CN106682581A - Suitcase identifying method and device - Google Patents
Suitcase identifying method and device Download PDFInfo
- Publication number
- CN106682581A CN106682581A CN201611081654.7A CN201611081654A CN106682581A CN 106682581 A CN106682581 A CN 106682581A CN 201611081654 A CN201611081654 A CN 201611081654A CN 106682581 A CN106682581 A CN 106682581A
- Authority
- CN
- China
- Prior art keywords
- case
- bag
- region
- type
- feature vector
- 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
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
- G06V20/47—Detecting features for summarising video content
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a suitcase identifying method. The method includes the following steps: acquiring characteristic regions and the ontology characteristics of the characteristic regions of a to-be-detected image through a preset suitcase identifying model, the characteristic regions including types divided into a human body region and a suitcase region; based on a first layer of a preset condition random field probability model, conducting first time analysis on the ontology characteristics of the suitcase region and the position relation between the suitcase region and the human body region, and determining the suitcase type to which the suitcase region belongs based on the result of the first time analysis. Since the ontology characteristics of the suitcase region are jointly analyzed with the position relation between the suitcase region and the human body region in the first time analysis, factors are more comprehensively analyzed and the result of suitcase prediction is increased. Also, the invention discloses a suitcase identifying device.
Description
Technical field
The present invention relates to communication technical field, more particularly to a kind of case and bag know method for distinguishing, while the application is also especially related to
And a kind of equipment of case and bag identification.
Background technology
Pedestrian's identification is the important component part of intelligent transportation system and intelligent monitor system, and case and bag identification is then that pedestrian knows
An other sub- problem, can provide more abundant feature for pedestrian's identification, improve the accuracy of identification, be that pedestrian's identification is non-
Often important supplement.
The method of current case and bag identification is a lot, and main flow is based on the scheme of deep learning.It is most typical to have based on faster
The case and bag technology of identification of rcnn.But it is too many in view of the influence factor of case and bag identification, including common illumination, rotation, block, contract
The factor, the also different style of case and bag such as put, different carrying modes and different environmental backgrounds can all significantly affect case and bag
The accuracy of identification.Therefore, want to learn these variables completely using deep learning to need the sample of high number.However, inspection
Test sample is originally not easy to make, the sample to manufacture enormous amount, needs to expend substantial amounts of manpower and materials.
In finding prior art, at least there is problems with during the application is realized in applicant:
Case and bag identifying schemes of the prior art, are using identification general target (the applicable mesh of all types of objects
Mark) identification model identifying schemes, consider case and bag relative seat feature in the picture, particularly case and bag and pedestrian
Position feature so that case and bag identification accuracy rate it is not high.
It can be seen that, how on the basis of existing case and bag technology of identification, with reference to case and bag relative seat feature in the picture, come
Case and bag region in image is identified, and then improves the accuracy of case and bag identification, become those skilled in the art and urgently solve
Technical problem certainly.
The content of the invention
The invention provides a kind of case and bag recognition methods, on the basis of existing case and bag technology of identification, with reference to case and bag
Relative seat feature in the picture, to be identified to the case and bag region in image, and then improves the accuracy of case and bag identification.
The present invention proposes a kind of case and bag and knows method for distinguishing, it is characterised in that methods described at least includes:
The characteristic area of altimetric image to be checked, and the body of the characteristic area are obtained by default case and bag identification model
Feature, the type of the characteristic area includes human region and case and bag region;
According to the ground floor of default condition random field probabilistic model to the main body characteristic in the case and bag region and described
Position relationship between case and bag region and the human region is once analyzed, and is determined according to the result once analyzed
Case and bag type belonging to the case and bag region.
Preferably, methods described also includes:
There is multiple case and bag regions, root in the altimetric image to be checked if determining according to the result once analyzed
According to the second layer of the condition random field probabilistic model to the main body characteristic in the case and bag region, the case and bag region and the people
Position relationship between body region and the position relationship between the case and bag region and other case and bag regions carry out secondary point
Analysis, and the case and bag type according to belonging to the result of the secondary analysis determines the case and bag region.
Preferably, methods described also includes:
The condition random field probabilistic model is trained by difficult example sample, the difficult example sample is by described
What the sample of training pattern identification mistake was obtained after being corrected.
Preferably, the condition random field probabilistic model is as follows:
Wherein, p (y | x, θ) is the probability for specifying case and bag type for case and bag region, Z (x) for case and bag region normalization because
Son, exp { φ (y, x;θ) } belong to the potential-energy function of specified case and bag type for case and bag region, y is the specified case and bag type, and x is
The input feature vector in the case and bag region, θ is default parameter, and the input feature vector in the case and bag region includes the case and bag region
The position relationship feature of main body characteristic and the case and bag region and the human region.
Preferably, in the condition random field probabilistic model
Wherein, y is the specified case and bag type, and x is the input feature vector in the case and bag region, and θ is default parameter, θ
(yv) it is characterized parameter, λi,jFor the weights of transfer function,It is apart from penalty coefficient, θ (xi,xj,xhead,yi,yj) it is to turn
Function is moved, v is case and bag region v, and V is the set of all case and bag region compositions in the picture to be detected, and i is characterized region i, j
It is characterized region j, E is the set that all characteristic areas connect two-by-two composition in picture to be detected, xvFor the input of case and bag region v
Feature, xiIt is characterized the input feature vector of region i, xjIt is characterized the input feature vector of region j, xheadIt is characterized the input of region head
Feature yvFor the case and bag type of case and bag region v, yiIt is characterized the attribute type of region i, yjIt is characterized the attribute type of region j, d
It is characterized the distance between region i and characteristic area j.
Accordingly, the invention allows for a kind of equipment of case and bag identification, it is characterised in that the equipment at least includes:
Acquisition module is for obtaining the characteristic area of altimetric image to be checked by default case and bag identification model and described
The main body characteristic of characteristic area, the type of the characteristic area includes human region and case and bag region;
First analysis module, for according to the ground floor of default condition random field probabilistic model to the case and bag region
Position relationship between main body characteristic and the case and bag region and the human region is once analyzed, and according to described one
The result of secondary analysis determines the case and bag type belonging to the case and bag region.
Preferably, the equipment also includes:
Second analysis module, it is multiple for existing in determining the altimetric image to be checked in the result once analyzed according to
Main body characteristic, institute during the case and bag region, according to the second layer of the condition random field probabilistic model to the case and bag region
State the position relationship between case and bag region and the human region and the position between the case and bag region and other case and bag regions
The relation of putting carries out secondary analysis, and the case and bag type according to belonging to the result of the secondary analysis determines the case and bag region.
Preferably, the equipment also includes:
Training module, for being trained to the condition random field probabilistic model by difficult example sample, the difficult example sample
Originally it is by recognizing that wrong sample is obtained after correcting to the training pattern.
Preferably, the condition random field probabilistic model is as follows:
Wherein, p (y | x, θ) is the probability for specifying case and bag type for case and bag region, Z (x) for case and bag region normalization because
Son, exp { φ (y, x;θ) } belong to the potential-energy function of specified case and bag type for case and bag region, y is the specified case and bag type, and x is
The input feature vector in the case and bag region, θ is default parameter, and the input feature vector in the case and bag region includes the case and bag region
The position relationship feature of main body characteristic and the case and bag region and the human region.
Preferably, in the condition random field probabilistic model
Wherein, y is the specified case and bag type, and x is the input feature vector in the case and bag region, and θ is default parameter, θ
(yv) it is characterized parameter, λi,jFor the weights of transfer function,It is apart from penalty coefficient, θ (xi,xj,xhead,yi,yj) it is to turn
Function is moved, v is case and bag region v, and V is the set of all case and bag region compositions in the picture to be detected, and i is characterized region i, j
It is characterized region j, E is the set that all characteristic areas connect two-by-two composition in picture to be detected, xvFor the input of case and bag region v
Feature, xiIt is characterized the input feature vector of region i, xjIt is characterized the input feature vector of region j, xheadIt is characterized the input of region head
Feature yvFor the case and bag type of case and bag region v, yiIt is characterized the attribute type of region i, yjIt is characterized the attribute type of region j, d
It is characterized the distance between region i and characteristic area j.
By using technical scheme proposed by the present invention, by default case and bag identification model the spy of altimetric image to be checked being obtained
Region, and the main body characteristic of characteristic area are levied, the type of characteristic area includes human region and case and bag region;According to default
Condition random field probabilistic model ground floor between the main body characteristic and case and bag region and human region in case and bag region
Position relationship is once analyzed, and the case and bag type according to belonging to the result once analyzed determines case and bag region.Due to one
The main body characteristic and position relationship between case and bag region and human region in binding analysis case and bag regions in secondary analysis so that
The factor of analysis is more comprehensive, so as to improve the result of case and bag prediction.
Description of the drawings
In order to be illustrated more clearly that the technical scheme of the application, embodiment will be described below needed for the accompanying drawing to be used
It is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the present application, general for this area
For logical technical staff, on the premise of not paying creative work, can be with according to these other accompanying drawings of accompanying drawings acquisition.
Fig. 1 is a kind of schematic flow sheet of case and bag recognition methods that the embodiment of the present application is proposed;
Fig. 2 is a kind of schematic diagram of picture after normalized that the application specific embodiment is proposed;
Fig. 3 is the schematic diagram that samples pictures are carried out with region division that the application specific embodiment is proposed;
Fig. 4 is a kind of schematic diagram of flase drop picture that the application specific embodiment is proposed;
Fig. 5 is a kind of schematic diagram of the distance between characteristic area that the application specific embodiment the is proposed computational methods of d;
Fig. 6 is a kind of schematic flow sheet of case and bag recognition methods that the application specific embodiment is proposed;
Fig. 7 is to propose a kind of structural representation of case and bag identification equipment in the application specific embodiment.
Specific embodiment
As stated in the Background Art, existing case and bag technology of identification, typically using the identification model of general target come to picture
In case and bag region be identified, because the identification model of general target is generally by case and bag region to be measured in detection picture
Main body characteristic to the attribute type belonging to case and bag being predicted.But position relationship of the case and bag in picture, particularly with figure
The position relationship of human body also has an impact to the attribute type belonging to case and bag in piece, for example, the upper edge of draw-bar box and handbag
Divide and be typically connected with arm, the upper rim of knapsack is generally proximate to head and two shoulders.It can be seen that, can be by case and bag in detection figure
Position relationship in piece is further verified to the attribute type belonging to case and bag.
Therefore, the application proposes a kind of case and bag recognition methods, on the basis of existing case and bag technology of identification, with reference to case
Bag relative seat feature in the picture, to be identified to the case and bag region in image, and then improves the accurate of case and bag identification
Property.First, the characteristic area of altimetric image to be checked, and the main body characteristic of characteristic area are obtained by default training pattern.Its
The type of middle characteristic area includes human region and case and bag region.Then according to the of default condition random field probabilistic model
One layer of main body characteristic and the position relationship between case and bag region and human region to case and bag region is once analyzed, and root
The case and bag type belonging to case and bag region is determined according to analysis result.Due to the binding analysis case and bag regions in once analyzing
Position relationship between main body characteristic and case and bag region and human region so that the factor of analysis is more comprehensive, so as to carry
The high result of case and bag predictions.
A kind of schematic flow sheet of case and bag recognition methods of the application proposition is illustrated in figure 1, it should be noted that this Shen
Please be applied to be predicted the case and bag region in picture, before the step of performing the application, need to picture to be predicted
Carry out pedestrian's picture normalized.Specifically, the application at least comprises the following steps:
S101, by default case and bag identification model the characteristic area of altimetric image to be checked, and the characteristic area are obtained
Main body characteristic.Wherein, the type of characteristic area includes human region and case and bag region.
In embodiments herein, after treating detection image and carrying out pedestrian's picture normalization, first by default
Training pattern obtain characteristic area in altimetric image to be checked, the type of characteristic area includes human region and case and bag region.
Wherein, default training pattern can be faster rcnn case and bag identification models, or ssd case and bag identification models.
Two kinds of case and bag identification models by more than can obtain characteristic area (human region and case and bag region) in altimetric image to be checked,
Characteristic area is obtained in altimetric image to be checked.
It should be noted that being come to be detected using faster rcnn case and bag identification models, or ssd case and bag identification models
The main body characteristic that characteristic area obtains characteristic area in altimetric image to be checked is obtained in image, simply the application proposition is a kind of excellent
The scheme of choosing, based on the core concept of the application, those skilled in the art can also be using other characteristic areas and characteristic area
The acquisition methods of the main body characteristic in domain, this can't affect the protection domain of the application.
Human region and case and bag region mentioned above is the preliminary identification of default case and bag identification model in the application
As a result, in embodiments herein, in addition it is also necessary to which above-mentioned recognition result is further verified.
The main body characteristic of characteristic area refers to the mark of the characteristic area that default case and bag identification model is extracted in the application
Know feature, which represent the feature that characteristic area possesses in itself, and not including the position feature of characteristic area.If using
Faster rcnn are specifically as follows a multi-C vector as default case and bag identification model, the then identification characteristics of characteristic area
(dimension can be specially 4096 dimensions).
The main body characteristic of characteristic area represents the feature of characteristic area itself, used as the condition random field probability of the application
A part of input feature vector of model is used.Also, in the preferred embodiment of the application, need to identification characteristics (specially
Multi-C vector) carry out dimension-reduction treatment after, then the part input feature vector of the condition random field probabilistic model as the application.
By S101 the step of the above, can treat detection image carries out the extraction of characteristic area, and obtains characteristic area
The main body characteristic in domain.In embodiments herein, it is exactly based on and the case and bag region in characteristic area is further entered
Row soundness verification, so as to increased the accuracy to case and bag region recognition.Specifically, below to specific soundness verification mistake
Journey is explained in detail.
S102, the main body characteristic and case and bag according to the ground floor of default condition random field probabilistic model to case and bag region
Position relationship between region and human region is once analyzed, and according to belonging to an analysis result determines case and bag region
Case and bag type.
In embodiments herein, in the main body characteristic of the characteristic area and characteristic area for obtaining altimetric image to be checked
Afterwards, the main body characteristic and case and bag region and people according to the ground floor of default condition random field probabilistic model to case and bag region
Position relationship between body region is once analyzed, and the case and bag class according to belonging to an analysis result determines case and bag region
Type.
Case and bag type belonging to altimetric image raising middle flask bag region to be checked is gone back in addition to related to the main body characteristic in case and bag region
Location is relevant in picture with case and bag region, and particularly the position relationship between case and bag region and human region seems outstanding
For important.For example, draw-bar box is typically connected with the upper rim of handbag with arm, and the upper rim of knapsack is general
Near head and two shoulders.It can be seen that, the position relationship between case and bag region and human region can be provided the affiliated type of case and bag
Important reference.Therefore, in embodiments herein, using the position relationship between case and bag region and human region as condition
One important input feature vector of random field probabilistic model is used, so as to the prediction of the affiliated type in case and bag region more
Accurately.
In the preferred embodiment of the application, if determining there are multiple casees in altimetric image to be checked according to the result once analyzed
Bag region, then according to the second layer of condition random field probabilistic model to the main body characteristic in case and bag region, case and bag region and human body area
The position relationship between position relationship and case and bag region and other case and bag regions between domain carries out secondary analysis, and according to
The result of secondary analysis determines the case and bag type belonging to case and bag region.
If the result according to once analyzing determines has multiple case and bag regions in altimetric image to be checked, need further
Judge whether the position relationship between each case and bag region is reasonable, namely further consider influencing each other between case and bag region.
Therefore, in the preferred embodiment of the application, if determining there are multiple casees in altimetric image to be checked according to the result once analyzed
Bag region, then according to the second layer of condition random field probabilistic model to the main body characteristic in case and bag region, case and bag region and human body area
The position relationship between position relationship and case and bag region and other case and bag regions between domain carries out secondary analysis, and according to
The result of secondary analysis determines the case and bag type belonging to case and bag region.Because comprehensive considers above-mentioned three kind factor, so as to
So that the prediction to the affiliated type in case and bag region is more accurate.
In the preferred embodiment of the application, condition random field probabilistic model is trained by difficult example sample.Wherein,
Difficult example sample is by recognizing that wrong sample is obtained after correcting to default case and bag identification model.
Before use condition random field probabilistic model carries out case and bag identification, it is necessary first to condition random field probabilistic model
It is trained, in the preferred embodiment of the application, except being carried out to condition random field probabilistic model using common sample
Outside training, also condition random field probabilistic model is trained using difficult example sample.Wherein, difficult example sample is by pre-
If case and bag identification model identification mistake sample corrected after obtain.
Also, due to the limited amount of difficult example sample, therefore difficult example sample can be carried out by PCA dimensionality reductions reconfiguration technique
Expand, extend to original 3~5 times.
Also condition random field probabilistic model is trained using difficult example sample by more than, enables to condition random
Training parameter is treated in probabilistic model more accurately, so as to condition random field probabilistic model to the identification in case and bag region more
Accurately.
In the preferred embodiment of the application, condition random field probabilistic model is specific as follows:
Wherein, p (y | x, θ) is the probability for specifying case and bag type for case and bag region, Z (x) for case and bag region normalization because
Son, exp { φ (y, x;θ) } belong to the potential-energy function of specified case and bag type for case and bag region, to specify case and bag type, x is case and bag to y
The input feature vector in region, θ is default parameter (leads to after being trained to condition random field probabilistic model and obtain).
Wherein, main body characteristic of the input feature vector in case and bag region comprising case and bag region and case and bag region and human region
Position relationship feature.
Main body characteristic of the input feature vector in case and bag region comprising case and bag region in the preferred embodiment of the application, with reference to case
Bag region and the position relationship feature of human region and obtain.
When the main body characteristic in case and bag region is multi-C vector (specially 4096 dimension), it is necessary first to which main body characteristic is carried out
Dimension-reduction treatment, in conjunction with 4, case and bag region corner location relative to head intermediate point relative displacement, and the length in the region
Degree, width information reconfigures a new multi-C vector as the input feature vector in case and bag region.
In the preferred embodiment of the application, in the probability function in condition random field probabilistic model
Wherein, y is to specify case and bag types, x for case and bag region input feature vector, θ is default parameter, θ (yv) it is state
Parameter, λi,jFor the weights of transfer function,It is apart from penalty coefficient, θ (xi,xj,xhead,yi,yj) be transfer function, v
For case and bag region v, V is the set of all case and bag region compositions in picture to be detected, and i is characterized region i, and j is characterized region j, E
For the set that all characteristic areas in picture to be detected connect two-by-two composition, xvFor the input feature vector of case and bag region v, xiIt is characterized
The input feature vector of region i, xjIt is characterized the input feature vector of region j, xheadIt is characterized the input feature vector of region head, yvFor case and bag
The case and bag type of region v, yiIt is characterized the attribute type of region i, yjBe characterized the attribute type of region j, d be characterized region i with
The distance between characteristic area j.
It should be noted that state parameter θ (yv) correspond to input feature vector xv, its dimension and xvDimension it is identical, shift letter
Number θ (xi,xj,xhead,yi,yj) correspond to xi、xjAnd xheadDeng three input feature vectors, its dimension is the 3 of input feature vector dimension
Times.
λi,jFor the weights of transfer function, only and yi, yjThere is relation, the size of its value can be by by condition random field
Probabilistic model is obtained after being trained.
In above formula,It is that, apart from penalty coefficient, two regions are separated by more remote, and its weight should be lower, wherein d's
Computational methods are as follows:
By observing a large amount of case and bag samples, the upper rim feature for finding case and bag region becomes apparent from and stablizes, for example
Draw-bar box is typically connected with the upper rim of handbag with arm, the upper rim of knapsack generally proximate to head and two shoulders,
But due to not knowing for case and bag size and location, lower edge portion position feature is just less obvious.Therefore, send to be elected in the application
The center position for taking case and bag region top edge is d (d for the minimum distance of reference point, other regions and this central point>=
0), partly in particular cases, the point is included by other regions, and now d should be negative value, and here we take d=0.
By the description of above example, the characteristic area of altimetric image to be checked is obtained by default training pattern,
And the main body characteristic of characteristic area.Wherein the type of characteristic area includes human region and case and bag region.Then according to pre-
If condition random field probabilistic model ground floor between the main body characteristic and case and bag region and human region in case and bag region
Position relationship once analyzed, and the case and bag type according to belonging to an analysis result determines case and bag region.Due to one
The main body characteristic and position relationship between case and bag region and human region in binding analysis case and bag regions in secondary analysis so that
The factor of analysis is more comprehensive, so as to improve the result of case and bag prediction.
In order to the technological thought of the present invention is expanded on further, in conjunction with specific implementing procedure, the technical side to the present invention
Case is illustrated.
Case and bag recognition methods described herein is made up of following steps, and entrance is the police of intelligent transportation electricity or garden etc.
The picture of collection, the region for then obtaining pedestrian place by pedestrian detection (suitably extends up and down a part, it is ensured that case and bag
Major part is included), by rectangle frame intercepting effect as shown in Fig. 2 the part is compared with mature technology, here is not superfluous
State, it is to be appreciated that no matter train or detect, the region that entrance picture is intercepted for pedestrian.Preferably, the figure for detecting and training
Piece is normalized to 400*200 sizes.
To the region segmentation of pedestrian in training process, one is divided into head, upper arm, underarm, leg and foot, except head
Outside, other parts are divided into left and right, are illustrated in fig. 3 shown below (pedestrian sample is obtained and is relatively easy to, and training difficulty is also below case and bag).
For case and bag classification, fall into 5 types altogether, respectively both shoulders bag, shoulder bag, handbag, cap and draw-bar box, wherein
Draw-bar box is confined as shown above, and other classifications are confined in the manner described above, after all area markings are completed, according to
Specific format generate training sample, input faster rcnn be trained (classification altogether be 5 classes of case and bag+human body attribute
Class=15 class of 9 classes+background 1).
By faster rcnn to training picture to carry out feature acquisition, all candidate frames pass through in training picture
Roipooling layers are normalized to the region of formed objects, then form 4096 dimensional vectors by full articulamentum and are used as through dimensionality reduction
A part of input feature vector of condition subsequent random field.
The input training sample of condition random field has two sources, and one is above-mentioned faster rcnn identical training samples;
Two is to confirm wrong sample in faster rcnn detection process, after this part sample is corrected, can as condition with
The difficult example sample input on airport, bearing calibration citing:As shown in figure 4, below this pictures having one more examine (backpack) and
The region of many inspections is just directly designated background frame by one flase drop (package), trimming process, is right by the area identification of flase drop
The classification (bag) answered, it is noted that after corrected, figure below is changed into two bag regional frames, and this is recognized by condition random field
Reasonably to there is relation, mainly for the higher situation of the false drop rate for being adapted to faster rcnn.For condition random field detection
The process of repeat region is exported afterwards, as long as these repeat regions belong to same category attribute, can be non-greatly by nms
Value Restrainable algorithms are eliminated.
Explanation is additionally needed, in view of detection sample is generally not a lot, this part hardly possible example sample is generally required through PCA drops
Dimension reconfiguration technique is expanded, and original 3~5 times is extended to, depending on actual conditions.
The condition random field probabilistic model that the application is adopted is as follows:This mould
Type can calculate some node its attribute probability, and (whether its attribute of case and bag region for arbitrarily detecting is reliable, in above formula
Y sequence only one of which it is unknown), also can be calculated the combined chance of several nodes that (i.e. multiple case and bag regions are each other
Position relationship whether rationally, the y sequences in above formula have multiple unknown).
Wherein, p (y | x, θ) is the probability for specifying case and bag type for case and bag region, Z (x) for case and bag region normalization because
Son, exp { φ (y, x;θ) } belong to the potential-energy function of specified case and bag type for case and bag region, to specify case and bag type, x is case and bag to y
The input feature vector in region, θ is default parameter (leads to after being trained to condition random field probabilistic model and obtain).Wherein x features
4096 dimensional feature vectors comprising full articulamentum output in aforementioned faster rcnn, 500 through PCA technology dimensionality reductions are tieed up
Vector, with reference to 4, region corner location relative to head intermediate point relative displacement, and the length in the region, width letter
Cease, totally 506 dimensional vectors composition.
In above-mentioned probabilistic model
Wherein, y is to specify case and bag types, x for case and bag region input feature vector, θ is default parameter, θ (yv) it is state
Parameter, λi,jFor the weights of transfer function,It is apart from penalty coefficient, θ (xi,xj,xhead,yi,yj) be transfer function, v
For case and bag region v, V is the set of all case and bag region compositions in picture to be detected, and i is characterized region i, and j is characterized region j, E
For the set that all characteristic areas in picture to be detected connect two-by-two composition, xvFor the input feature vector of case and bag region v, xiIt is characterized
The input feature vector of region i, xjIt is characterized the input feature vector of region j, xheadIt is characterized the input feature vector of region head, yvFor case and bag
The case and bag type of region v, yiIt is characterized the attribute type of region i, yjBe characterized the attribute type of region j, d be characterized region i with
The distance between characteristic area j.
It should be noted that for ground floor condition random field, inevitable one of i and j is case and bag region, and another is human body
Attribute region, for second layer condition random field, it is only necessary to ensure that one of them is case and bag region, then traversal is possible to group
Close.
λ mentioned abovei,jIt is the weights of transfer function, only and yi,yjIt is relevant,It is two apart from penalty coefficient
Region is separated by more remote, and its weight should be lower, and the wherein computational methods of d are as follows.
The computational methods of d:By observing a large amount of case and bag samples, the upper rim feature for finding case and bag region becomes apparent from
With it is stable, such as draw-bar box is typically connected with the upper rim of handbag with arm, the upper rim of knapsack generally proximate to
Head and two shoulders, but not knowing due to case and bag size and location, lower edge portion position feature is just less obvious.Therefore,
As shown in figure 5, the application send to be elected take case and bag region top edge center position be reference point, other regions and this central point
Minimum distance be d (d>=0).In particular cases, the point is included by other regions, and now d should be negative value for part, this
In we take d=0.
In above-mentioned probabilistic modelWherein y ' represents case and bag region and is likely to belong to
Case and bag type.
In this application, case and bag region is predicted using the condition random field of two-layer, takes two-layer condition random field
The reason for be faster rcnn to the prediction reliability of human body attribute apparently higher than the prediction to case and bag attribute, if at the beginning
Just human body attribute and all case and bag attributes are put together and be predicted, shadow can be predicted due to the case and bag of faster rcnn mistakes
The judgement of condition random field is rung, so the prediction of the condition random field of ground floor is based on all human body attributes for detecting
Region and single case and bag attribute region, confirm respectively the confidence level in each case and bag region, and flase drop therein and many inspections are carried out
Correct.The condition random field of the second layer is then that all case and bag regions are carried out with comprehensive descision, that is, consider the phase between case and bag region
Mutually affect, especially case and bag overlapping cases, can effectively be corrected by the condition random field of the second layer.If a width picture is detected
Out only one of which case and bag attribute region, then need not carry out the judgement of the second layer.
As shown in fig. 6, the schematic flow sheet of the case and bag identification for the application specific embodiment, as seen from the figure including following
Step:
S601, receives detection picture.
S602, to detecting that picture carries out pedestrian's picture normalized.
S603, obtains testing result and roi features.
The characteristic area of detection image, and the main body characteristic of characteristic area are obtained by Faster rcnn training patterns.
S604, feature PCA dimensionality reduction.
The dimension-reduction treatment of feature PCA is carried out to the main body characteristic of characteristic area.
S605, condition random field ground floor differentiates.
S606, if only one of which case and bag node, if so, then performs S607;If it is not, then performing S608.
S607, output result.
S608, the condition random field second layer differentiates.
S609, output result.
Flow process description by more than, by default case and bag identification model the characteristic area of altimetric image to be checked is obtained
Domain, and the main body characteristic of characteristic area, the type of characteristic area includes human region and case and bag region;According to default bar
Main body characteristic and position case and bag region and human region between of the ground floor of part random field probabilistic model to case and bag region
Relation is once analyzed, and the case and bag type according to belonging to the result once analyzed determines case and bag region.Due to once dividing
The main body characteristic and position relationship between case and bag region and human region in binding analysis case and bag regions in analysis so that analysis
Factor more comprehensively, so as to improve the result of case and bag prediction.
Accordingly, the invention allows for a kind of case and bag identification equipment, as shown in fig. 7, be the application specific embodiment
It is middle to propose a kind of structural representation of case and bag identification equipment.The equipment at least includes:
Acquisition module 701, for obtaining the characteristic area of altimetric image to be checked, Yi Jisuo by default case and bag identification model
The main body characteristic of characteristic area is stated, the type of the characteristic area includes human region and case and bag region;
First analysis module 702, for according to the ground floor of default condition random field probabilistic model to the case and bag area
Position relationship between the main body characteristic in domain and the case and bag region and the human region is once analyzed, and according to institute
State the result once analyzed and determine case and bag type belonging to the case and bag region.
Preferably, the equipment also includes:
Second analysis module, it is multiple for existing in determining the altimetric image to be checked in the result once analyzed according to
Main body characteristic, institute during the case and bag region, according to the second layer of the condition random field probabilistic model to the case and bag region
State the position relationship between case and bag region and the human region and the position between the case and bag region and other case and bag regions
The relation of putting carries out secondary analysis, and the case and bag type according to belonging to the result of the secondary analysis determines the case and bag region.
Preferably, the equipment also includes:
Training module, for being trained to the condition random field probabilistic model by difficult example sample, the difficult example sample
Originally it is by recognizing that wrong sample is obtained after correcting to the training pattern.
Preferably, the condition random field probabilistic model is as follows:
Wherein, p (y | x, θ) is the probability for specifying case and bag type for case and bag region, Z (x) for case and bag region normalization because
Son, exp { φ (y, x;θ) } belong to the potential-energy function of specified case and bag type, exp { φ (y, x for case and bag region;θ) } it is the case
Bag region belongs to the probability function of the specified case and bag type, and y is the specified case and bag type, and x is the defeated of the case and bag region
Enter feature, θ is default parameter, the main body characteristic of the input feature vector comprising the case and bag region in the case and bag region and described
The position relationship feature of case and bag region and the human region.
Preferably, in the condition random field probabilistic model
Wherein, y is the specified case and bag type, and x is the input feature vector in the case and bag region, and θ is default parameter, θ
(yv) it is characterized parameter, λi,jFor the weights of transfer function,It is apart from penalty coefficient, θ (xi,xj,xhead,yi,yj) be
Transfer function, v is case and bag region v, and V is the set of all case and bag region compositions in the picture to be detected, and i is characterized region i,
J is characterized region j, and E is the set that all characteristic areas connect two-by-two composition in picture to be detected, xvFor the defeated of case and bag region v
Enter feature, xiIt is characterized the input feature vector of region i, xjIt is characterized the input feature vector of region j, xheadIt is characterized the defeated of region head
Enter feature yvFor the case and bag type of case and bag region v, yiIt is characterized the attribute type of region i, yjThe attribute type of region j is characterized,
D is characterized the distance between region i and characteristic area j.
The description of concrete equipment, by default case and bag identification model the feature of altimetric image to be checked is obtained by more than
Region, and the main body characteristic of characteristic area, the type of characteristic area includes human region and case and bag region;According to default
Main body characteristic and position case and bag region and human region between of the ground floor of condition random field probabilistic model to case and bag region
Put relation once to be analyzed, and the case and bag type according to belonging to the result once analyzed determines case and bag region.Due to once
The main body characteristic and position relationship between case and bag region and human region in binding analysis case and bag regions in analysis so that point
The factor of analysis is more comprehensive, so as to improve the result of case and bag prediction.
What is finally illustrated is:Various embodiments above only to illustrate technical scheme, rather than a limitation;Although
The present invention has been described in detail with reference to foregoing embodiments, it will be understood by those within the art that;It is still
Technical scheme described in foregoing embodiments can be modified, either which part or all technical characteristic are carried out
Equivalent;And these are changed or are replaced, the essence disengaging the claims in the present invention for not making appropriate technical solution are limited
Scope.
Through the above description of the embodiments, those skilled in the art can be understood that the present invention can lead to
Cross hardware realization, it is also possible to realize by the mode of software plus necessary general hardware platform.Based on such understanding, this
Bright technical scheme can be embodied in the form of software product, and the software product can be stored in a non-volatile memories
In medium (can be CD-ROM, USB flash disk, portable hard drive etc.), including some instructions are used so that a computer equipment (can be
Personal computer, server, or network equipment etc.) perform method described in each implement scene of the invention.
It will be appreciated by those skilled in the art that accompanying drawing is a schematic diagram for being preferable to carry out scene, module in accompanying drawing or
Flow process is not necessarily implemented necessary to the present invention.
It will be appreciated by those skilled in the art that the module in the device in implement scene can according to implement scene describe into
Row is distributed in the device of implement scene, it is also possible to carry out one or more dresses that respective change is disposed other than this implement scene
In putting.The module of above-mentioned implement scene can merge into a module, it is also possible to be further split into multiple submodule.
The invention described above sequence number is for illustration only, does not represent the quality of implement scene.
Disclosed above is only that the several of the present invention are embodied as scene, but, the present invention is not limited to this, Ren Heben
What the technical staff in field can think change should all fall into protection scope of the present invention.
Claims (10)
1. a kind of case and bag know method for distinguishing, it is characterised in that methods described at least includes:
The characteristic area of altimetric image to be checked, and the body spy of the characteristic area are obtained by default case and bag identification model
Levy, the type of the characteristic area includes human region and case and bag region;
Main body characteristic and the case and bag according to the ground floor of default condition random field probabilistic model to the case and bag region
Position relationship between region and the human region is once analyzed, and according to the result once analyzed determines
Case and bag type belonging to case and bag region.
2. the method for claim 1, it is characterised in that methods described also includes:
There are multiple case and bag regions in the altimetric image to be checked if determining according to the result once analyzed, according to institute
State main body characteristic of the second layer to the case and bag region, the case and bag region and the human body area of condition random field probabilistic model
Position relationship between domain and the position relationship between the case and bag region and other case and bag regions carry out secondary analysis, and
Case and bag type according to belonging to the result of the secondary analysis determines the case and bag region.
3. the method as described in right wants 2, it is characterised in that methods described also includes:
The condition random field probabilistic model is trained by difficult example sample, the difficult example sample is by the training
What the sample of Model Identification mistake was obtained after being corrected.
4. the method as described in any one of claim 1-3, it is characterised in that the condition random field probabilistic model is as follows:
Wherein, p (y | x, θ) is the probability for specifying case and bag type for case and bag region, Z (x) for case and bag region normalization factor,
exp{φ(y,x;θ) } belong to the potential-energy function of specified case and bag type for case and bag region, y is the specified case and bag type, and x is institute
The input feature vector in case and bag region is stated, θ is default parameter, and the input feature vector in the case and bag region includes the sheet in the case and bag region
The position relationship feature of body characteristicses and the case and bag region and the human region.
5. method as claimed in claim 4, it is characterised in that in the condition random field probabilistic model
Wherein, y is the specified case and bag type, and x is the input feature vector in the case and bag region, and θ is default parameter, θ (yv) it is spy
Levy parameter, λi,jFor the weights of transfer function,It is apart from penalty coefficient, θ (xi,xj,xhead,yi,yj) it is transfer function,
V is case and bag region v, and V is the set of all case and bag region compositions in the picture to be detected, and i is characterized region i, and j is characterized area
Domain j, E are the set that all characteristic areas connect two-by-two composition in picture to be detected, xvFor the input feature vector of case and bag region v, xiFor
The input feature vector of characteristic area i, xjIt is characterized the input feature vector of region j, xheadIt is characterized the input feature vector y of region headvFor case
The case and bag type of bag region v, yiIt is characterized the attribute type of region i, yjThe attribute type of region j is characterized, d is characterized region i
The distance between with characteristic area j.
6. the equipment that a kind of case and bag are recognized, it is characterised in that the equipment at least includes:
Acquisition module, for obtaining the characteristic area of altimetric image to be checked, and the feature by default case and bag identification model
The main body characteristic in region, the type of the characteristic area includes human region and case and bag region;
First analysis module, for the body according to the ground floor of default condition random field probabilistic model to the case and bag region
Position relationship between feature and the case and bag region and the human region is once analyzed, and is once divided according to described
The result of analysis determines the case and bag type belonging to the case and bag region.
7. equipment as claimed in claim 6, it is characterised in that the equipment also includes:
Second analysis module, for according to the result once analyzed determine exist in the altimetric image to be checked it is multiple described in
Main body characteristic, case during case and bag region, according to the second layer of the condition random field probabilistic model to the case and bag region
Position relationship between bag region and the human region and the position between the case and bag region and other case and bag regions are closed
System carries out secondary analysis, and the case and bag type according to belonging to the result of the secondary analysis determines the case and bag region.
8. equipment as claimed in claim 7, it is characterised in that the equipment also includes:
Training module, for being trained to the condition random field probabilistic model by difficult example sample, the difficult example sample is
Obtain after by recognizing that the sample of mistake is corrected to the training pattern.
9. the equipment as described in any one of claim 6-8, it is characterised in that the condition random field probabilistic model is as follows:
Wherein, p (y | x, θ) is the probability for specifying case and bag type for case and bag region, Z (x) for case and bag region normalization factor,
exp{φ(y,x;θ) } belong to the potential-energy function of specified case and bag type for case and bag region, y is the specified case and bag type, and x is institute
The input feature vector in case and bag region is stated, θ is default parameter, and the input feature vector in the case and bag region includes the sheet in the case and bag region
The position relationship feature of body characteristicses and the case and bag region and the human region.
10. equipment as claimed in claim 9, it is characterised in that in the condition random field probabilistic model
Wherein, y is the specified case and bag type, and x is the input feature vector in the case and bag region, and θ is default parameter, θ (yv) it is spy
Levy parameter, λi,jFor the weights of transfer function,It is apart from penalty coefficient, θ (xi,xj,xhead,yi,yj) it is transfer function,
V is case and bag region v, and V is the set of all case and bag region compositions in the picture to be detected, and i is characterized region i, and j is characterized area
Domain j, E are the set that all characteristic areas connect two-by-two composition in picture to be detected, xvFor the input feature vector of case and bag region v, xiFor
The input feature vector of characteristic area i, xjIt is characterized the input feature vector of region j, xheadIt is characterized the input feature vector y of region headvFor case
The case and bag type of bag region v, yiIt is characterized the attribute type of region i, yjThe attribute type of region j is characterized, d is characterized region i
The distance between with characteristic area j.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611081654.7A CN106682581B (en) | 2016-11-30 | 2016-11-30 | Luggage identification method and equipment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611081654.7A CN106682581B (en) | 2016-11-30 | 2016-11-30 | Luggage identification method and equipment |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106682581A true CN106682581A (en) | 2017-05-17 |
CN106682581B CN106682581B (en) | 2021-04-20 |
Family
ID=58866196
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611081654.7A Active CN106682581B (en) | 2016-11-30 | 2016-11-30 | Luggage identification method and equipment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106682581B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108959197A (en) * | 2018-09-10 | 2018-12-07 | 杭州后博科技有限公司 | A kind of stolen probability calculation system and method for luggage |
CN113378752A (en) * | 2021-06-23 | 2021-09-10 | 济南博观智能科技有限公司 | Pedestrian backpack detection method and device, electronic equipment and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP2709042A1 (en) * | 2012-09-17 | 2014-03-19 | Sihl GmbH | Multilayer RFID loop tag comprising cold seal adhesive layers |
CN104535594A (en) * | 2015-01-08 | 2015-04-22 | 安徽中福光电科技有限公司 | Luggage security check method based on X-rays, DR and CBS and intelligent X-ray machine |
CN105469087A (en) * | 2015-07-13 | 2016-04-06 | 百度在线网络技术(北京)有限公司 | Method for identifying clothes image, and labeling method and device of clothes image |
-
2016
- 2016-11-30 CN CN201611081654.7A patent/CN106682581B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP2709042A1 (en) * | 2012-09-17 | 2014-03-19 | Sihl GmbH | Multilayer RFID loop tag comprising cold seal adhesive layers |
CN104535594A (en) * | 2015-01-08 | 2015-04-22 | 安徽中福光电科技有限公司 | Luggage security check method based on X-rays, DR and CBS and intelligent X-ray machine |
CN105469087A (en) * | 2015-07-13 | 2016-04-06 | 百度在线网络技术(北京)有限公司 | Method for identifying clothes image, and labeling method and device of clothes image |
Non-Patent Citations (1)
Title |
---|
周海: ""基于条件随机场和空间推理的地理编码方法"", 《中国优秀硕士学位论文全文数据库 基础科学辑》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108959197A (en) * | 2018-09-10 | 2018-12-07 | 杭州后博科技有限公司 | A kind of stolen probability calculation system and method for luggage |
CN113378752A (en) * | 2021-06-23 | 2021-09-10 | 济南博观智能科技有限公司 | Pedestrian backpack detection method and device, electronic equipment and storage medium |
CN113378752B (en) * | 2021-06-23 | 2022-09-06 | 济南博观智能科技有限公司 | Pedestrian backpack detection method and device, electronic equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN106682581B (en) | 2021-04-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Eisenbach et al. | How to get pavement distress detection ready for deep learning? A systematic approach | |
WO2019140767A1 (en) | Recognition system for security check and control method thereof | |
Turker et al. | Field-based sub-boundary extraction from remote sensing imagery using perceptual grouping | |
CN108830188A (en) | Vehicle checking method based on deep learning | |
CN109766936B (en) | Image change detection method based on information transfer and attention mechanism | |
CN108334849A (en) | A kind of recognition methods again of the pedestrian based on Riemann manifold | |
CN109285139A (en) | A kind of x-ray imaging weld inspection method based on deep learning | |
CN109446925A (en) | A kind of electric device maintenance algorithm based on convolutional neural networks | |
CN110018524A (en) | A kind of X-ray safety check contraband recognition methods of view-based access control model-attribute | |
CN108830285A (en) | A kind of object detection method of the reinforcement study based on Faster-RCNN | |
CN111611874B (en) | Face mask wearing detection method based on ResNet and Canny | |
CN106529499A (en) | Fourier descriptor and gait energy image fusion feature-based gait identification method | |
CN106022317A (en) | Face identification method and apparatus | |
CN108710913A (en) | A kind of switchgear presentation switch state automatic identification method based on deep learning | |
CN105574550A (en) | Vehicle identification method and device | |
CN106096561A (en) | Infrared pedestrian detection method based on image block degree of depth learning characteristic | |
CN104598908A (en) | Method for recognizing diseases of crop leaves | |
CN107358182A (en) | Pedestrian detection method and terminal device | |
CN105260738A (en) | Method and system for detecting change of high-resolution remote sensing image based on active learning | |
CN107833213A (en) | A kind of Weakly supervised object detecting method based on pseudo- true value adaptive method | |
CN105427309A (en) | Multiscale hierarchical processing method for extracting object-oriented high-spatial resolution remote sensing information | |
CN105160317A (en) | Pedestrian gender identification method based on regional blocks | |
CN112560675B (en) | Bird visual target detection method combining YOLO and rotation-fusion strategy | |
CN107506793A (en) | Clothes recognition methods and system based on weak mark image | |
CN106023257A (en) | Target tracking method based on rotor UAV platform |
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 | ||
TR01 | Transfer of patent right | ||
TR01 | Transfer of patent right |
Effective date of registration: 20221107 Address after: 15 / F, Zhongchuang Plaza, 385 Hangtian Middle Road, national civil aerospace industrial base, Xi'an City, Shaanxi Province 710100 Patentee after: Xi'an Yu vision Mdt InfoTech Ltd. Address before: 2 / F, South Block, building 10, wanlun Science Park, 88 Jiangling Road, Binjiang District, Hangzhou City, Zhejiang Province, 310051 Patentee before: ZHEJIANG UNIVIEW TECHNOLOGIES Co.,Ltd. |