CN107038400A - Face identification device and method and utilize its target person tracks of device and method - Google Patents
Face identification device and method and utilize its target person tracks of device and method Download PDFInfo
- Publication number
- CN107038400A CN107038400A CN201610079687.1A CN201610079687A CN107038400A CN 107038400 A CN107038400 A CN 107038400A CN 201610079687 A CN201610079687 A CN 201610079687A CN 107038400 A CN107038400 A CN 107038400A
- Authority
- CN
- China
- Prior art keywords
- face
- collection
- target person
- face collection
- several
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
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
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
-
- 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
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N7/00—Television systems
- H04N7/18—Closed-circuit television [CCTV] systems, i.e. systems in which the video signal is not broadcast
- H04N7/181—Closed-circuit television [CCTV] systems, i.e. systems in which the video signal is not broadcast for receiving images from a plurality of remote sources
Abstract
The invention provides face identification device and method and utilize its target person tracks of device and method.Face identification device includes face collection generation module, is configured as generating face collection based on several video images;Characteristic measure normalizes module, is configured as carrying out metric space conversion to identical facial image under several different cameras, to eliminate the otherness under several different cameras between identical facial image;And characteristic envelope formation module, be configured as to metric space convert several facial images carry out envelope processing, using by multiple different face collection Feature Space Transformations of face collection as an identical face collection feature space.The otherness that the present invention is directed to different cameras carries out characteristic measure spatial normalization, and binding characteristic metric space and face image set are identified, and the precision of the recognition of face under monitoring scene greatly improved.
Description
Technical field
This invention relates generally to computer video processing technology field, filled more particularly, to recognition of face
Put, face identification method, the target person tracks of device using face identification device and utilize recognition of face side
The target person tracking of method, above-mentioned apparatus and method are adapted to urban safety management, such as airport, prison
With the public domain such as library.
Background technology
Existing face recognition technology has higher identification for the preferable face of picture quality.For
(it has been arranged the databases such as LFW by Massachusetts, USA university Amster branch school technology visual experiment room
Into) recognition of face precision can reach more than 99%.But if such algorithm is directly applied to monitoring
In scene, its effect will have a greatly reduced quality.By taking LFW databases as an example, although it is generally acknowledged that this database
It is the image shot under conditions of uncontrolled, but in resolution ratio, real colour degree, the side such as human face posture
Face is well many compared with the facial image of monitoring scene.Therefore, if directly transplanted general face recognition algorithms
Into the recognition of face of monitoring scene, its effect will decline a lot.
It is at following 2 points using the subject matter of above-mentioned LFW databases:1st, conventional recognition of face is general
Using single image, and in monitoring scene, because picture quality is poor, therefore single image information content has
Limit, it is impossible to ensure the accuracy of identification;2nd, the image imaging difference that conventional recognition of face is faced is little,
Therefore different camera images will not be typically directed to and carry out otherness reduction processing, and in monitoring scene,
Because face resolution ratio, color, posture between different cameras etc. can all have larger difference, it is therefore desirable to
Otherness is first reduced, the accuracy of identification is just can guarantee that.Therefore, LFW databases can not be applied to scene
Monitoring scene.
The more common recognition of face of recognition of face of monitoring scene has the characteristics of some are new.1st, single goal object
Can be described by multiple sequential charts, and scheme the description target characteristic that extracted feature can be more complete more.
At present, image-set method into the space similar to higher-dimension envelope can describe list by multiple figures
One target, calculates the minimum range between two envelopes for two different targets, is used as the two targets
Diversity factor, presently relevant algorithm has the image set based on affine hull apart from AHISD (Affine Hull
Based Image Set Distance), sparse approximate KNN point SANP, CRNP (Collaboratively
Regularized Nearest Points) and binary linear regression classification DLRC (Dual Linear Regression
Classification) etc..Although such algorithm can be than more rich description clarification of objective space, this
The envelope that class algorithm is formed for different target is easy to overlapping, and reason is that its feature has similitude, therefore
For different targets, its feature difference degree extracted must be larger, and the envelope of feature based formation is just not
Can be overlapping.2nd, there is different illumination, visual angle/posture and image resolution ratio across the monitoring scene of camera,
The precision of these meeting strong influence identifications.At present, metric-learning (that is, metric learning) side
Method can solve the problem that this problem.This kind of algorithm can measure the Projection Character in different characteristic space to identical
Space, can reduce the difference that same target is caused due to the difference of camera and posture/illumination/expression, increase
Plus the feature difference of different target.This feature of metric-learning methods can make up image-set (i.e.,
Image set, also known as atlas) method deficiency.Presently relevant algorithm has large-spacing nearest-neighbors LMNN
(Large margin nearest neighbor) etc..3rd, because monitoring camera needs continuously to record
Data, therefore the data volume of its storage is surprising, and target image is searched out in substantial amounts of data, need
The algorithm that efficiency is exceedingly fast is wanted, and is recognized again after being compressed to face characteristic, algorithm effect can greatly be improved
Rate.
Therefore, the problem of there is multi-cam recognition of face under the monitoring scene using multi-cam, due to
It is larger with conventional scenario otherness, therefore there is relatively big difference with conventional recognition of face in processing method.
The object tracked across camera is usually multi-source image, and the difference (manufacturer and model) of camera causes figure
As resolution ratio, color and deformation have relatively big difference;In addition, environment light source difference, artificially focusing are poor
Different and camera setting angle difference etc. can also allow the generation of different cameras image in same target difference
Property is larger;Secondly, in monitoring scene, the target range camera distance change of motion is larger, can cause
Imaging definition and color also have larger difference.The above difference is all the recognition of face institute of monitoring scene
The oddity problem faced.In addition, across camera tracking is that the mass data of multiple cameras is handled,
Therefore quick processing mass data is also the key issue that across camera tracking needs to solve.
The content of the invention
For the otherness and pin of the multi-cam recognition of face in the monitoring scene in the presence of prior art
To the technical problem of the mass data processing of multiple cameras, the invention provides can solve the problem that above-mentioned skill problem
A kind of face identification device and method, and target person tracks of device and method.
In order to solve drawbacks described above, present invention proposition is a kind of to be applied under the monitoring scene of large scale data
Across camera knowledge method for distinguishing.Advantage is provided in particular by following three kinds of modes:1st, it is directed to multiple
Camera, in order to reduce its otherness, uses characteristic measure spatial normalization;2nd, it is directed to frame of video
Continuity, extracts and merges multiple face characteristics to describe same target, make target signature space more complete;
And 3, for large-scale data set operation, use Feature Compression and two-stage matching process.
Include according to an aspect of the present invention there is provided a kind of face identification device:Face collection generation module,
It is configured as generating face collection based on several video images;Characteristic measure normalizes module, is configured as to many
Identical facial image carries out metric space conversion under width difference camera, to eliminate phase under several different cameras
With the otherness between facial image;And characteristic envelope formation module, it is configured as converting metric space
Several facial images carry out envelope processing, multiple different face collection feature spaces of face collection are become
It is changed to the feature space of an identical face collection.
Preferably, the otherness under several different cameras between identical facial image includes the difference of background light source
The otherness of the opposite sex and shooting angle.
Preferably, metric space conversion is carried out to identical facial image under several different cameras, it is many to eliminate
Otherness between width facial image further comprises:Several facial images are subjected to metric space conversion, with
Make identical facial image under several different cameras that there is identical background light source;And by several facial images
Metric space conversion is carried out, so that identical facial image has identical shooting angle under several different cameras.
Preferably, face identification device also includes:Identification module, is configured as the target person based on setting,
The target person matching degree highest target face collection selected with setting is concentrated from multiple faces.
Preferably, face collection generation module includes:Face detection module, is configured as obtaining with predetermined frame period
The video image taken carries out Face datection, wherein, predetermined frame period is determined according to video frame rate.
Preferably, face collection generation module also includes face alignment module, is configured as to carrying out Face datection
Video image carry out following image procossing:It is configured as from the extraction key point in video image, and root
Change of scale is carried out to face according to pupil of both eyes distance;And the position of the key point according to extraction, estimate people
The angle of row partially of face, and face is rotated to be by frontal faces according to the yaw angle of face, generation includes same mesh
The face collection of several facial images of people is marked, wherein, key point includes eyes, nose and face.
Preferably, characteristic measure normalization module also includes:Characteristic extracting module, feature is carried out to face collection
Extract, to form primitive character face collection;And compression module, it is configured as being compressed primitive character
Processing, to form compressive features face collection.
Preferably, identification module also includes:Thick match cognization module, is configured as the target person based on setting,
It is determined that the face collection of orderly predetermined quantity;And smart match cognization module, it is configured as the mesh based on setting
People is marked, quick determination target face collection is concentrated from the face of orderly predetermined quantity.
The target person of setting is preferably based on, the people for determining predetermined quantity is concentrated from multiple compressive features faces
Face collection;According to the matching degree with the target person of setting, with the order of matching degree from high to low to predetermined quantity
Face collection is ranked up, to generate the face collection of orderly predetermined quantity;And the target person based on setting,
Concentrate quick true from the primitive character face of the predetermined quantity corresponding with the face collection of orderly predetermined quantity
Set the goal face collection.
According to another aspect of the present invention there is provided a kind of face identification method, comprise the following steps:It is based on
Several video images generate face collection;Metric space conversion is carried out to several facial images that face is concentrated, with
Eliminate the otherness between identical facial image under several different cameras;And to many of metric space conversion
Width facial image carries out envelope processing, is by multiple different face collection Feature Space Transformations of face collection
The feature space of one identical face collection.
Preferably, the otherness under several different cameras between identical facial image includes the difference of background light source
The otherness of the opposite sex and shooting angle.
Preferably, metric space conversion is carried out to identical facial image under several different cameras, it is many to eliminate
Otherness under width difference camera between identical facial image further comprises:By under several different cameras
Identical facial image carries out metric space conversion, so that identical facial image has phase under several different cameras
Same background light source;And several facial images are subjected to metric space conversion, so that several facial images have
There is identical shooting angle.
Preferably, face identification method also includes:Target person based on setting, concentrates from multiple faces and selects
Go out the target person matching degree highest target face collection with setting.
Preferably, face identification method also includes:Before generation face collection, with predetermined frame interval acquiring
Video image carries out Face datection, wherein, predetermined frame period is determined according to video frame rate.
Preferably, before metric space conversion is carried out to several facial images, regarded to carrying out Face datection
Frequency image carries out following image procossing:From the extraction key point in video image, and according to pupil of both eyes away from
Change of scale is carried out to face;And the position of the key point according to extraction, estimate the angle of row partially of face,
And face is rotated to be by frontal faces according to the yaw angle of face, generation includes several faces of same target person
The face collection of image, wherein, key point includes eyes, nose and face.
Preferably, also include before several facial images concentrated to face carry out metric space conversion following
Step:Feature extraction is carried out to face collection, to form primitive character face collection;And be configured as to original
Feature is compressed processing, to form compressive features face collection.
Preferably, face identification method further comprises:Target person based on setting, it is determined that orderly is predetermined
The face collection of quantity;And the target person based on setting, concentrate quick true from the face of orderly predetermined quantity
Set the goal face collection.
Preferably, face identification method further comprises:Target person based on setting, from multiple compressive features
Face concentrates the face collection for determining predetermined quantity;According to the matching degree with the target person of setting, with matching degree from
High to Low order is ranked up to the face collection of predetermined quantity, to generate the face collection of orderly predetermined quantity;
And the target person based on setting, from the original of the predetermined quantity corresponding with the face collection of orderly predetermined quantity
Beginning eigenface concentrates quick determination target face collection.
According to another aspect of the invention there is provided a kind of target person tracks of device, including:Face collection is generated
Module, is configured as generating face collection based on several video images;Characteristic measure normalizes module, is configured
To carry out metric space conversion to several facial images, to eliminate the otherness between several facial images;It is special
Envelope formation module is levied, is configured as carrying out envelope processing to several facial images that metric space is converted,
So that the feature of multiple different face collection Feature Space Transformations of face collection as identical face collection is empty
Between;Face recognition module, the target person based on setting concentrates the target selected with setting from multiple faces
The facial image matching degree highest target face collection of people;Tracking module, according to each of target face concentration
The camera site of width facial image and time, determine the route of target person.
According to another aspect of the invention there is provided a kind of target person tracking, comprise the following steps:Base
Face collection is generated in several video images;Metric space conversion is carried out to several facial images that face is concentrated,
To eliminate the otherness between several facial images;And several facial images that metric space is converted are carried out
Envelopeization processing, using by multiple different face collection Feature Space Transformations of face collection as an identical face
The feature space of collection;Target person based on setting, concentrates from multiple faces and selects and the target person of setting
Facial image matching degree highest target face collection;The bat for each width facial image concentrated according to target face
Act as regent and put and the time, determine the route of target person.
Present invention is specifically directed to the face identification device of monitoring scene and method.It is directed to different cameras
Otherness carry out characteristic measure spatial normalization, and binding characteristic metric space and face image set enter
Row identification, greatly improved the precision of the recognition of face under monitoring scene.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to implementing
The accompanying drawing used required in example is briefly described, it should be apparent that, drawings in the following description are only
Some embodiments of the present invention, for those of ordinary skill in the art, are not paying creative work
Under the premise of, other accompanying drawings can also be obtained according to these accompanying drawings.
Fig. 1 is the block diagram of face identification device according to an embodiment of the invention;
Fig. 2 is the flow chart of face identification method according to an embodiment of the invention;
Fig. 3 is that the face identification method according to an embodiment of the invention in Fig. 2 further comprises
Feature extraction and the schematic diagram of compression step.
Fig. 4 is that the envelopeization processing according to an embodiment of the invention in face identification method is generated
The schematic diagram of characteristic envelope;
Fig. 5 is the block diagram of target person tracks of device according to an embodiment of the invention;
Fig. 6 is according to an embodiment of the invention in the flow chart of target person tracking;
Fig. 7 is two stages according to an embodiment of the invention conducted in target person tracking process
The schematic diagram of matching;
Fig. 8 A, 8B and 8C respectively show target person tracking process according to an embodiment of the invention
In in a camera track target person, in multi-cam continue track target person and by with
The shown system diagram in track of the track in map.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is entered
Row is clearly and completely described, it is clear that described embodiment is only a part of embodiment of the invention,
Rather than whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art are obtained
The every other embodiment obtained, belongs to the scope of protection of the invention.
Fig. 1 is the block diagram of face identification device according to an embodiment of the invention.Fig. 2 is according to this hair
The flow chart of the face identification method of bright embodiment.
With reference to Fig. 1, face identification device includes face collection generation module 102, characteristic measure normalization mould
Block 104 and characteristic envelope formation module 106.Specifically, face collection generation module 102 is configured as
Face collection is generated based on several video images.Characteristic measure normalization module 104 is configured as to several
Identical facial image carries out metric space conversion under different cameras, to eliminate under several different cameras
Otherness between identical facial image.Characteristic envelope formation module 106 is configured as to metric space
Several facial images of conversion carry out envelope processing, by multiple different face Ji Te of face collection
Levy feature space of the spatial alternation for an identical face collection.
In one embodiment, face collection generation module 102 also includes:Face detection module, by with
It is set to and Face datection is carried out with the video image of predetermined frame interval acquiring, wherein, entered according to video frame rate
Row determines predetermined frame period.In addition, face collection generation module 102 also includes face alignment module, quilt
It is configured to carry out following image procossing to the video image for carrying out Face datection:It is configured as from video figure
Extraction key point as in, and change of scale is carried out to face according to pupil of both eyes distance;And root
According to the position of the key point of extraction, the angle of row partially of face is estimated, and will according to the yaw angle of face
Face rotates to be frontal faces, and generation includes the face collection of several facial images of same target person, wherein,
Key point includes eyes, nose and face.
Specifically, the otherness under several different cameras between identical facial image includes the difference of background light source
The otherness of the opposite sex and shooting angle.Module 104 is normalized under several different cameras using characteristic measure
Identical facial image carries out metric space conversion, is further wrapped with eliminating the otherness between several facial images
Include:Several facial images are subjected to metric space conversion, so that identical facial image under several different cameras
With identical background light source;And several facial images are subjected to metric space conversion, so that several are different
Identical facial image has identical shooting angle under camera.In addition, characteristic measure normalizes module 104
Also include characteristic extracting module, feature extraction is carried out to face collection, to form primitive character face collection;And
Compression module, is configured as being compressed primitive character processing, to form compressive features face collection.
In addition, face identification device also includes identification module, the target person based on setting is configured as, from many
Individual face concentrates the target person matching degree highest target face collection selected with setting.Identification module also includes:
Thick match cognization module, is configured as the target person based on setting, it is determined that the face collection of orderly predetermined quantity;
And smart match cognization module, the target person based on setting is configured as, from the face of orderly predetermined quantity
Concentrate quick determination target face collection.Specifically, the target person based on setting, from multiple compressive features faces
Concentrate the face collection for determining predetermined quantity;According to the matching degree with the target person of setting, with matching degree from height to
Low order is ranked up to the face collection of predetermined quantity, to generate the face collection of orderly predetermined quantity;With
And the target person based on setting, from the original of the predetermined quantity corresponding with the face collection of orderly predetermined quantity
Eigenface concentrates quick determination target face collection.
With reference to Fig. 2, face identification method comprises the following steps:In step 210, regarded based on several
Frequency image generates face collection;In a step 220, several facial images that face is concentrated are measured
Spatial alternation, to eliminate the otherness under several different cameras between identical facial image;And
In step 230, envelope processing is carried out to several facial images that metric space is converted, by face
Multiple different face collection Feature Space Transformations of collection are the feature space of an identical face collection.
Face identification method also includes:Before generation face collection, with the video image of predetermined frame interval acquiring
Face datection is carried out, wherein, predetermined frame period is determined according to video frame rate.In addition, to several people
Face image is carried out before metric space conversion, and following image procossing is carried out to the video image for carrying out Face datection:
Change of scale is carried out to face from the extraction key point in video image, and according to pupil of both eyes distance;With
And the position of the key point according to extraction, estimate the angle of row partially of face, and will according to the yaw angle of face
Face rotates to be frontal faces, and generation includes the face collection of several facial images of same target person, wherein, close
Key point includes eyes, nose and face.
Otherness under several different cameras between identical facial image include background light source otherness and
The otherness of shooting angle.Before several facial images concentrated to face carry out metric space conversion also
Including:Feature extraction is carried out to face collection, to form primitive character face collection;And be configured as to original
Feature is compressed processing, to form compressive features face collection.To identical face figure under several different cameras
As carrying out metric space conversion, entered with eliminating the otherness under several different cameras between identical facial image
One step includes:Identical facial image under several different cameras is subjected to metric space conversion, so that several are not
With under camera, identical facial image has identical background light source;And measured several facial images
Spatial alternation, so that several facial images have identical shooting angle.
In addition, face identification method also includes:Target person based on setting, concentrates from multiple faces and selects
With the target person matching degree highest target face collection of setting.Face identification method further comprises:Based on setting
Fixed target person, it is determined that the face collection of orderly predetermined quantity;And the target person based on setting, from orderly
The face of predetermined quantity concentrate and quick determine target face collection.Specifically, face identification method is further wrapped
Include:Target person based on setting, the face collection for determining predetermined quantity is concentrated from multiple compressive features faces;Root
According to the matching degree with the target person of setting, the face collection of predetermined quantity is entered with the order of matching degree from high to low
Row sequence, to generate the face collection of orderly predetermined quantity;And the target person based on setting, from it is orderly
The primitive character face of the corresponding predetermined quantity of face collection of predetermined quantity concentrate and quick determine target person
Face collection.
According to the abovementioned embodiments of the present invention, using the face identification device and method of monitoring scene, its
In, the otherness for being directed to different cameras carries out characteristic measure spatial normalization, and binding characteristic degree
Quantity space and face image set are identified, and can increase substantially the precision of recognition of face.
Referring below to Fig. 3 to Fig. 8 C to target person tracks of device and target person tracking, then
Included face identification device is described in detail and to target person in target person tracks of device
Face identification method used in tracking is described in detail.
Fig. 5 is the block diagram of target person tracks of device according to an embodiment of the invention.With reference first to Fig. 5
Target person tracks of device is described in detail.With reference to Fig. 5, according to another embodiment of the present invention,
Target person tracks of device includes face collection generation module 502, characteristic measure normalization module 504, feature
Envelope formation module 506, face recognition module 510 and tracking module 512.Face collection generation module
502nd, characteristic measure normalization module 504, characteristic envelope formation module 506 are included in recognition of face dress
In putting.In a preferred embodiment, face collection generation module 502, the characteristic measure normalizing shown in Fig. 5
Mould can be generated with the face collection shown in Fig. 1 respectively by changing module 504, characteristic envelope formation module 506
Block 102, characteristic measure normalization module 104 and characteristic envelope formation module 106 are same or similar.
Hereinafter, the modules in target person tracks of device will be respectively described in detail.
Face collection generation module 502 is configured as generating face collection based on several video images.At one
In embodiment, face collection generation module 502 also includes:Face detection module, is configured as with predetermined
The video image that frame period is obtained carries out Face datection, wherein, it is determined according to video frame rate predetermined
Frame period.In addition, face collection generation module 502 also includes face alignment module, it is configured as to entering
The video image of row Face datection carries out following image procossing:It is configured as from the extraction in video image
Key point, and change of scale is carried out to face according to pupil of both eyes distance;And according to the pass of extraction
The position of key point, estimates the angle of row partially of face, and is rotated to be face according to the yaw angle of face
Frontal faces, generation includes the face collection of several facial images of same target person, wherein, key point bag
Include eyes, nose and face.
Hereinafter, the example to face generation module 502 is described in detail.In instantiation,
Face detection and tracking is carried out by face detection module first, then carried out by face alignment module
Face aligns.Specifically, monitor video is decomposed into picture, pedestrian is entered to video pictures every several frames
Face is detected.The frame number at interval is determined according to actual video frame rate.The human face region detected, as far as possible
Few includes background information, reduces influence of the ambient noise for identification.In order to obtain same person
Plurality of pictures to facial image, it is necessary to be tracked.Face tracking might have two kinds of mistakes:1st, with
The target of track is not belonging to same person;Solution is:Detection consecutive frame facial image is similar in real time
Degree, if similarity be less than threshold value, the ID new to tracked image setting, wherein, the threshold value can
To be needed to be configured according to user;2nd, the target of tracking belongs to same person, but tracing figure picture without
Method accurately covers human face region.Solution is:Utilize the nearest Face datection figure of face tracking frame
As amendment face tracking result.
In instantiation, alignd followed by face.Face alignment influences larger to recognition result.
Eyes, nose, the extraction of four key points of face are carried out to face first, then according to two eye pupil holes
Distance to face carry out change of scale;So the angle between latter two interpupillary line and horizontal line, right
Facial image rotates to be level;According to the position of this four key points, the yaw angle of face is estimated,
Face is rotated into frontal faces.If this four key points of face can not be extracted, then to these people
Face image is without processing., can be to this in the case where the amount of images of single image collection is relatively more
The image that a little key points can not be extracted carries out discard processing, but is the need to ensure that the image of single image collection
Quantity is more than 15.
Target person tracks of device also includes characteristic measure and normalizes module 504, is configured as to several people
Face image carries out metric space conversion, to eliminate the otherness between several facial images, specifically,
Otherness under several different cameras between identical facial image includes otherness and the bat of background light source
Take the photograph the otherness of angle etc..Specifically, identical facial image under several different cameras is measured
Spatial alternation, is further comprised with eliminating the otherness between several facial images:By several face figures
As carrying out metric space conversion, so that identical facial image has the identical back of the body under several different cameras
Scape light source;And several facial images are subjected to metric space conversion, so that under several different cameras
Identical facial image has identical shooting angle.In addition, characteristic measure normalization module 504 is also wrapped
Characteristic extracting module is included, feature extraction is carried out to face collection, to form primitive character face collection;And
Compression module, is configured as being compressed primitive character processing, to form compressive features face collection.
Otherness and the shooting of background light source can be greatly lowered using this feature measurement normalization module 504
The otherness of angle etc..
Hereinafter, the example that module 504 is normalized to characteristic measure is described in detail.Specific
In example, when carrying out the multi-cam characteristic measure spatial normalization based on face collection, pass through first
Characteristic extracting module and compression module face detection module carry out Face detection and tracking, then pass through spy
Levy measurement normalization module 504 and carry out multi-cam characteristic measure spatial normalization.
Fig. 3 is that the face identification method according to an embodiment of the invention in Fig. 2 further comprises
Feature extraction and the schematic diagram of compression step.With reference to Fig. 3, pass through characteristic extracting module and compression module
Face detection module carries out Face detection and tracking.In instantiation, to the facial image after alignment
Collection carries out feature extraction, and it can be LBP (that is, local binary patterns), Gabor to extract feature
(that is, covering rich) or DCT (that is, discrete cosine transform) feature.Generally, for 32 × 32
Or 64 × 64 image, its characteristic dimension can reach more than 1000, although single image processing consumption
When it is few, but if when amount of images reaches 100,000 grades, its time-consuming length will be extremely considerable,
Therefore need to carry out Feature Compression processing to extracted feature.The feature extracted is sampled;So
The feature sampled is compressed using WTA hash compress techniques afterwards, compression result is:It is local
Eigenvalue of maximum is 1, remaining characteristic value boil down to 0;Finally, the compression of all regional areas is connected
Get up, its result is the compression result of this characteristic vector.During Feature Compression, between feature sampling
Determined according to the actual requirements every, regional area length.Characteristic vector after compression is compressed for two-value, pressure
Characteristic dimension after contracting will be reduced.
In instantiation, multi-cam characteristic measure spatial normalization is described in detail.This process
Need first to train and then reuse.Transformation matrix instruction is carried out first by the sample between different cameras
Practice, then this transformation matrix is applied on corresponding camera in use.Characteristic measure is empty
Between normalized purpose be reduce same person under different cameras because the difference of illumination and angle is led
The otherness of cause.For example, it is the right side to normalize module 504 by left side light change by this feature measurement
Sidelight is left side illumination according to or by right side light change.In addition, for example by different cameras relative to
The different angular transformations that target person is shot are that target person is shot in equal angular.This process
Characteristic vector after compression is subjected to metric space conversion, by the changing features of different metric spaces to together
One metric space, i.e., with identical metric.The method that this process has used metric-learning.
Metric-learning can be by input feature vector spatial alternation to the space with yardstick meaning:
L is transformation matrix., it is necessary to full when by the characteristic measure standard projection of different spaces to the same space
The following two conditions of foot:Different ID clarifications of objective space lengths are become big by 1, and apart from more than safety
Distance;2 reduce identical ID target signatures space length.In this manner it is possible to avoid the feelings of misclassification
Condition occurs.Before its error function is described, symbol is first defined as follows:
Input feature value,
yij:IfWithIt is 1 with same label, is otherwise 0,
ηij:IfAdjacent target beWhen be 1, be otherwise 0.
Its error function is as follows:
ε (L)=εpull(L)+εpush(L)
Wherein
εpullFor the error function that same target in neighborhood furthers, εpushFor different target in neighborhood is pushed away far
Error function.Because direct solution L may result in locally optimal solution, therefore entered line translation:
M=LTL, such ε (M) will have globally optimal solution, and after corresponding eigentransformation is apart from formal argument:Final error functional form is:
Minimize:
ε (M)=εpull(M)+εpush(M)
Subject to:
M≥0
Optimal value is solved is approached using tree recurrence progress gradient, can in complex situations such as linearly inseparable
Introduce geo-nuclear tracin4.
Next, characteristic envelope formation module 506 is configured as several faces converted to metric space
Image carry out envelope processing, using by multiple different face collection Feature Space Transformations of face collection as one
The feature space 508 of individual identical face collection.Specifically, to identical face under several different cameras
Image carries out metric space conversion, is further comprised with eliminating the otherness between several facial images:
Several facial images are subjected to metric space conversion, so that identical facial image under several different cameras
With identical background light source;And several facial images are subjected to metric space conversion, so that several
Identical facial image has identical shooting angle under different cameras.
Fig. 4 is that the envelopeization processing according to an embodiment of the invention in face identification method is generated
The schematic diagram of characteristic envelope.Hereinafter, 4 pairs of reference picture is carried out by characteristic envelope formation module
Envelopeization processing be described in detail.
With some images it is image set to describe face be in order to expand the feature space of face, more entirely
The description face in face.It is special to compression face on the basis of its multiple face figure is extracted to same target
Levy after progress metric space conversion, envelope processing is carried out to it using image-set methods, most terminated
Fruit be by multiple image appearances of same target be different weighted array.So-called envelope refers to identical
The different picture formation of people the feature space that is formed of an image set.Assuming that XcFor comprising some
The feature set of characteristics of image, Xc,iFor one of sample, wherein c=1 ..., C, i=1 ..., nc。
For example, ncFor 5-50 or any other positive integer, it is preferable that be 15,20 and 50.More
Preferably, n is worked ascFor 5 when, efficiency highest.Envelope is expressed as Hc, concrete form is:
Wherein, αc,iFor the weight of each sample.There are different acquisition methods for weight, in the present invention
In, using the method for DLRC weight calculations.
Face recognition module (also known as identification module) 510 target persons based on setting, from multiple faces
Concentrate the facial image matching degree highest target face collection selected with the target person of setting.Recognize mould
Block also includes:Thick match cognization module, is configured as the target person based on setting, it is determined that orderly is pre-
The face collection of fixed number amount;And smart match cognization module, the target person based on setting is configured as, from
The face of orderly predetermined quantity concentrates quick determination target face collection.More specifically, based on setting
Target person, the face collection for determining predetermined quantity is concentrated from multiple compressive features faces;According to setting
The matching degree of target person, is ranked up with the order of matching degree from high to low to the face collection of predetermined quantity,
To generate the face collection of orderly predetermined quantity;And the target person based on setting, from it is orderly pre-
The primitive character face of the corresponding predetermined quantity of the face collection of fixed number amount, which is concentrated, quick determines target face
Collection.
In the following example, 7 and 8A to 8C knows to across camera face collection envelope with reference to the accompanying drawings
It is not described in detail.
The characteristic envelope of each face collection forms rear, it is necessary to measure the otherness of different face collection features,
Otherness is determined by calculating the minimum range of face collection envelope, and minimum range is added by asking two groups
The minimum residual error of feature is weighed to determine.Fig. 3 show the characteristic envelope schematic diagram of different face collection,
There are three face collection in figure, everyone is made up of face collection some facial images.In actual identification process
In, each face collection is lineup's face figure feature, and each face figure feature includes two features:It is original
Feature after feature and compression.Primitive character is that image set feature and compressive features in accompanying drawing 3 are figure
Image set Hash coding characteristic in 3.The characteristics of primitive character and compressive features is respectively:Original spy
Levy discrimination higher, but required time is long;Compressive features discrimination is less than primitive character, but institute
Take time short.Identification process is divided into two stages:
1st, thick matching stage:Using feature after compression, object is all identification face collection;This stage speed
Degree is preferential;
2nd, accurate matching stage:Using primitive character, object is N earlier in thick matching result
Individual face collection;This stage precision is preferential.
Fig. 4 is the two-stage to match schematic diagram, in thick matching stage, and all test samples are used to match,
The distance for carrying out face image set using the feature after compression is calculated, then basis and face to be matched
Distance is ranked up to test sample.It may be made a mistake matching problem in thick matching stage, such as more
Behind similar face comes, it is therefore desirable to accurate to match to adjust the result slightly matched.Accurate matching
It is used to be characterized as original only to carrying out computing according to several image sets of distance-taxis up front
Beginning feature.For example, in thick matching stage, can concentrate fast from all faces compressed to be recognized
Speed selects the orderly face collection of 2-50 or any other quantity;It is any in accurate matching stage,
The orderly face obtained from thick matching stage, which is concentrated, determines target face collection.The number of orderly face collection
Amount can need to be determined according to user, for example, 5,10 and 15.
The camera site for each width facial image that tracking module 512 is concentrated according to target face and time,
Determine the route of target person.Specifically, the target person with setting selected in face recognition module
After facial image matching degree highest target face collection, tracking module 512 is according to the target obtained
The camera site for each width facial image that face is concentrated and time, the fortune of this person is shown on map
Dynamic rail mark.
Under the scene that across camera camera is tracked, the present invention can be effectively to different cameral camera
In same person be tracked and recognize, the movement locus of this person is finally shown on map.Figure
8A to Fig. 8 C respectively show following result:(a) target person is tracked in a camera camera
Thing;(b) continue to track target person in multiple camera cameras;(c) show that tracked people exists
Track in map.
Fig. 6 is according to an embodiment of the invention in the flow chart of target person tracking.Below will ginseng
Target person tracking is described in detail according to Fig. 6.
Target person tracking comprises the following steps:
In step 610, face collection is generated based on several video images;Before generation face collection,
Face datection is carried out with the video image of predetermined frame interval acquiring, wherein, carried out according to video frame rate true
Fixed predetermined frame period.Before metric space conversion is carried out to several facial images, to carrying out face inspection
The video image of survey carries out following image procossing:From the extraction key point in video image, and according to
Pupil of both eyes distance carries out change of scale to face;And the position of the key point according to extraction, estimation
The angle of row partially of face, and face is rotated to be by frontal faces according to the yaw angle of face, generation includes
The face collection of several facial images of same target person, wherein, key point includes eyes, nose and mouth
Bar.
In step 620, metric space conversion is carried out to several facial images that face is concentrated, to disappear
Except the otherness between several facial images;Before metric space conversion is carried out, face collection is carried out
Feature extraction, to form primitive character face collection;And be configured as being compressed primitive character at place
Reason, to form compressive features face collection.Wherein, under several different cameras between identical facial image
Otherness including background light source otherness and the otherness of shooting angle.Specifically, to several not
With under camera, identical facial image carries out metric space conversion, to eliminate phase under several different cameras
Further comprise with the otherness between facial image:By identical facial image under several different cameras
Metric space conversion is carried out, so that identical facial image has identical background under several different cameras
Light source;And several facial images are subjected to metric space conversion, so that several facial images have phase
Same shooting angle.
In act 630, envelope processing is carried out to several facial images that metric space is converted, with
By the feature sky that multiple different face collection Feature Space Transformations of face collection are an identical face collection
Between;
In step 640, the target person based on setting is concentrated from multiple faces and selected and setting
The facial image matching degree highest target face collection of target person.Specifically, the target person based on setting,
It is determined that the face collection of orderly predetermined quantity;And the target person based on setting, from orderly predetermined number
The face of amount concentrates quick determination target face collection.More specifically, the target person based on setting, from many
Individual compressive features face concentrates the face collection for determining predetermined quantity;Matched according to the target person of setting
Degree, is ranked up with the order of matching degree from high to low to the face collection of predetermined quantity, to generate in order
Predetermined quantity face collection;And the target person based on setting, from the people with orderly predetermined quantity
The primitive character face of the corresponding predetermined quantity of face collection concentrates quick determination target face collection.
In step 650, according to target face concentrate each width facial image camera site and when
Between, determine the route of target person.
Therefore, embodiments of the invention are compressed to the invariant features extracted first, Ran Hou
On the basis of image-set methods and metric-learning, across camera recognition methods is proposed, prison is solved
Control under scene the problem of camera is tracked.
Therefore, it can be improved using the face identification device and face identification method of embodiments of the invention
Recognition of face precision under monitoring scene.Filled in addition, being tracked using the target person of the face identification device
Put and can not only provide recognition of face essence using the target person tracking of face identification method
Degree, and can be exceedingly fast by the matched and searched target person in two stages in substantial amounts of data
Speed search out target image.That is, in actual applications, the present invention can aid in using
Person finds target person from high-volume database and extracts its time and routing information, greatly improves work
Efficiency.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in this hair
Within bright spirit and principle, any modification, equivalent substitution and improvements made etc. should be included in this hair
Within bright protection domain.
Claims (20)
1. a kind of face identification device, it is characterised in that including:
Face collection generation module, is configured as generating face collection based on several video images;
Characteristic measure normalizes module, is configured as entering identical facial image under several described different cameras
Row metric space is converted, to eliminate the otherness under several described different cameras between identical facial image;
And
Characteristic envelope formation module, is configured as wrapping several facial images described in metric space conversion
Networkization processing, using by multiple different face collection Feature Space Transformations of the face collection as an identical people
The feature space of face collection.
2. face identification device according to claim 1, it is characterised in that several described differences are taken the photograph
Include the otherness of background light source and the difference of shooting angle as the otherness between identical facial image under head
Property.
3. face identification device according to claim 2, it is characterised in that to several described differences
Identical facial image carries out metric space conversion under camera, to eliminate the difference between several described facial images
The opposite sex further comprises:
Several described facial images are subjected to metric space conversion, so that identical under several described different cameras
Facial image has identical background light source;And
Several described facial images are subjected to metric space conversion, so that identical under several described different cameras
Facial image has identical shooting angle.
4. face identification device according to claim 1, it is characterised in that also include:Recognize mould
Block, is configured as the target person based on setting, and the target person selected with the setting is concentrated from multiple faces
Matching degree highest target face collection.
5. face identification device according to claim 1, it is characterised in that the face collection generation
Module includes:Face detection module, is configured as carrying out face inspection with the video image of predetermined frame interval acquiring
Survey, wherein, the predetermined frame period is determined according to video frame rate.
6. face identification device according to claim 5, it is characterised in that the face collection generation
Module also includes face alignment module, is configured as follows to the video image progress for carrying out Face datection
Image procossing:
It is configured as from the extraction key point in the video image, and according to pupil of both eyes distance to face
Carry out change of scale;And
According to the position of the key point of extraction, the angle of row partially of face is estimated, and according to the face
Face is rotated to be frontal faces by yaw angle, and generation includes the people of several facial images of the same target person
Face collection, wherein, the key point includes eyes, nose and face.
7. face identification device according to claim 1, it is characterised in that the characteristic measure is returned
One change module also includes:
Characteristic extracting module, carries out feature extraction, to form primitive character face collection to the face collection;With
And
Compression module, is configured as being compressed processing to the primitive character, to form compressive features face
Collection.
8. face identification device according to claim 4, it is characterised in that the identification module is also
Including:
Thick match cognization module, is configured as the target person based on the setting, it is determined that orderly predetermined quantity
Face collection;And
Smart match cognization module, is configured as the target person based on the setting, from the orderly predetermined number
The face of amount is concentrated and quickly determines the target face collection.
9. the face identification device according to any one of claim 7 and 8, it is characterised in that
Based on the target person of the setting, the face collection for determining predetermined quantity is concentrated from multiple compressive features faces;
According to the matching degree with the target person of the setting, with the order of matching degree from high to low to described predetermined
The face collection of quantity is ranked up, to generate the face collection of the orderly predetermined quantity;And
Based on the target person of the setting, make a reservation for from corresponding with the face collection of the orderly predetermined quantity
The primitive character face of quantity is concentrated and quickly determines the target face collection.
10. a kind of face identification method, it is characterised in that comprise the following steps:
Face collection is generated based on several video images;
Metric space conversion is carried out to several facial images that the face is concentrated, to eliminate several described differences
Otherness under camera between identical facial image;And
Envelope processing is carried out to several facial images described in metric space conversion, by the face collection
Multiple different face collection Feature Space Transformations are the feature space of an identical face collection.
11. face identification method according to claim 10, it is characterised in that several described differences
Otherness under camera between identical facial image includes the otherness of background light source and the difference of shooting angle
The opposite sex.
12. face identification method according to claim 11, it is characterised in that to it is described several not
With under camera, identical facial image carries out metric space conversion, to eliminate phase under several described different cameras
Further comprise with the otherness between facial image:
Metric space conversion will be carried out by identical facial image under several described different cameras so that it is described several
Identical facial image has identical background light source under different cameras;And
Several described facial images are subjected to metric space conversion so that several described facial images have it is identical
Shooting angle.
13. face identification method according to claim 10, it is characterised in that also include:It is based on
The target person of setting, the target person matching degree highest target selected with the setting is concentrated from multiple faces
Face collection.
14. face identification method according to claim 10, it is characterised in that also include:In life
Into before the face collection, Face datection is carried out with the video image of predetermined frame interval acquiring, wherein, according to
Video frame rate is determined the predetermined frame period.
15. face identification method according to claim 14, it is characterised in that to it is described several
Facial image is carried out before metric space conversion, and below figure is carried out to the video image for carrying out Face datection
As processing:
Yardstick is carried out to face from the extraction key point in the video image, and according to pupil of both eyes distance
Conversion;And
According to the position of the key point of extraction, the angle of row partially of face is estimated, and according to the face
Face is rotated to be frontal faces by yaw angle, and generation includes the people of several facial images of the same target person
Face collection, wherein, the key point includes eyes, nose and face.
16. face identification method according to claim 10, it is characterised in that to the face collection
In several facial images carry out metric space conversion it is further comprising the steps of:
Feature extraction is carried out to the face collection, to form primitive character face collection;And
It is configured as being compressed processing to the primitive character, to form compressive features face collection.
17. the face identification method according to claim 13, it is characterised in that further comprise:
Based on the target person of the setting, it is determined that the face collection of orderly predetermined quantity;And
Based on the target person of the setting, concentrated from the face of the orderly predetermined quantity described in quick determine
Target face collection.
18. the face identification method according to any one of claim 16 and 17, it is characterised in that
Further comprise:
Based on the target person of the setting, the face collection for determining predetermined quantity is concentrated from multiple compressive features faces;
According to the matching degree with the target person of the setting, with the order of matching degree from high to low to described predetermined
The face collection of quantity is ranked up, to generate the face collection of the orderly predetermined quantity;And
Based on the target person of the setting, make a reservation for from corresponding with the face collection of the orderly predetermined quantity
The primitive character face of quantity is concentrated and quickly determines the target face collection.
19. a kind of target person tracks of device, it is characterised in that including:
Face collection generation module, is configured as generating face collection based on several video images;
Characteristic measure normalizes module, is configured as carrying out metric space conversion to several described facial images,
To eliminate the otherness between several described facial images;
Characteristic envelope formation module, is configured as wrapping several facial images described in metric space conversion
Networkization processing, using by multiple different face collection Feature Space Transformations of the face collection as an identical people
The feature space of face collection;
Face recognition module, the target person based on setting is concentrated from multiple faces and selected and the setting
The facial image matching degree highest target face collection of target person;
Tracking module, the camera site for each width facial image concentrated according to the target face and time,
Determine the route of the target person.
20. a kind of target person tracking, it is characterised in that comprise the following steps:
Face collection is generated based on several video images;
Metric space conversion is carried out to several facial images that the face is concentrated, to eliminate several described faces
Otherness between image;And
Envelope processing is carried out to several facial images described in metric space conversion, by the face collection
Multiple different face collection Feature Space Transformations are the feature space of an identical face collection;
Target person based on setting, the face figure selected with the target person of the setting is concentrated from multiple faces
As matching degree highest target face collection;
The camera site for each width facial image concentrated according to the target face and time, determine the mesh
Mark the route of people.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610079687.1A CN107038400A (en) | 2016-02-04 | 2016-02-04 | Face identification device and method and utilize its target person tracks of device and method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610079687.1A CN107038400A (en) | 2016-02-04 | 2016-02-04 | Face identification device and method and utilize its target person tracks of device and method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107038400A true CN107038400A (en) | 2017-08-11 |
Family
ID=59532526
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610079687.1A Pending CN107038400A (en) | 2016-02-04 | 2016-02-04 | Face identification device and method and utilize its target person tracks of device and method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107038400A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109670451A (en) * | 2018-12-20 | 2019-04-23 | 天津天地伟业信息系统集成有限公司 | Automatic face recognition tracking |
CN109753857A (en) * | 2017-11-07 | 2019-05-14 | 北京虹图吉安科技有限公司 | A kind of three-dimensional face identification apparatus and system based on photometric stereo visual imaging |
CN109993039A (en) * | 2018-01-02 | 2019-07-09 | 上海银晨智能识别科技有限公司 | Portrait identification method and device, computer readable storage medium |
CN110472460A (en) * | 2018-05-11 | 2019-11-19 | 北京京东尚科信息技术有限公司 | Face image processing process and device |
CN110826435A (en) * | 2019-10-23 | 2020-02-21 | 上海能塔智能科技有限公司 | Method and device for identifying and verifying face in driving, vehicle-mounted equipment and storage medium |
CN111131700A (en) * | 2019-12-25 | 2020-05-08 | 重庆特斯联智慧科技股份有限公司 | Hidden tracking equipment for intelligent security and use method |
WO2020094091A1 (en) * | 2018-11-07 | 2020-05-14 | 杭州海康威视数字技术股份有限公司 | Image capturing method, monitoring camera, and monitoring system |
CN112307868A (en) * | 2019-07-31 | 2021-02-02 | 百度(美国)有限责任公司 | Image recognition method, electronic device, and computer-readable medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101373514A (en) * | 2007-08-24 | 2009-02-25 | 李树德 | Method and system for recognizing human face |
CN102567716A (en) * | 2011-12-19 | 2012-07-11 | 中山爱科数字科技股份有限公司 | Face synthetic system and implementation method |
CN102682276A (en) * | 2011-12-20 | 2012-09-19 | 河南科技大学 | Face recognition method and base image synthesis method under illumination change conditions |
CN104008370A (en) * | 2014-05-19 | 2014-08-27 | 清华大学 | Video face identifying method |
-
2016
- 2016-02-04 CN CN201610079687.1A patent/CN107038400A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101373514A (en) * | 2007-08-24 | 2009-02-25 | 李树德 | Method and system for recognizing human face |
CN102567716A (en) * | 2011-12-19 | 2012-07-11 | 中山爱科数字科技股份有限公司 | Face synthetic system and implementation method |
CN102682276A (en) * | 2011-12-20 | 2012-09-19 | 河南科技大学 | Face recognition method and base image synthesis method under illumination change conditions |
CN104008370A (en) * | 2014-05-19 | 2014-08-27 | 清华大学 | Video face identifying method |
Non-Patent Citations (2)
Title |
---|
胡德文 等: "平均脸", 《生物特征识别技术与方法》 * |
黄祥林 等: "图像检索中的高维索引", 《图像检索原理与实践》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109753857A (en) * | 2017-11-07 | 2019-05-14 | 北京虹图吉安科技有限公司 | A kind of three-dimensional face identification apparatus and system based on photometric stereo visual imaging |
CN109993039A (en) * | 2018-01-02 | 2019-07-09 | 上海银晨智能识别科技有限公司 | Portrait identification method and device, computer readable storage medium |
CN110472460A (en) * | 2018-05-11 | 2019-11-19 | 北京京东尚科信息技术有限公司 | Face image processing process and device |
WO2020094091A1 (en) * | 2018-11-07 | 2020-05-14 | 杭州海康威视数字技术股份有限公司 | Image capturing method, monitoring camera, and monitoring system |
CN111163259A (en) * | 2018-11-07 | 2020-05-15 | 杭州海康威视数字技术股份有限公司 | Image capturing method, monitoring camera and monitoring system |
CN109670451A (en) * | 2018-12-20 | 2019-04-23 | 天津天地伟业信息系统集成有限公司 | Automatic face recognition tracking |
CN112307868A (en) * | 2019-07-31 | 2021-02-02 | 百度(美国)有限责任公司 | Image recognition method, electronic device, and computer-readable medium |
CN112307868B (en) * | 2019-07-31 | 2024-04-19 | 百度(美国)有限责任公司 | Image recognition method, electronic device, and computer-readable medium |
CN110826435A (en) * | 2019-10-23 | 2020-02-21 | 上海能塔智能科技有限公司 | Method and device for identifying and verifying face in driving, vehicle-mounted equipment and storage medium |
CN111131700A (en) * | 2019-12-25 | 2020-05-08 | 重庆特斯联智慧科技股份有限公司 | Hidden tracking equipment for intelligent security and use method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107038400A (en) | Face identification device and method and utilize its target person tracks of device and method | |
CN104881637B (en) | Multimodal information system and its fusion method based on heat transfer agent and target tracking | |
CN104517102B (en) | Student classroom notice detection method and system | |
CN110490158B (en) | Robust face alignment method based on multistage model | |
CN107895160A (en) | Human face detection and tracing device and method | |
US9330312B2 (en) | Multispectral detection of personal attributes for video surveillance | |
CN105138954A (en) | Image automatic screening, query and identification system | |
CN109255289B (en) | Cross-aging face recognition method based on unified generation model | |
Kaâniche et al. | Recognizing gestures by learning local motion signatures of HOG descriptors | |
US20220180534A1 (en) | Pedestrian tracking method, computing device, pedestrian tracking system and storage medium | |
CN102938065A (en) | Facial feature extraction method and face recognition method based on large-scale image data | |
CN110008909A (en) | A kind of real-time audit system of system of real name business based on AI | |
CN114998934B (en) | Clothes-changing pedestrian re-identification and retrieval method based on multi-mode intelligent perception and fusion | |
CN108171223A (en) | A kind of face identification method and system based on multi-model multichannel | |
Wang et al. | Use fast R-CNN and cascade structure for face detection | |
CN110858276A (en) | Pedestrian re-identification method combining identification model and verification model | |
CN108647621A (en) | A kind of video analysis processing system and method based on recognition of face | |
CN111639580A (en) | Gait recognition method combining feature separation model and visual angle conversion model | |
CN112989889B (en) | Gait recognition method based on gesture guidance | |
Jiang et al. | Application of a fast RCNN based on upper and lower layers in face recognition | |
Nikitin et al. | Face anti-spoofing with joint spoofing medium detection and eye blinking analysis | |
CN110969101A (en) | Face detection and tracking method based on HOG and feature descriptor | |
Andiani et al. | Face recognition for work attendance using multitask convolutional neural network (MTCNN) and pre-trained facenet | |
CN110909678B (en) | Face recognition method and system based on width learning network feature extraction | |
CN109492513B (en) | Face space duplication eliminating method for light field monitoring |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170811 |