en) * | 2018-04-13 | 2018-09-14 | 盐城师范学院 | A kind of gait recognition method based on the study of big spacing depth measure |
| CN108764065A (en) * | 2018-05-04 | 2018-11-06 | 华中科技大学 | A kind of method of pedestrian's weight identification feature fusion assisted learning |
| CN109190646A (en) * | 2018-06-25 | 2019-01-11 | 北京达佳互联信息技术有限公司 | A kind of data predication method neural network based, device and nerve network system |
| CN109190646B (en) * | 2018-06-25 | 2019-08-20 | 北京达佳互联信息技术有限公司 | A kind of data predication method neural network based, device and nerve network system |
| CN108960140B (en) * | 2018-07-04 | 2021-04-27 | 国家新闻出版广电总局广播科学研究院 | Pedestrian re-identification method based on multi-region feature extraction and fusion |
| CN108960140A (en) * | 2018-07-04 | 2018-12-07 | 国家新闻出版广电总局广播科学研究院 | The pedestrian's recognition methods again extracted and merged based on multi-region feature |
| CN109190446A (en) * | 2018-07-06 | 2019-01-11 | 西北工业大学 | Pedestrian's recognition methods again based on triple focused lost function |
| CN109214271A (en) * | 2018-07-17 | 2019-01-15 | 北京迈格威科技有限公司 | The method and device that loss function for identifying again determines |
| CN109063607A (en) * | 2018-07-17 | 2018-12-21 | 北京迈格威科技有限公司 | The method and device that loss function for identifying again determines |
| CN109063607B (en) * | 2018-07-17 | 2022-11-25 | 北京迈格威科技有限公司 | Method and device for determining loss function for re-identification |
| CN109214271B (en) * | 2018-07-17 | 2022-10-18 | 北京迈格威科技有限公司 | Method and device for determining loss function for re-identification |
| CN109165589B (en) * | 2018-08-14 | 2021-02-23 | 北京颂泽科技有限公司 | Vehicle weight recognition method and device based on deep learning |
| CN109165589A (en) * | 2018-08-14 | 2019-01-08 | 北京颂泽科技有限公司 | Vehicle based on deep learning recognition methods and device again |
| CN110858295B (en) * | 2018-08-24 | 2021-04-20 | 广州汽车集团股份有限公司 | Traffic police gesture recognition method and device, vehicle control unit and storage medium |
| CN110858295A (en) * | 2018-08-24 | 2020-03-03 | 广州汽车集团股份有限公司 | Traffic police gesture recognition method and device, vehicle control unit and storage medium |
| CN109446898A (en) * | 2018-09-20 | 2019-03-08 | 暨南大学 | A kind of recognition methods again of the pedestrian based on transfer learning and Fusion Features |
| CN109446898B (en) * | 2018-09-20 | 2021-10-15 | 暨南大学 | Pedestrian re-identification method based on transfer learning and feature fusion |
| CN111091020A (en) * | 2018-10-22 | 2020-05-01 | 百度在线网络技术（北京）有限公司 | Automatic driving state distinguishing method and device |
| CN109614853A (en) * | 2018-10-30 | 2019-04-12 | 国家新闻出版广电总局广播科学研究院 | It is a kind of based on body structure divide bilinearity pedestrian identify network establishing method again |
| CN109299707A (en) * | 2018-10-30 | 2019-02-01 | 天津师范大学 | A kind of unsupervised pedestrian recognition methods again based on fuzzy depth cluster |
| CN109614853B (en) * | 2018-10-30 | 2023-05-05 | 国家新闻出版广电总局广播科学研究院 | Bilinear pedestrian re-identification network construction method based on body structure division |
| CN109508663B (en) * | 2018-10-31 | 2021-07-13 | 上海交通大学 | Pedestrian re-identification method based on multi-level supervision network |
| CN109508663A (en) * | 2018-10-31 | 2019-03-22 | 上海交通大学 | A kind of pedestrian's recognition methods again based on multi-level supervision network |
| CN109583315A (en) * | 2018-11-02 | 2019-04-05 | 北京工商大学 | A kind of quick human posture recognition method of multichannel towards intelligent video monitoring |
| CN109583315B (en) * | 2018-11-02 | 2023-05-12 | 北京工商大学 | Multichannel rapid human body posture recognition method for intelligent video monitoring |
| CN109492583A (en) * | 2018-11-09 | 2019-03-19 | 安徽大学 | A kind of recognition methods again of the vehicle based on deep learning |
| CN109522850A (en) * | 2018-11-22 | 2019-03-26 | 中山大学 | A kind of movement similarity estimating method based on small-sample learning |
| CN109522850B (en) * | 2018-11-22 | 2023-03-10 | 中山大学 | Action similarity evaluation method based on small sample learning |
| CN109472248A (en) * | 2018-11-22 | 2019-03-15 | 广东工业大学 | A kind of pedestrian recognition methods, system and electronic equipment and storage medium again |
| CN109472248B (en) * | 2018-11-22 | 2022-03-25 | 广东工业大学 | Pedestrian re-identification method and system, electronic equipment and storage medium |
| CN109583502A (en) * | 2018-11-30 | 2019-04-05 | 天津师范大学 | A kind of pedestrian's recognition methods again based on confrontation erasing attention mechanism |
| CN109583502B (en) * | 2018-11-30 | 2022-11-18 | 天津师范大学 | Pedestrian re-identification method based on anti-erasure attention mechanism |
| CN111310518B (en) * | 2018-12-11 | 2023-12-08 | 北京嘀嘀无限科技发展有限公司 | Picture feature extraction method, target re-identification method, device and electronic equipment |
| CN111310518A (en) * | 2018-12-11 | 2020-06-19 | 北京嘀嘀无限科技发展有限公司 | Image feature extraction method, target re-identification method and device and electronic equipment |
| CN109800794B (en) * | 2018-12-27 | 2021-10-22 | 上海交通大学 | Cross-camera re-identification fusion method and system for appearance similar targets |
| CN109800794A (en) * | 2018-12-27 | 2019-05-24 | 上海交通大学 | A kind of appearance similar purpose identifies fusion method and system across camera again |
| CN109711366A (en) * | 2018-12-29 | 2019-05-03 | 浙江大学 | A kind of recognition methods again of the pedestrian based on group information loss function |
| CN111401113A (en) * | 2019-01-02 | 2020-07-10 | 南京大学 | Pedestrian re-identification method based on human body posture estimation |
| CN109711386A (en) * | 2019-01-10 | 2019-05-03 | 北京达佳互联信息技术有限公司 | Obtain method, apparatus, electronic equipment and the storage medium of identification model |
| CN109919320A (en) * | 2019-01-23 | 2019-06-21 | 西北工业大学 | Triplet online learning methods based on Semantic hierarchy |
| CN109919320B (en) * | 2019-01-23 | 2022-04-01 | 西北工业大学 | Triplet network learning method based on semantic hierarchy |
| CN109902573B (en) * | 2019-01-24 | 2023-10-31 | 中国矿业大学 | Multi-camera non-labeling pedestrian re-identification method for video monitoring under mine |
| CN109902573A (en) * | 2019-01-24 | 2019-06-18 | 中国矿业大学 | Multiple-camera towards video monitoring under mine is without mark pedestrian's recognition methods again |
| CN109886141B (en) * | 2019-01-28 | 2023-06-06 | 同济大学 | Pedestrian re-identification method based on uncertainty optimization |
| CN109886141A (en) * | 2019-01-28 | 2019-06-14 | 同济大学 | A kind of pedestrian based on uncertainty optimization discrimination method again |
| CN110046553A (en) * | 2019-03-21 | 2019-07-23 | 华中科技大学 | A kind of pedestrian weight identification model, method and system merging attributive character |
| CN109934197A (en) * | 2019-03-21 | 2019-06-25 | 深圳力维智联技术有限公司 | Training method, device and the computer readable storage medium of human face recognition model |
| CN109993116A (en) * | 2019-03-29 | 2019-07-09 | 上海工程技术大学 | A kind of pedestrian mutually learnt based on skeleton recognition methods again |
| CN110110755A (en) * | 2019-04-04 | 2019-08-09 | 长沙千视通智能科技有限公司 | Based on the pedestrian of PTGAN Regional disparity and multiple branches weight recognition detection algorithm and device |
| CN110110755B (en) * | 2019-04-04 | 2021-02-26 | 长沙千视通智能科技有限公司 | Pedestrian re-identification detection method and device based on PTGAN region difference and multiple branches |
| CN109919141A (en) * | 2019-04-09 | 2019-06-21 | 广东省智能制造研究所 | A kind of recognition methods again of the pedestrian based on skeleton pose |
| CN111832348A (en) * | 2019-04-17 | 2020-10-27 | 中国科学院宁波材料技术与工程研究所 | Pedestrian re-identification method based on pixel and channel attention mechanism |
| CN111832348B (en) * | 2019-04-17 | 2022-05-06 | 中国科学院宁波材料技术与工程研究所 | Pedestrian re-identification method based on pixel and channel attention mechanism |
| CN110309701A (en) * | 2019-04-17 | 2019-10-08 | 武汉大学 | It is a kind of based on the pedestrian across visual angle the same area again recognition methods |
| CN110309701B (en) * | 2019-04-17 | 2022-08-05 | 武汉大学 | Pedestrian re-identification method based on same cross-view-angle area |
| CN110163110B (en) * | 2019-04-23 | 2023-06-06 | 中电科大数据研究院有限公司 | Pedestrian re-recognition method based on transfer learning and depth feature fusion |
| CN110163110A (en) * | 2019-04-23 | 2019-08-23 | 中电科大数据研究院有限公司 | A kind of pedestrian's recognition methods again merged based on transfer learning and depth characteristic |
| CN111738039A (en) * | 2019-05-10 | 2020-10-02 | 北京京东尚科信息技术有限公司 | Pedestrian re-identification method, terminal and storage medium |
| CN111738039B (en) * | 2019-05-10 | 2024-05-24 | 北京京东尚科信息技术有限公司 | Pedestrian re-identification method, terminal and storage medium |
| CN111783506A (en) * | 2019-05-17 | 2020-10-16 | 北京京东尚科信息技术有限公司 | Method and device for determining target characteristics and computer-readable storage medium |
| CN110288677A (en) * | 2019-05-21 | 2019-09-27 | 北京大学 | It is a kind of based on can deformation structure pedestrian image generation method and device |
| CN110232330A (en) * | 2019-05-23 | 2019-09-13 | 复钧智能科技(苏州)有限公司 | A kind of recognition methods again of the pedestrian based on video detection |
| CN110232330B (en) * | 2019-05-23 | 2020-11-06 | 复钧智能科技(苏州)有限公司 | Pedestrian re-identification method based on video detection |
| CN110334738A (en) * | 2019-06-05 | 2019-10-15 | 大连理工大学 | The method of more sorter networks for image recognition |
| CN110321813A (en) * | 2019-06-18 | 2019-10-11 | 南京信息工程大学 | Cross-domain pedestrian recognition methods again based on pedestrian's segmentation |
| CN110321813B (en) * | 2019-06-18 | 2023-06-20 | 南京信息工程大学 | Cross-domain pedestrian re-identification method based on pedestrian segmentation |
| CN110458004A (en) * | 2019-07-02 | 2019-11-15 | 浙江吉利控股集团有限公司 | A kind of recongnition of objects method, apparatus, equipment and storage medium |
| CN110321862A (en) * | 2019-07-09 | 2019-10-11 | 天津师范大学 | A kind of pedestrian's recognition methods again based on the loss of compact ternary |
| CN110334675A (en) * | 2019-07-11 | 2019-10-15 | 山东大学 | A kind of pedestrian's recognition methods again based on skeleton key point segmentation and column convolution |
| CN110334675B (en) * | 2019-07-11 | 2022-12-27 | 山东大学 | Pedestrian re-identification method based on human skeleton key point segmentation and column convolution |
| CN110490901A (en) * | 2019-07-15 | 2019-11-22 | 武汉大学 | The pedestrian detection tracking of anti-attitudes vibration |
| CN110543817A (en) * | 2019-07-25 | 2019-12-06 | 北京大学 | Pedestrian re-identification method based on posture guidance feature learning |
| CN110688888A (en) * | 2019-08-02 | 2020-01-14 | 浙江省北大信息技术高等研究院 | Pedestrian attribute identification method and system based on deep learning |
| CN110688888B (en) * | 2019-08-02 | 2022-08-05 | 杭州未名信科科技有限公司 | Pedestrian attribute identification method and system based on deep learning |
| CN110619271A (en) * | 2019-08-12 | 2019-12-27 | 浙江浩腾电子科技股份有限公司 | Pedestrian re-identification method based on depth region feature connection |
| CN112417932B (en) * | 2019-08-23 | 2023-04-07 | 中移雄安信息通信科技有限公司 | Method, device and equipment for identifying target object in video |
| CN112417932A (en) * | 2019-08-23 | 2021-02-26 | 中移雄安信息通信科技有限公司 | Method, device and equipment for identifying target object in video |
| WO2021082078A1 (en) * | 2019-10-30 | 2021-05-06 | 平安科技（深圳）有限公司 | Pedestrian re-recognition method and apparatus, computer device, and readable storage medium |
| CN110968734A (en) * | 2019-11-21 | 2020-04-07 | 华东师范大学 | Pedestrian re-identification method and device based on depth measurement learning |
| CN110968734B (en) * | 2019-11-21 | 2023-08-04 | 华东师范大学 | Pedestrian re-recognition method and device based on deep measurement learning |
| CN111126198A (en) * | 2019-12-11 | 2020-05-08 | 中山大学 | Pedestrian re-identification method based on deep representation learning and dynamic matching |
| CN111126198B (en) * | 2019-12-11 | 2023-05-09 | 中山大学 | Pedestrian re-identification method based on deep representation learning and dynamic matching |
| CN111274958B (en) * | 2020-01-20 | 2022-10-04 | 福州大学 | Pedestrian re-identification method and system with network parameter self-correction function |
| CN111274958A (en) * | 2020-01-20 | 2020-06-12 | 福州大学 | Pedestrian re-identification method and system with network parameter self-correction function |
| CN111597876A (en) * | 2020-04-01 | 2020-08-28 | 浙江工业大学 | Cross-modal pedestrian re-identification method based on difficult quintuple |
| CN111582154A (en) * | 2020-05-07 | 2020-08-25 | 浙江工商大学 | Pedestrian re-identification method based on multitask skeleton posture division component |
| CN111598037A (en) * | 2020-05-22 | 2020-08-28 | 北京字节跳动网络技术有限公司 | Method and device for obtaining human body posture predicted value, server and storage medium |
| CN111598037B (en) * | 2020-05-22 | 2023-04-25 | 北京字节跳动网络技术有限公司 | Human body posture predicted value acquisition method, device, server and storage medium |
| CN111657926A (en) * | 2020-07-08 | 2020-09-15 | 中国科学技术大学 | Arrhythmia classification method based on multi-lead information fusion |
| CN111657926B (en) * | 2020-07-08 | 2021-04-23 | 中国科学技术大学 | Arrhythmia classification method based on multi-lead information fusion |
| CN111797813B (en) * | 2020-07-21 | 2022-08-02 | 天津理工大学 | Partial pedestrian re-identification method based on visible perception texture semantic alignment |
| CN111797813A (en) * | 2020-07-21 | 2020-10-20 | 天津理工大学 | Partial pedestrian re-identification method based on visible perception texture semantic alignment |
| CN111861335A (en) * | 2020-07-23 | 2020-10-30 | 广元量知汇科技有限公司 | Industrial interconnection material management system |
| CN112084917B (en) * | 2020-08-31 | 2024-06-04 | 腾讯科技（深圳）有限公司 | Living body detection method and device |
| CN112084917A (en) * | 2020-08-31 | 2020-12-15 | 腾讯科技（深圳）有限公司 | Living body detection method and device |
| CN112307979A (en) * | 2020-10-31 | 2021-02-02 | 成都新潮传媒集团有限公司 | Personnel attribute identification method and device and computer equipment |
| CN112382068A (en) * | 2020-11-02 | 2021-02-19 | 陈松山 | Station waiting line crossing detection system based on BIM and DNN |
| CN112101300A (en) * | 2020-11-02 | 2020-12-18 | 北京妙医佳健康科技集团有限公司 | Medicinal material identification method and device and electronic equipment |
| CN112381859A (en) * | 2020-11-20 | 2021-02-19 | 公安部第三研究所 | System, method, device, processor and storage medium for realizing intelligent analysis, identification and processing for video image data |
| CN112381859B (en) * | 2020-11-20 | 2024-06-04 | 公安部第三研究所 | System, method, device, processor and storage medium for implementing intelligent analysis and identification processing for video image data |
| CN112733594A (en) * | 2020-12-01 | 2021-04-30 | 贵州电网有限责任公司 | Machine room figure re-identification method based on deformable convolutional network |
| CN112989911A (en) * | 2020-12-10 | 2021-06-18 | 奥比中光科技集团股份有限公司 | Pedestrian re-identification method and system |
| CN112488071B (en) * | 2020-12-21 | 2021-10-26 | 重庆紫光华山智安科技有限公司 | Method, device, electronic equipment and storage medium for extracting pedestrian features |
| CN112488071A (en) * | 2020-12-21 | 2021-03-12 | 重庆紫光华山智安科技有限公司 | Method, device, electronic equipment and storage medium for extracting pedestrian features |
| CN112597944A (en) * | 2020-12-29 | 2021-04-02 | 北京市商汤科技开发有限公司 | Key point detection method and device, electronic equipment and storage medium |
| CN112597944B (en) * | 2020-12-29 | 2024-06-11 | 北京市商汤科技开发有限公司 | Key point detection method and device, electronic equipment and storage medium |
| CN112733921A (en) * | 2020-12-31 | 2021-04-30 | 深圳辰视智能科技有限公司 | Neural network loss function calculation method and system for predicting rigid body 6D posture |
| CN112733921B (en) * | 2020-12-31 | 2024-05-17 | 深圳辰视智能科技有限公司 | Neural network loss function calculation method and system for predicting rigid body 6D posture |
| CN112733707B (en) * | 2021-01-07 | 2023-11-14 | 浙江大学 | Pedestrian re-recognition method based on deep learning |
| CN112733707A (en) * | 2021-01-07 | 2021-04-30 | 浙江大学 | Pedestrian re-identification method based on deep learning |
| CN112784772A (en) * | 2021-01-27 | 2021-05-11 | 浙江大学 | In-camera supervised cross-camera pedestrian re-identification method based on contrast learning |
| CN112784772B (en) * | 2021-01-27 | 2022-05-27 | 浙江大学 | In-camera supervised cross-camera pedestrian re-identification method based on contrast learning |
| CN112990120A (en) * | 2021-04-25 | 2021-06-18 | 昆明理工大学 | Cross-domain pedestrian re-identification method using camera style separation domain information |
| CN113408351A (en) * | 2021-05-18 | 2021-09-17 | 河南大学 | Pedestrian re-recognition method for generating confrontation network based on attitude guidance |
| CN113255598A (en) * | 2021-06-29 | 2021-08-13 | 南京视察者智能科技有限公司 | Pedestrian re-identification method based on Transformer |
| CN113963206A (en) * | 2021-10-20 | 2022-01-21 | 中国石油大学(华东) | Posture guidance-based target detection method for fast skating athletes |
| CN114067356A (en) * | 2021-10-21 | 2022-02-18 | 电子科技大学 | Pedestrian re-identification method based on joint local guidance and attribute clustering |
| CN114120665A (en) * | 2022-01-29 | 2022-03-01 | 山东科技大学 | Intelligent phase control method and system based on pedestrian number |
| CN114550220A (en) * | 2022-04-21 | 2022-05-27 | 中国科学技术大学 | Training method of pedestrian re-recognition model and pedestrian re-recognition method |
| CN115762172A (en) * | 2022-11-02 | 2023-03-07 | 济南博观智能科技有限公司 | Method, device, equipment and medium for identifying vehicles entering and exiting parking places |
| CN115631464A (en) * | 2022-11-17 | 2023-01-20 | 北京航空航天大学 | Pedestrian three-dimensional representation method oriented to large space-time target association |
| CN115631464B (en) * | 2022-11-17 | 2023-04-04 | 北京航空航天大学 | Pedestrian three-dimensional representation method oriented to large space-time target association |
| CN116192997B (en) * | 2023-02-21 | 2023-12-01 | 兴容(上海)信息技术股份有限公司 | Event detection method and system based on network flow |
| CN116192997A (en) * | 2023-02-21 | 2023-05-30 | 上海兴容信息技术有限公司 | Event detection method and system based on network flow |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107832672B (en) | 2020-07-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107832672A (en) | A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information | |
| CN106709568B (en) | The object detection and semantic segmentation method of RGB-D image based on deep layer convolutional network | |
| CN107392964B (en) | The indoor SLAM method combined based on indoor characteristic point and structure lines | |
| CN104715493B (en) | A kind of method of movement human Attitude estimation | |
| CN104463117B (en) | A kind of recognition of face sample collection method and system based on video mode | |
| CN105869178B (en) | A kind of complex target dynamic scene non-formaldehyde finishing method based on the convex optimization of Multiscale combination feature | |
| CN106767812B (en) | A kind of indoor semantic map updating method and system based on Semantic features extraction | |
| CN100472556C (en) | Testing apparatus and method for special object | |
| CN108009509A (en) | Vehicle target detection method | |
| CN107145845A (en) | The pedestrian detection method merged based on deep learning and multi-characteristic points | |
| CN107871106A (en) | Face detection method and device | |
| CN112784736B (en) | Character interaction behavior recognition method based on multi-modal feature fusion | |
| CN109446925A (en) | A kind of electric device maintenance algorithm based on convolutional neural networks | |
| CN106845357A (en) | A kind of video human face detection and recognition methods based on multichannel network | |
| CN107358223A (en) | A kind of Face datection and face alignment method based on yolo | |
| CN105160310A (en) | 3D (three-dimensional) convolutional neural network based human body behavior recognition method | |
| CN103279768B (en) | A kind of video face identification method based on incremental learning face piecemeal visual characteristic | |
| CN109190508A (en) | A kind of multi-cam data fusion method based on space coordinates | |
| CN108734737A (en) | The method that view-based access control model SLAM estimation spaces rotate noncooperative target shaft | |
| CN104021381B (en) | Human movement recognition method based on multistage characteristics | |
| CN110532970A (en) | Age-sex's property analysis method, system, equipment and the medium of face 2D image | |
| CN103853794B (en) | Pedestrian retrieval method based on part association | |
| CN109934047A (en) | Face identification system and its face identification method based on deep learning | |
| CN109101864A (en) | The upper half of human body action identification method returned based on key frame and random forest | |
| CN106127733B (en) | The method and apparatus of human body target identification |
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 |