Specific embodiment
To keep the purposes, technical schemes and advantages of the application clearer, below in conjunction with attached drawing to the application embodiment party
Formula is described in further detail.
Referring to FIG. 1, the schematic diagram of the implementation environment provided it illustrates the application one embodiment.The implementation environment can
To include: unmanned counter 10 and computer equipment 20.
It include cabinet body 101, camera 102 and gravity pallet 103 in unmanned counter 10.Camera 102 is for acquiring nothing
The image of article in people's counter 10, gravity pallet 103 are used for overall A/W changing value corresponding to unmanned counter 10 and carry out
Measurement.Optionally, unmanned counter 10 is 5 layers of counter, is respectively installed on the cabinet body 101 of the left and right sides of first layer and third layer counter
There is a camera 102, is fitted with a gravity pallet 103 below each layer of counter.Unmanned counter 10 supports 15SKU
(Stock Keeping Unit, keeper unit), i.e., the operation of 15 kinds goods categories.
Computer equipment 20 refers to the electronic equipment for having calculating and processing capacity, for example, PC (Personal
Computer, personal computer), server etc..Server can be a server, be also possible to be made of multiple servers
Server cluster or a cloud computing service center.
After camera in unmanned counter 10 collects the image of article, computer equipment 20 is sent an image to, is calculated
Machine equipment 20 identifies the image of article, obtains visual identity result.
Computer equipment 20 can integrate in unmanned counter 10, can also exist independently of unmanned counter 10.Unmanned goods
It can be communicated by network between cabinet 10 and computer equipment 20, which can be cable network, be also possible to wireless
Network.
For ease of description, it in following methods embodiment, is only carried out by computer equipment of the executing subject of each step
Illustrate, but does not constitute restriction to this.
Referring to FIG. 2, it illustrates the item identification methods applied to unmanned counter that the application one embodiment provides
Flow chart.This method can be executed by computer equipment.This method may include following steps:
Step 201, by Visual identification technology, identify that the n times article that user executes in unmanned counter picks and places behavior pair
The pick-and-place article summation answered obtains visual identity as a result, n is positive integer.
Visual identification technology refers to by the way that after camera acquisition identification image, collected identification image is sent to identification
Model, the technology that identification image is identified by identification model.In the embodiment of the present application, Visual identification technology is for identification
The article that user executes in unmanned counter picks and places the corresponding pick-and-place article summation of behavior.Visual identity by vision the result is that known
The pick-and-place article result that other technology identifies.
User selects nobody lift-on-lift-off article, for example, user taken from unmanned counter one bottle can happy one
After bottle Sprite, and Sprite is put back in unmanned counter, has taken one bottle of fruit juice, then user performs 3 objects in unmanned counter
Product pick and place behavior, and 3 articles pick and place the corresponding pick-and-place article summation of behavior can happy one bottle of fruit juice for one bottle.
Step 202, the corresponding article Gravity changer value of detection visual identity result is corresponding with unmanned counter overall practical
Whether Gravity changer value matches.
The corresponding article Gravity changer value of visual identity result refers to all items of the corresponding pick-and-place of visual identity result
Mark gravity value summation.
The corresponding overall A/W changing value of unmanned counter refers to the gravity value and use of nobody lift-on-lift-off original article
Family picks and places the difference of the gravity value of nobody the lift-on-lift-off remaining articles after article.Optionally, the corresponding totality of unmanned counter
A/W changing value can be obtained according to the corresponding measured value of nobody lift-on-lift-off gravity pallet.
In one example, when the corresponding article Gravity changer value of visual identity result is corresponding with unmanned counter overall real
When border Gravity changer value is consistent, the corresponding article Gravity changer value of visual identity result overall practical weight corresponding with unmanned counter
The matching of power changing value.For example, visual identity the result is that user has taken one bottle of cola, the gravity value of one bottle of laughable mark is 5, then
The corresponding article Gravity changer value of visual identity is 5.The gravity value of nobody lift-on-lift-off original article is 20, and user picks and places article
The gravity value of nobody lift-on-lift-off remaining articles later is 15, then the corresponding overall A/W changing value of unmanned counter is 5.
The corresponding article Gravity changer value of visual identity result overall A/W changing value corresponding with unmanned counter is consistent, then vision
The overall A/W changing value matching corresponding with unmanned counter of the corresponding article Gravity changer value of recognition result.
In another example, when the corresponding article Gravity changer value of visual identity result totality corresponding with unmanned counter
When A/W changing value differs default value, the corresponding article Gravity changer value of visual identity result is corresponding with unmanned counter
Overall A/W changing value matching, the default value are the gravity error values itself allowed when article factory.For example, vision is known
Not one bottle of cola the result is that user has taken, the gravity value marked on cola is 5, then the corresponding article Gravity changer of visual identity
Value is 5.Nobody gravity value of lift-on-lift-off original article is 95, and user picks and places nobody lift-on-lift-off remaining articles after article
Gravity value be 91, then the corresponding overall A/W changing value of unmanned counter is 4.But visual identity result it is corresponding one bottle can
The actual gravity value of pleasure is 4,1 is differed with the gravity value of one bottle of cola mark, but value of this difference belongs to when cola dispatches from the factory certainly
The gravity error value that body allows, so the corresponding article Gravity changer value of visual identity result is corresponding with unmanned counter overall real
The matching of border Gravity changer value.
Step 203, if the corresponding article Gravity changer value of visual identity result and overall A/W changing value mismatch,
It then generates n times article and picks and places the corresponding a variety of prediction results for picking and placing article summation of behavior.
According to the gravity value for each article that visual identity result, overall A/W changing value and unmanned counter provide,
Generate a variety of prediction results.The corresponding overall gravity value of a variety of prediction results of above-mentioned generation and overall A/W changing value phase
Matching.For example, having cola, Sprite and fruit juice, laughable gravity value in unmanned counter is 5, the gravity value of Sprite is 4, fruit juice
Gravity value is 3, and visual identity cola the result is that user has taken, overall A/W changing value is 9, then a kind of prediction generated
The result is that user taken one bottle can happy one bottle of Sprite.
Step 204, the corresponding confidence level of a variety of prediction results is calculated.
Confidence level can be referred to as confidence level or reliability again, indicate that population parameter value falls in a certain area of sample statistics value
Interior probability.In the embodiment of the present application, the corresponding confidence level of a variety of prediction results refers to that the article of user's pick-and-place is
The probability of a certain prediction result.A possibility that confidence level is bigger, and the article for showing that user picks and places is the prediction result is bigger.
In the embodiment of the present application, according to the corresponding confidence level of each of prediction result article, it is pre- to calculate this
Survey the corresponding confidence level of result.The corresponding confidence level of each article be according in gravity value, visual identity feature extremely
It is few a kind of to obtain.Optionally, the corresponding confidence level of each of prediction result article is averaging, obtains prediction result pair
The confidence level answered.
Step 205, by the maximum target prediction of confidence level in a variety of prediction results as a result, being determined as n times article picks and places row
For corresponding pick-and-place article summation.
Optionally, it is ranked up the corresponding confidence level of a variety of prediction results is descending, chooses the in sequence
The corresponding prediction result of one confidence level is as target prediction result.
The summation that behavior is picked and placed using the maximum prediction result of confidence level as n times article ensure that identification user picked and placed
The accuracy rate of goods categories.
In conclusion in technical solution provided by the embodiments of the present application, by working as the corresponding article weight of visual identity result
When power changing value overall A/W changing value corresponding with unmanned counter mismatches, a variety of prediction results are generated, and calculate more
The kind corresponding confidence level of prediction result, chooses the maximum prediction result of confidence level in a variety of prediction results, as final
Article summation is picked and placed, ensure that the accuracy rate in the goods categories that identification user takes.
It is found through experiment that unmanned counter after application the present processes, has been taken a kind of merchandise classification in identification user
Accuracy rate is 100%, is 95% or so in the accuracy rate of identification user while 2 kinds of merchandise classifications of having taken.
In addition, Visual identification technology provided by the embodiments of the present application not only can recognize that the article that user takes, Er Qieneng
Identify that the article that user puts back to, the i.e. article of user pick and place behavior and can be identified.
Referring to FIG. 3, it illustrates the article identification sides applied to unmanned counter that another embodiment of the application provides
Method.This method may include following steps:
Step 301, by Visual identification technology, identify that the n times article that user executes in unmanned counter picks and places behavior pair
The pick-and-place article summation answered obtains visual identity as a result, n is positive integer.
Step 302, the corresponding article Gravity changer value of detection visual identity result is corresponding with unmanned counter overall practical
Whether Gravity changer value matches.
The introduction explanation of step 301 to 302 can be found in foregoing embodiments, and details are not described herein again.
Step 303, if the corresponding article Gravity changer value of visual identity result is matched with overall A/W changing value,
Visual identity result is determined as n times article and picks and places the corresponding pick-and-place article summation of behavior.
When the corresponding article Gravity changer value of visual identity result is matched with overall A/W changing value, show vision
Recognition result is accurately, visual identity result can be determined as to n times article and pick and place the corresponding pick-and-place article summation of behavior.
Step 304, if the corresponding article Gravity changer value of visual identity result and overall A/W changing value mismatch,
It then detects visual identity result and whether overall A/W changing value meets default segmentation rule.
Default segmentation rule is the rule for presetting out according to the behavior of user by technical staff.Technical staff can be with
According to following principle, a plurality of default segmentation rule is set:
1, do not conflict mutually between default segmentation rule.
2, executing the mistake occurred after default segmentation rule is acceptable, for example, goods categories conjecture can occur
Mistake, but the goods categories and the practical goods categories price picked and placed of user guessed are consistent.
3, the more specific default segmentation rule the better.
4, Article 2 is not just executed when meeting first default segmentation rule to default segmentation rule setting priority
Default segmentation rule.
Technical staff can add different default segmentation rules according to the case where test, so as to quickly solve article
The some unreasonable situations occurred in identification process.Default segmentation rule can continuous adjusting and optimizing during the test, from
And it is correct for guaranteeing final default segmentation rule.Default segmentation rule can be preferentially configured from the angle of quantity.Example
Such as, overall A/W changing value is m times of the corresponding article gravity value of visual identity result, and m is the positive integer no more than 6.
Step 305, if meeting default segmentation rule, it is corresponding that the pick-and-place behavior of n times article is obtained according to default segmentation rule
Pick-and-place article summation.
In one example, it is the corresponding object of visual identity result that default segmentation rule, which includes overall A/W changing value,
The integral multiple of product Gravity changer value.For example, visual identity is the result is that user has taken, one bottle of cola, one bottle of laughable gravity value are
5, it is four times of 5 that overall A/W changing value, which is 20,20, then visual identity result and overall A/W changing value meet pre-
If fusion rule, it is four bottles of colas that article, which picks and places the corresponding pick-and-place article summation of behavior,.
In another example, default segmentation rule include when the corresponding article of visual identity result is specified article,
The gravity value for the specified article that the corresponding Gravity changer value of visual identity result is twice.For example, visual identity the result is that with
Family has been taken one bottle of cola, and cola has the activity bought one and got one free, and one bottle of laughable gravity value is 5, overall A/W changing value
It is 10,10 is 2 times of 5, then visual identity result and overall A/W changing value meet default segmentation rule, and article picks and places row
It is two bottles of colas for corresponding pick-and-place article summation.
Step 306, it if being unsatisfactory for default segmentation rule, generates n times article and picks and places the corresponding pick-and-place article summation of behavior
A variety of prediction results.
Optionally, according to each article provided in visual identity result, overall A/W changing value and unmanned counter
Gravity value, generate a variety of prediction results.
For example, having cola, Sprite, Yoghourt and fruit juice in unmanned counter, laughable gravity value is 5, and the gravity value of Sprite is
4, the gravity value of Yoghourt is 4, and the gravity value of fruit juice is 3, visual identity cola the result is that user has taken, and overall A/W becomes
Change value is 9, then a kind of prediction result be user taken one bottle can happy one bottle of Sprite, another prediction result is that user takes
One bottle can happy one bottle of Yoghourt.
In one possible implementation, when the corresponding article Gravity changer value of visual identity result and overall practical weight
When power changing value mismatches, does not need detection visual identity result and whether overall A/W changing value meets default fusion rule
Then, but directly according to the weight of each article provided in visual identity result, overall A/W changing value and unmanned counter
Force value generates a variety of prediction results.
Step 307, for i-th kind of prediction result in a variety of prediction results, include in i-th kind of prediction result every is generated
The corresponding confidence level of one article, i is positive integer.
Optionally, the corresponding confidence level of each article is generated according to following steps:
1, for j-th of article for including in i-th kind of prediction result, classification belonging to j-th of article is determined, j is positive whole
Number.
Optionally, classification belonging to article has following five kinds:
First category refers to identifies obtain and Gravity Matching by Visual identification technology in single article pick-and-place behavior
Article;
Second category refers to does not identify obtain and gravity not by Visual identification technology in single article pick-and-place behavior
The article matched;
Third classification refers to the article obtained in single article pick-and-place behavior based on gravity prediction;
4th classification refers to the article identified in n times article pick-and-place behavior by Visual identification technology;
5th classification refers to through Visual identification technology the unidentified article being obtained in n times article pick-and-place behavior.
2, the corresponding confidence level of the classification according to belonging to j-th of article assigns rule, generates the corresponding confidence of j-th of article
Degree.
Optionally, this step includes following sub-step:
If 2.1, j-th of article belongs to first category, by the corresponding visual identity confidence level of j-th of article and the first threshold
The sum of value, is determined as the corresponding confidence level of j-th of article;Wherein, first category refers to through Visual identification technology in single object
Article obtain and Gravity Matching is identified in product pick-and-place behavior.
In the embodiment of the present application, the corresponding visual identity confidence level of j-th of article, refers to and is existed by Visual identification technology
Identification obtains the confidence level exported when j-th of article.
The corresponding article of single visual identity result is article A, and the gravity value of article A totality corresponding with unmanned counter
A/W changing value matches, then article A belongs to first category, illustrate user taken article A a possibility that it is very big, therefore
By the sum of the corresponding visual identity confidence level of article A and first threshold as the corresponding confidence level of article A.
If 2.2, j-th of article belongs to second category, by the corresponding visual identity confidence level of j-th of article and the second threshold
The difference of value, is determined as the corresponding confidence level of j-th of article;Wherein, second category refers to through Visual identification technology in single object
The obtain and unmatched article of gravity is identified in product pick-and-place behavior.
The corresponding article of single visual identity result is article A, and the gravity value of article A totality corresponding with unmanned counter
A/W changing value mismatches, then article A belongs to second category, illustrate user taken article A a possibility that be not it is very big,
Therefore using the difference of the corresponding visual identity confidence level of article A and second threshold confidence level corresponding as article A.
If 2.3, j-th of article belongs to third classification, third threshold value is determined as the corresponding confidence level of j-th of article;
Wherein, third classification refers to the article obtained in single article pick-and-place behavior based on gravity prediction.
The corresponding article of visual identity result is article B, and the sum of gravity value of the gravity value of article A and article B is for nobody
The corresponding overall A/W changing value of counter illustrates that user may take article A and article B, then article A belongs to third class
Not, using third threshold value as the corresponding confidence level of article A.
If 2.4, j-th of article belongs to the 4th classification, the corresponding visual identity confidence level of j-th of article is determined as
The corresponding confidence level of j-th of article;Wherein, the 4th classification refers to is known in n times article pick-and-place behavior by Visual identification technology
The article not being obtained.
Visual identification technology identifies in 1 article pick-and-place behavior obtains article A, then article A belongs to the 4th classification, by object
The corresponding visual identity confidence level of product A is as the corresponding confidence level of article A.
If 2.5, j-th of article belongs to the 5th classification, the 4th threshold value is determined as the corresponding confidence level of j-th of article;
Wherein, the 5th classification refers to through Visual identification technology the unidentified article being obtained in n times article pick-and-place behavior.
Visual identification technology is unidentified in 1 article pick-and-place behavior to be obtained article A, then article A belongs to the 5th classification,
Using the 4th threshold value as the corresponding confidence level of article A.
First threshold, second threshold, third threshold value and the 4th threshold value are obtained according to test of many times one reasonable
Numerical value.
Step 308, the average value for calculating the corresponding confidence level of each article for including in i-th kind of prediction result, obtains i-th
The corresponding confidence level of kind prediction result.
For example, containing Sprite and cola in the first prediction result, the corresponding confidence level of Sprite is 0.7, and cola is corresponding
Confidence level be 0.5, then the corresponding confidence level of the first prediction result is (0.7+0.5)/2=0.6.
Step 309, by the maximum target prediction of confidence level in a variety of prediction results as a result, being determined as n times article picks and places row
For corresponding pick-and-place article summation.
In conclusion in technical solution provided by the embodiments of the present application, by according to each object for including in prediction result
The corresponding confidence level of classification belonging to product assigns rule, generates the corresponding confidence level of each article, the classification according to belonging to article
Assign article corresponding confidence level, the corresponding confidence level of article is more reasonable, so that the corresponding confidence level of prediction result is more
Rationally, finally make the accuracy rate for picking and placing article in identification user higher.
In another alternative embodiment provided based on any of the above-described embodiment, visual identity result can be by as follows
Step obtains:
The first, by Visual identification technology, it is corresponding that the single article that identification user executes in unmanned counter picks and places behavior
Pick-and-place article.
Single article picks and places behavior and refers to that single takes article behavior or single puts back to article behavior.Single article picks and places row
Refer to the article that user once takes or the article that user once puts back to for corresponding pick-and-place article.For example, user is from unmanned goods
It has taken in cabinet one bottle of cola, by Visual identification technology, has identified that user has taken one bottle of cola from unmanned counter.
Optionally, this step includes following sub-step:
1, Articles detecting is carried out to the image that camera in unmanned counter is shot by the first article identification model.
First article identification model is the model for identifying merchandise classification.First article identification model can pass through machine
Study and neural network model obtain.To the first article identification model input one with identification Item Information image, first
Article identification model can recognize that the corresponding goods categories of the image.
If 2, the first article identification model detects article from image, by the first article identification model to image into
The identification of row article, obtains the first recognition result and the corresponding confidence level of the first recognition result.
Assuming that there is 4 cameras in unmanned counter, the corresponding frame per second of each camera is 120fps, then the first article is known
Other model needs 480 frames/second recognition rate, so when the first article identification model detects article from image,
It identifies that image carries out article identification by the first article, saves the workload of computer equipment.
Optionally, article identification is carried out at least one image by the first article identification model, obtains at least one figure
As corresponding article and the corresponding confidence level of the corresponding article of at least one image, when at least one image is each
When self-corresponding article is inconsistent, the corresponding article of the maximum image of confidence level is chosen as the first recognition result, this is maximum
Confidence level is as the corresponding confidence level of the first recognition result.
3, article identification is carried out to the article region in image by the second article identification model, obtains the second recognition result
And the second corresponding confidence level of recognition result.
Second article identification model is the model for identifying merchandise classification, the accuracy of identification of the second article identification model
Greater than the accuracy of identification of the first article identification model.
Optionally, article identification is carried out at least one article region by the second article identification model, obtains at least one
A corresponding article in article region and at least one corresponding confidence level of the corresponding article in article region, when extremely
When a few corresponding article in article region is inconsistent, the corresponding article in the maximum article region of confidence level is chosen as the
Two recognition results, the maximum confidence level is as the corresponding confidence level of the second recognition result.
As shown in figure 4, the article it illustrates a kind of schematic diagram of image of the application one embodiment offer, in image
For cola 41, the first article identification model is to identify to image 40, and the second article identification model is to article region 42
It is identified.Therefore, the accuracy of identification of the second article identification model is greater than the accuracy of identification of the first article identification model.
The article region in image is carried out except through the second article identification model to pass through phase outside article identification
Article identification is carried out to the article region in image like the mode of retrieval.To the article area in image by way of similar to search
Domain carry out article identification be by article region and nobody the lift-on-lift-off all items picture library that is in advance collected in image into
Row comparison, if the article region in image is similar to the image of article A in all items picture library, then it is assumed that the image
In the corresponding article in article region be article A.
If 4, the corresponding confidence level of the second recognition result is greater than the corresponding confidence level of the first recognition result, by the second identification
As a result it is determined as single article and picks and places the corresponding pick-and-place article of behavior.
If 5, the corresponding confidence level of the second recognition result confidence level corresponding less than the first recognition result, by the first identification
As a result it is determined as single article and picks and places the corresponding pick-and-place article of behavior.
Confidence level is higher, and recognition result is more accurate, picks and places behavior pair for the high recognition result of confidence level as final single
The pick-and-place article answered, more rationally.
The second, the corresponding pick-and-place article of behavior is picked and placed according to single article, determines that n times article picks and places the corresponding pick-and-place of behavior
Article summation obtains visual identity result.
The article obtained each time is picked and placed the corresponding pick-and-place article of behavior to add up, n times article is obtained and picks and places behavior
Corresponding pick-and-place article summation.
For example, it is one bottle of cola that first time article, which picks and places the corresponding pick-and-place article of behavior, second of article picks and places behavior pair
The pick-and-place article answered is one bottle of Sprite, then twice article pick and place corresponding the pick-and-places article summation of behavior be one bottle can happy one bottle avenge
It is green.
In conclusion in technical solution provided by the embodiments of the present application, when the first article identification model is detected from image
When to article, just article is identified, alleviates the workload of computer equipment.By respectively being obtained to two identification models
The corresponding confidence level of recognition result be compared, choose the corresponding article of maximum confidence level as single article pick and place behavior
Corresponding pick-and-place article, it is more reasonable, so that the corresponding pick-and-place article summation of n times article pick-and-place behavior is more reasonable, therefore final
Obtained visual identity result is more accurate.
In the embodiment of the present application, article identification model can be obtained by computer equipment training, can also be set by other
Standby training obtains, and is then forwarded to computer equipment, the present embodiment does not limit the source of article identification model.
When article identification model is obtained by computer equipment training, computer equipment first creates article identification model;Again
Obtain training sample, the training sample include comprising article image and for marking the classification of each article and position in the image
The markup information set;Finally article identification model is trained according to the training sample.
Wherein, computer equipment can construct article identification model, this reality with the structural model of any convolutional neural networks
Example is applied to be not construed as limiting.In one possible implementation, computer equipment is with FSSD (Feature Fusion Single
Shot Multibox Detector, the single-lens more box detectors of Fusion Features) algorithm building article identification model, FSSD algorithm
The feature of more lower layers (low-level) can be introduced into high level, although available so more accurately examine Small object
It surveys, still, it is larger to will lead to calculation amount.
Due to requiring the computational efficiency of computer equipment higher (such as 480 frames/second), so, it can be in FSSD algorithm
Certain layers of port number (channel) is cut, to meet the needs of computational efficiency.
Optionally, the embodiment of the present application also passes through track algorithm, as Staple algorithm constructs article identification model.Staple
Algorithm is improved based on correlation filtering, and by HOG, (Histogram of Oriented Gradient, direction gradient are straight
Side's figure)-KCF (Kernelized Correlation Filters, core correlation filter) feature and Color (color)-KCF be special
Sign is combined to track to article.HOG feature is more sensitive to deformation and motion blur, but tracks and imitate to color change
Fruit is preferable;On the contrary, Color feature is more sensitive to color, but it is more preferable to deformation and motion blur tracking effect.Therefore, two
The fusion of person is able to solve the problems such as deformation encountered in most of tracking, dimensional variation, motion blur, color change.Therefore,
The article identification model constructed by Staple algorithm can also be by according to target image frame and an at least history image frame
The motion track for generating article determines that user behavior is that article takes behavior or article puts back to behavior.For example, the movement of article
Track is the interior from unmanned counter, it is determined that user behavior is that article is taken behavior;The motion track of article be from
The outside of unmanned counter is internally, it is determined that user behavior is that article puts back to behavior.
As shown in figure 5, it illustrates the item identification methods applied to unmanned counter that the application one embodiment provides
Timing diagram.
1: Articles detecting is carried out to the image that camera in unmanned counter is shot by the first article identification model;
2: if the first article identification model detects article from image, by the first article identification model to image into
The identification of row article, obtains the first recognition result and the corresponding confidence level of the first recognition result;
3: article identification being carried out to the article region in image by the second article identification model, obtains the second recognition result
And the second corresponding confidence level of recognition result;
4: if the corresponding confidence level of the second recognition result is greater than the corresponding confidence level of the first recognition result, by the second identification
As a result it is determined as single article and picks and places the corresponding pick-and-place article of behavior;
5: if the corresponding confidence level of the second recognition result confidence level corresponding less than the first recognition result, by the first identification
As a result it is determined as single article and picks and places the corresponding pick-and-place article of behavior;
6: the corresponding pick-and-place article of behavior being picked and placed according to single article, determines that n times article picks and places the corresponding pick-and-place object of behavior
Product summation obtains visual identity result;
7: the corresponding article Gravity changer value of detection visual identity result overall A/W corresponding with unmanned counter becomes
Whether change value matches;
8: if the corresponding article Gravity changer value of visual identity result is matched with overall A/W changing value, by vision
Recognition result is determined as n times article and picks and places the corresponding pick-and-place article summation of behavior;
9: if the corresponding article Gravity changer value of visual identity result and overall A/W changing value mismatch, detecting
Whether visual identity result and overall A/W changing value meet default segmentation rule;
10: if meeting default segmentation rule, n times article being obtained according to default segmentation rule and picks and places the corresponding pick-and-place of behavior
Article summation;
11: if being unsatisfactory for default segmentation rule, generating n times article and pick and place a variety of of the corresponding pick-and-place article summation of behavior
Prediction result;
12: for j-th of article for including in i-th kind of prediction result, determining classification belonging to j-th of article, j is positive whole
Number;
13: assigning rule according to the corresponding confidence level of classification belonging to j-th of article, generate that j-th of article is corresponding to be set
Reliability;
14: calculating the average value for the corresponding confidence level of each article for including in i-th kind of prediction result, obtain i-th kind in advance
Survey the corresponding confidence level of result;
15: the maximum target prediction of confidence level in a variety of prediction results is corresponded to as a result, being determined as n times article and picking and placing behavior
Pick-and-place article summation.
In a practical application scene, by taking user buys nobody lift-on-lift-off article as an example.It can be put in unmanned counter
Put a variety of commodity that purchase is selected for user, such as beverage, snacks, daily necessities extensive stock.For example, unmanned counter is divided into multilayer
Counter is placed with the commodity sold on each layer of counter, and is provided with gravity pallet below each layer of counter.In addition, nothing
At least one camera is installed, unmanned counter is internally integrated the computer for having calculating and processing capacity and sets in people's counter
It is standby, and it is arranged or shows graphic code on the cabinet body of unmanned counter.Illustratively, primary complete article purchasing process is as follows:
1, user scans above-mentioned graphic code using terminals such as mobile phones and carries out authentication, user bound account, and unlocks nothing
People's counter.After unlocking, cabinet door can be opened unmanned counter by user.
2, user opens cabinet door, executes article at least once and picks and places behavior, the object to be bought is selected in unmanned counter
Product.
3, during user selects article, unmanned counter can pass through Visual identification technology and gravity described above
Matching etc. means, identify user article each time pick and place behavior be taking or putting back to article, and identify take or
The classification for the article put back to, and then show that the article of the user during the cabinet door of unmanned counter is opened picks and places the corresponding pick-and-place of behavior
Article summation.
4, after the completion of selecting, user closes the cabinet door of unmanned counter.
5, after detecting that the cabinet door of unmanned counter is closed, computer equipment calculates all items that user takes
Total price, and deduct from user account the total price of all items that user takes.
It is applied to be illustrated in unmanned supermarket with unmanned counter below.An at least unmanned goods in unmanned supermarket
Cabinet.Unmanned supermarket is provided with access control equipment, and access control equipment built-in camera can carry out recognition of face to user, if user
It is not the member of the unmanned supermarket, then meeting display reminding information on the panel of access control equipment, prompts user's registration face to become and be somebody's turn to do
The member of unmanned supermarket, and open account and exempt from close withhold.After user registration success, access control equipment is opened, and user is super into nobody
City;If user has been the member of unmanned supermarket, after the camera of access control equipment carries out recognition of face to user, access control equipment
It automatically opens, user enters unmanned supermarket.User's picking commodities in unmanned counter, be integrated in unmanned counter have calculating and
The computer equipment of processing capacity.Unmanned counter obtains user in unmanned goods by means such as Visual identification technology and Gravity Matchings
The article executed in cabinet picks and places the corresponding pick-and-place article summation of behavior, sends checkout station for the pick-and-place article summation.Checkout station
It identifies that the confirmation of user's face and user pay gesture by camera, withholds from user account automatically.If commodity are
Through settling accounts, then access control equipment is opened, and user may exit off unmanned supermarket.
Following is the application Installation practice, can be used for executing the application embodiment of the method.It is real for the application device
Undisclosed details in example is applied, the application embodiment of the method is please referred to.
Referring to FIG. 6, it illustrates the item identification devices applied to unmanned counter that the application one embodiment provides.
The device, which has, realizes the exemplary function of the above method, and the function can also be executed corresponding by hardware realization by hardware
Software realization.The device 600 may include: result obtain module 610, gravity detection module 620, prediction of result module 630,
Confidence calculations module 640 and result determining module 650.
The result obtains module 610, for by Visual identification technology, identification user to execute in the unmanned counter
N times article pick and place corresponding the pick-and-places article summation of behavior, obtain visual identity as a result, the n is positive integer.
The gravity detection module 620, for detecting the corresponding article Gravity changer value of the visual identity result and institute
State whether the corresponding overall A/W changing value of unmanned counter matches.
The prediction of result module 630, for when the corresponding article Gravity changer value of the visual identity result with it is described
When overall A/W changing value mismatches, generates the n times article and pick and place a variety of pre- of the corresponding pick-and-place article summation of behavior
Survey result.
The confidence calculations module 640, for calculating the corresponding confidence level of a variety of prediction results.
The result determining module 650 is used for the maximum target prediction of confidence level described in a variety of prediction results
As a result, being determined as the n times article picks and places the corresponding pick-and-place article summation of behavior.
In conclusion in technical solution provided by the embodiments of the present application, by working as the corresponding article weight of visual identity result
When power changing value overall A/W changing value corresponding with unmanned counter mismatches, a variety of prediction results are generated, and calculate more
The kind corresponding confidence level of prediction result, chooses the maximum prediction result of confidence level in a variety of prediction results, as final
Article summation is picked and placed, ensure that the accuracy rate in the goods categories that identification user takes.
Optionally, as shown in fig. 7, the confidence calculations module 640, comprising: confidence level generation unit 641 and confidence level
Computing unit 642.
The confidence level generation unit 641, for generating for i-th kind of prediction result in a variety of prediction results
The corresponding confidence level of each article for including in i-th kind of prediction result, the i are positive integer.
The confidence computation unit 642, it is corresponding for calculating each article for including in i-th kind of prediction result
The average value of confidence level obtains the corresponding confidence level of i-th kind of prediction result.
Optionally, the confidence level generation unit 641, comprising: it is single that classification determines that subelement 6411 and confidence level generate son
Member 6412.
The classification determines subelement 6411, for j-th of article for including in i-th kind of prediction result, really
Classification belonging to fixed j-th of article, the j are positive integer.
The confidence level generates subelement 6412, for the corresponding confidence level of classification according to belonging to j-th of article
Rule is assigned, the corresponding confidence level of j-th of article is generated.
Optionally, the confidence level generates subelement 6412, is used for:
When j-th of article belongs to first category, by the corresponding visual identity confidence level of j-th of article and the
The sum of one threshold value is determined as the corresponding confidence level of j-th of article;Wherein, the first category refers to through the vision
Identification technology identifies article obtain and Gravity Matching in single article pick-and-place behavior;
When j-th of article belongs to second category, by the corresponding visual identity confidence level of j-th of article and the
The difference of two threshold values is determined as the corresponding confidence level of j-th of article;Wherein, the second category refers to through the vision
Identification technology identifies the obtain and unmatched article of gravity in single article pick-and-place behavior;
When j-th of article belongs to third classification, third threshold value is determined as the corresponding confidence of j-th of article
Degree;Wherein, the third classification refers to the article obtained in single article pick-and-place behavior based on gravity prediction;
When j-th of article belongs to four classifications, by the corresponding visual identity confidence level of j-th of article, really
It is set to the corresponding confidence level of j-th of article;Wherein, the 4th classification refers to through the Visual identification technology described
The article being obtained is identified in n times article pick-and-place behavior;
When j-th of article belongs to five classifications, the 4th threshold value is determined as the corresponding confidence of j-th of article
Degree;Wherein, the 5th classification refers to obtains by the way that the Visual identification technology is unidentified in the n times article pick-and-place behavior
The article crossed;
Wherein, the corresponding visual identity confidence level of j-th of article, refers to and is being identified by the Visual identification technology
Obtain the confidence level exported when j-th of article.
Optionally, the prediction of result module 630, for according to the visual identity result, the overall A/W
The gravity value of each article provided in changing value and the unmanned counter generates a variety of prediction results.
Optionally, the result determining module 650, comprising: article recognition unit 651 and result determination unit 652.
The article recognition unit 651, for identifying the user in the unmanned goods by the Visual identification technology
The single article executed in cabinet picks and places the corresponding pick-and-place article of behavior.
The result determination unit 652 determines institute for picking and placing the corresponding pick-and-place article of behavior according to the single article
It states n times article and picks and places the corresponding pick-and-place article summation of behavior, obtain the visual identity result.
Optionally, the article recognition unit 651, is used for:
Articles detecting is carried out to the image that camera in the unmanned counter is shot by the first article identification model;
When the first article identification model detects article from described image, mould is identified by first article
Type carries out article identification to described image, obtains the first recognition result and the corresponding confidence level of first recognition result;
Article identification is carried out to the article region in described image by the second article identification model, obtains the second identification knot
Fruit and the corresponding confidence level of second recognition result;
When the corresponding confidence level of second recognition result confidence level corresponding greater than first recognition result, by institute
It states the second recognition result and is determined as the corresponding pick-and-place article of the single article pick-and-place behavior;
When the corresponding confidence level of second recognition result confidence level corresponding less than first recognition result, by institute
It states the first recognition result and is determined as the corresponding pick-and-place article of the single article pick-and-place behavior.
Optionally, described device 600 further include: first detection module 660 and article guess module 670.
The first detection module 660, for when the corresponding article Gravity changer value of the visual identity result with it is described
When overall A/W changing value mismatches, detects the visual identity result and whether the overall A/W changing value is full
Sufficient default segmentation rule.
The article guesses module 670, for when meeting the default segmentation rule, according to the default segmentation rule
It obtains the n times article and picks and places the corresponding pick-and-place article summation of behavior.
The prediction of result module 630, is also used to when being unsatisfactory for the default segmentation rule, generates the n times article
The corresponding a variety of prediction results for picking and placing article summation of pick-and-place behavior.
Optionally, the result determining module 650 is also used to when the corresponding article Gravity changer of the visual identity result
When value is matched with the overall A/W changing value, the visual identity result is determined as the n times article and picks and places behavior
Corresponding pick-and-place article summation.
It should be noted that device provided by the above embodiment, when realizing its function, only with above-mentioned each functional module
It divides and carries out for example, can according to need in practical application and be completed by different functional modules above-mentioned function distribution,
The content structure of equipment is divided into different functional modules, to complete all or part of the functions described above.In addition,
Apparatus and method embodiment provided by the above embodiment belongs to same design, and specific implementation process is detailed in embodiment of the method, this
In repeat no more.
Referring to FIG. 8, the structural block diagram of the computer equipment 800 provided it illustrates the application one embodiment.The meter
It calculates machine equipment 800 and refers to the electronic equipment for having calculating and processing capacity, such as PC, server etc..The computer equipment 800 can
It is applied to nobody lift-on-lift-off item identification method for implement to provide in above-described embodiment.
In general, computer equipment 800 includes: processor 801 and memory 802.
Processor 801 may include one or more processing cores, such as 4 core processors, 8 core processors etc..Place
Reason device 801 can use DSP (Digital Signal Processing, Digital Signal Processing), FPGA (Field
Programmable Gate Array, field programmable gate array), PLA (Programmable Logic Array, may be programmed
Logic array) at least one of example, in hardware realize.Processor 801 also may include primary processor and coprocessor, master
Processor is the processor for being handled data in the awake state, also referred to as CPU (Central Processing
Unit, central processing unit);Coprocessor is the low power processor for being handled data in the standby state.?
In some embodiments, processor 801 can be integrated with GPU (Graphics Processing Unit, image processor),
GPU is used to be responsible for the rendering and drafting of content to be shown needed for display screen.In some embodiments, processor 801 can also be wrapped
AI (Artificial Intelligence, artificial intelligence) processor is included, the AI processor is for handling related machine learning
Calculating operation.
Memory 802 may include one or more computer readable storage mediums, which can
To be non-transient.Memory 802 may also include high-speed random access memory and nonvolatile memory, such as one
Or multiple disk storage equipments, flash memory device.In some embodiments, the non-transient computer in memory 802 can
Storage medium is read for storing at least one instruction, at least one instruction performed by processor 801 for realizing this Shen
Please in embodiment of the method provide the item identification method applied to unmanned counter.
In some embodiments, terminal 800 is also optional includes: peripheral device interface 803 and at least one peripheral equipment.
It can be connected by bus or signal wire between processor 801, memory 802 and peripheral device interface 803.Each peripheral equipment
It can be connected by bus, signal wire or circuit board with peripheral device interface 803.Specifically, peripheral equipment may include: display
At least one of screen 804, voicefrequency circuit 805, communication interface 806 and power supply 807.
It will be understood by those skilled in the art that structure shown in Fig. 8 does not constitute the restriction to computer equipment 800,
It may include perhaps combining certain components than illustrating more or fewer components or being arranged using different components.
In this example in embodiment, a kind of computer equipment is additionally provided, the computer equipment includes processor and deposits
Reservoir is stored at least one instruction, at least a Duan Chengxu, code set or instruction set in the memory.Described at least one
Instruction, at least a Duan Chengxu, code set or instruction set are configured to be executed by one or more than one processor, on realizing
State the item identification method applied to unmanned counter.
In the exemplary embodiment, a kind of computer readable storage medium is additionally provided, is stored in the storage medium
At least one instruction, at least a Duan Chengxu, code set or instruction set, at least one instruction, an at least Duan Chengxu, institute
It states code set or described instruction collection and realizes the above-mentioned article applied to unmanned counter when being executed by the processor of computer equipment
Recognition methods.
Optionally, above-mentioned computer readable storage medium can be ROM, random access memory (RAM), CD-ROM, magnetic
Band, floppy disk and optical data storage devices etc..
In the exemplary embodiment, a kind of computer program product is additionally provided, when the computer program product is performed
When, for realizing the above-mentioned item identification method applied to unmanned counter.
It should be understood that referenced herein " multiple " refer to two or more."and/or", description association
The incidence relation of object indicates may exist three kinds of relationships, for example, A and/or B, can indicate: individualism A exists simultaneously A
And B, individualism B these three situations.Character "/" typicallys represent the relationship that forward-backward correlation object is a kind of "or".In addition, herein
Described in number of steps, the merely exemplary a kind of possible execution sequencing shown between step, in some other implementations
In example, above-mentioned steps can not also be executed according to number order, and such as the step of two different numbers is performed simultaneously or two
The step of different numbers, executes according to the sequence opposite with diagram, and the embodiment of the present application is not construed as limiting this.
Those of ordinary skill in the art will appreciate that realizing that all or part of the steps of above-described embodiment can pass through hardware
It completes, relevant hardware can also be instructed to complete by program, the program can store in a kind of computer-readable
In storage medium, storage medium mentioned above can be read-only memory, disk or CD etc..
The foregoing is merely the exemplary embodiments of the application, all in spirit herein not to limit the application
Within principle, any modification, equivalent replacement, improvement and so on be should be included within the scope of protection of this application.