CN106951930A - A kind of instrument localization method suitable for Intelligent Mobile Robot - Google Patents
A kind of instrument localization method suitable for Intelligent Mobile Robot Download PDFInfo
- Publication number
- CN106951930A CN106951930A CN201710239499.5A CN201710239499A CN106951930A CN 106951930 A CN106951930 A CN 106951930A CN 201710239499 A CN201710239499 A CN 201710239499A CN 106951930 A CN106951930 A CN 106951930A
- Authority
- CN
- China
- Prior art keywords
- sample
- instrument
- training
- classifier
- window
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of instrument localization method suitable for Intelligent Mobile Robot, image to be detected carries out Primary Location first with Adaboost graders, position obtained multiple candidate regions and reuse the secondary positioning of SVM classifier progress, the region for being determined as instrument twice is the regional location of image where instrument.The region similar to instrument is oriented using Adaboost graders, the recall rate of instrument is greatly improved；The deficiency of Adaboost graders is made up using the graders of SVM bis-, color characteristic and textural characteristics has been merged, finally accurately candidate region is classified.The scalability that verification and measurement ratio also increases system is not only increased, the requirement that transformer station positions to instrument is met.
Description
Technical field
The invention belongs to image identification technical field, a kind of more particularly to instrument suitable for Intelligent Mobile Robot is determined
Position method.
Background technology
Transformer station is the hinge that power system carries out electrical energy transportation, and inside has substantial amounts of instrument, switch, oil level indicator etc. secondary
Equipment.To ensure the safety of power system, it is necessary to periodically carry out inspection and record to the equipment in transformer station.At present, China
Substation inspection mode is from traditional artificial inspection on foot to automation inspection transition.Intelligent Mobile Robot is exactly special
Door is the automation inspection device that transformer station services, and one of its major function is to complete the instrument inspection in transformer station.
Instrument inspection is including taking pictures, instrument is positioned, camera parameter is adjusted, meter reading recognizes several flows.Wherein, instrument
The effect of positioning, by the adjustment for directly determining camera parameter and the effect of Recognition of Reading, is the key link for completing instrument inspection.
And the instrument in transformer station is positioned over outdoor environment, the Instrument image that crusing robot is shot can be influenceed by external environment.
These influences include：Instrument intensity of illumination has apparatus glass cover under obvious gap, direct projection light to produce instead caused by different weather
Instrument surface blur caused by optical phenomenon, overcast and rainy foggy weather, diversified equipment causes to include in shooting image in transformer station
The background of large amount of complex.These unfavorable factors, significantly increase the difficulty of instrument positioning.
For instrument orientation problem, existing main stream approach has two kinds：Localization method based on template matches and based on spy
Levy the localization method of Point matching.Localization method based on template matches is in altimetric image to be checked using a fixed instrument template
In all regions of altimetric image to be checked are traveled through in the way of change of scale and sliding window, one is calculated to each region traversed
The individual similarity between template, final similarity highest region of choosing completes instrument with this and determined as instrument region
Position.This method is although simple in construction, but more sensitive to light change and noise, light and actual bat when template image
When taking the photograph the light of image and having during larger difference or there is similar area in the background of shooting image, it may appear that error hiding phenomenon,
Cause instrument positioning failure.
When carrying out instrument positioning using the method for Feature Points Matching, extracted first with feature operator in altimetric image to be checked
Characteristic point, is then matched with the characteristic point of template image, and multiple characteristic points finally according to matching carry out affine transformation, positioning
Go out meter location.Wherein conventional feature operator is SIFT feature operator.It is use in current converting station instrument alignment system more
This localization method based on SIFT feature Point matching.This method can overcome illumination effect to a certain extent, and have
Yardstick and rotational invariance, have stronger adaptability to the situation of On Local Fuzzy.But, using this localization method, still deposit
Problem both ways：On the one hand requirement of this method to characteristic point is high, is obscured or image when the image of shooting has bulk
When characteristic point is less, easily occur mismatching phenomenon；When the background of shooting image has similar features point, affine change can be caused
Change unsuccessfully, it is impossible to position instrument.On the other hand, to ensure accuracy rate, it is necessary to provide the specific of multiple yardsticks for each instrument
Masterplate.By taking lightning-arrest instruments and meters as an example, there is more or less a hundred lightning-arrest instruments and meters in a transformer station, although each instrument appearance is identical, but
It is because it shoots orientation and the difference of instrument putting position, it is necessary to be the multiple yardstick templates of each arrester instrument to collect.When
When some instrument occurs change in location or did the dust cleaning on surface, it is necessary to resurvey the template of the instrument.So not
But add the scalability that artificial operation also greatly reduces system.
The content of the invention
It is an object of the invention to provide a kind of adaptable, positioning precision is high, scalability is good is applied to transformer station
The instrument localization method of crusing robot.
Therefore, the technical scheme is that：A kind of instrument localization method suitable for Intelligent Mobile Robot, including
Following steps：
1）Collect sample：Sample includes positive sample and negative sample, and positive sample is the image of the only dial plate containing instrument, and negative sample is not wrap
Background image containing instrument；
2）Adaboost classifier trainings：Utilize step 1）The sample of collection is trained, selection of the training process including feature,
The training of Weak Classifier, the training of strong classifier；The sample used during training is constant, but sample weight according to the effect of classification
And constantly change, so it is considered as constantly being trained grader using new sample；Finally classifier stage is associated in
Together, Decision Classfication device is formed；
3）Two graders based on SVM；Construction decision function is calculated using the color characteristic and textural characteristics of image, two are obtained
Grader；
4）Instrument is positioned：One-level detection is carried out first by the Adaboost graders with high detection rate, can be with Primary Location
Go out the multiple regions similar with instrument；That is one-level detection improves the probability for orienting instrument；Then two are done using SVM classifier
Secondary positioning, makes up the deficiency of Adaboost graders, and then orients from candidate region instrument.
Further, the step 4）Described instrument localization method specifically includes following steps：
D1 picture to be detected) is inputted（Dimension of picture is）, and image is pre-processed；Turn including cromogram at gray-scale map
Reason, medium filtering and gaussian filtering process；
D2 detection window size) is initialized；Window selection uses width and height identical size, and by detection window width
It is initialized as 20 pixels；
D3) it is using sizeDetection window, since initial position, with two pixels of horizontal direction and Vertical Square
Translated to two pixels for unit on whole pictures, often translate and can once obtain a new detection window；
D4) for a detection window, corresponding all Haar-like subcharacters under the window are calculated；
D5 classification judgement) is carried out to the image-region covered under detection window using Adaboost graders, if being judged as instrument
Table, then record position and the size of the window；
D6) adjustment window size is：If met, then into step d7), otherwise enter step
Rapid d3)；
D7 detection window) is merged；The size of detection window is different when being positioned because of instrument, can be detected near same position many
Individual region, therefore the window of analogous location is merged, it is to avoid repeat；
D8 all candidate's instrument regions) are extracted, and size normalization is carried out to it；
D9 classification judgement, output result) are carried out to candidate region using SVM classifier；
D10) output meter is in the position of image.
Further, the step 1）Described in positive sample need to carry out dimension normalization, negative sample need not carry out
Normalization, but its size is bigger than normalized positive sample, and positive and negative samples are both needed to be labeled；The ratio of positive and negative samples is 1：
2.5~1：Between 3.
Further, the step 2）The specific training algorithm of described Adaboost classifier trainings includes following step
Suddenly：
A1 training sample) is inputted, whereinRepresent that sample belongs to negative sample
This,Expression sample is positive sample, and n is sample size；
A2) initialization sample weight, if positive sample, then sets sample weights；If negative sample, then weigh
Reset and be set to, whereinWithThe number of positive negative sample is represented respectively；
a3）The optimal Weak Classifier of training；
a4）Weak Classifier is cascaded into strong classifier
；
Represent Weak Classifier；Represent the weight of this Weak Classifier；
a5）Classification and Identification is done to sample using strong classifier, and calculates weighting fault rate, if error rate is default more than reaching
Standard and frequency of training are less than default value, then proceed the sample of classification error as additional sample to train；Otherwise
Deconditioning, obtains Adaboost graders.
Further, the step 2）Middle a3）The optimal Weak Classifier of the training comprises the following steps：
b1）Sample weights are normalized, and utilize formula：Sample weights are normalized；
b2）Using Haar-like features according to Pan and Zoom form calculus and all subcharacters are produced in sample window；
b3）By each class subcharacterIt is trained to a Weak Classifier, parameterRepresentative sample,Represent
Subcharacter,Majorization inequality direction is used to refer to,It is Weak Classifier threshold value, to subcharacterWhen training Weak Classifier,
First the corresponding characteristic value of all training samples is ranked up, by way of traversal, successively using each characteristic value as threshold value,
Error in classification when using this threshold value is calculated, finally compares the corresponding error amount of each threshold value, selects optimal threshold, so that
Train a Weak Classifier；
b4）Calculate the weighting fault rate of Weak Classifier：
；
b5）By comparing weighting fault rate, optimal Weak Classifier is selected from all Weak Classifiers,
It has the weighting fault rate of minimum；
b6）According to estimate of situation of this optimal Weak Classifier to sample, to readjust the weight of each sample：, whereinRepresent that sample is correctly classified,Represent by mistake classification,。
Further, the step 3）Described in the calculating of two graders based on SVM comprise the following steps：
C1 SVM input feature value) is calculated；In feature selecting, the color characteristic and textural characteristics of image are used；Wherein,
Color characteristic uses the second moment with variance characteristic and the third moment with gradient characteristic as color moment characteristics；Textural characteristics
Using the energy based on gray level co-occurrence matrixes, entropy, the moment of inertia, related 4 parametric textures, wherein gray level co-occurrence matrixes use 0 °,
45 °, 90 °, 135 ° of four angles finally calculate the average and variance of four parametric textures；Color characteristic and textural characteristics are melted
It is combined into 14 dimensional feature vectors comprising color characteristic and textural characteristics；
C2) according to step c1) choose each sample of feature extraction characteristic vector, composing training sample set, wherein,,；Generation
Space where table x is the real number space of n dimensions,There was only 1 and -1 liang of number in space where representing y；
C3) construct and solve convex quadratic programming problem：
It must solve：；
C4) basisCalculate：；
C5 decision function) is constructed, obtain two graders.
The present invention utilizes Adaboost graders and SVM classifier, constitutes the instrument alignment system of a two-stage, carries out instrument
Table is positioned.The one-time detection of instrument is carried out using the Adaboost graders of high detection rate, one or more instrument can be oriented
Table section；The reason for orienting multiple candidate regions has two：First, the high detection rate of grader will cause high false drop rate, the
Two, pending image usually contain complexity and meters as background.Accordingly, it would be desirable to increase on the basis of one-time detection
Plus secondary detection, using two graders based on SVM as the grader of secondary detection, secondary detection is carried out to candidate region,
It is final to determine meter location.To sum up, image to be detected carries out Primary Location first with Adaboost graders, and positioning is obtained
Multiple candidate regions reuse SVM classifier and carry out secondary positioning, the region for being determined as instrument twice is instrument place
The regional location of image.
The present invention orients the region similar to instrument using Adaboost graders, greatly improves the detection of instrument
Rate；The deficiency of Adaboost graders is made up using the graders of SVM bis-, color characteristic and textural characteristics has been merged, it is final accurate
Candidate region is classified；The scalability that verification and measurement ratio also increases system is not only increased, transformer station is met fixed to instrument
The requirement of position.The present invention realizes that instrument is positioned using grader, it is not necessary to provide instrument template, so simplifies artificial treatment
Flow improve the scalability of system again；Moreover, Detection results are not limited by meter size in picture to be detected, the party
Method can be accurately positioned the instrument of various resolution ratio in picture；Meanwhile, use the Adaboost graders and SVM point of high detection rate
Class device carries out two-stage classification detection, respectively using different types of characteristics of image, realizes manifold fusion and utilizes, improves
The locating accuracy of instrument；It is trained under grader is online, is not take up the time positioned in real time, can quickly realizes instrument
Table is positioned.
Brief description of the drawings
It is described in further detail below in conjunction with accompanying drawing and embodiments of the present invention
Fig. 1 is system block diagram of the invention；
Fig. 2 is Adaboost classifier training flow charts of the invention；
Fig. 3 is instrument positioning flow figure of the invention.
Embodiment
Referring to accompanying drawing.The present embodiment utilizes Adaboost graders and SVM classifier, the instrument positioning of one two-stage of composition
System, carries out instrument positioning.First, in order to train Adaboost graders, it is necessary to gather substantial amounts of sample, including instrument is only included
The positive sample picture of table dial plate and the negative sample picture for only including background.Then, to reduce the probability of instrument missing inspection, set higher
Recall rate, utilize the Adaboost graders of one high detection rate of these sample trainings；Wherein, training process includes feature
Selection, the training of Weak Classifier, the training of strong classifier.The grader of pending this high detection rate of imagery exploitation carries out instrument
The one-time detection of table, can orient one or more instrument regions；Secondary detection is used as using two graders based on SVM
Grader, carries out secondary detection to candidate region, finally determines meter location.
As shown in figure 1, Fig. 1 Block Diagrams left part is training process schematic diagram, right side is detection process schematic diagram.Training
Process contains the training of Adaboost graders and SVM classifier trains two parts, during detection, while passing through two classification
The window of device is considered as just window where instrument.Whole flow process fusion has used local gray level feature, color characteristic, texture special
Levy, therefore preferable positioning precision can be reached.
The overall flow of instrument positioning comprises the following steps：
1）Collect sample：Sample includes positive sample and negative sample, and positive sample is the image of the only dial plate containing instrument, and negative sample is not wrap
Background image containing instrument；Wherein positive sample needs to carry out dimension normalization, and negative sample need not be normalized, but its size
Bigger than normalized positive sample, positive and negative sample is both needed to be labeled.Positive sample quantity is tried one's best more than 1000, positive negative sample
Ratio 1：2.5 to 1：Between 3.
2）Adaboost classifier trainings：
Adaboost is a kind of iterative algorithm, and core is to be concatenated together different Weak Classifiers, constitutes a classifying quality
It is much better than the strong classifier of Weak Classifier.The sample used during training is constant, but sample weight according to the effect of classification without
It is disconnected to change, so it is considered as constantly being trained grader using new sample.Finally classifier stage is linked togather,
Form Decision Classfication device.This feature of Adaboost algorithm, is constantly focused on the crucial sample being difficult to differentiate between, and is realized
Excellent classifying quality.
Flow is trained as shown in Fig. 2 in Fig. 2For the frequency of training of strong classifier, when starting training, first willValue
0 is initialized as, after strong classifier is utilized every time to sample classification,Jia one with regard to progress to operate.N is the individual of optimal Weak Classifier
Number, when starting training, is initialized as 0 by N, trains every time after an optimal Weak Classifier, N value adds one.T is strong to constitute
The number of optimal Weak Classifier required for grader, is initialized as 5, when the error rate for the strong classifier that training is obtained can not reach
To default requirement when, T value adds one.For maximum training series, a larger number is could be arranged to, such as
It is set to 20.For the classification error rate of strong classifier,The highest classification error rate set for system, is traditionally arranged to be
0.001 or smaller value.
Specific training algorithm is as follows：
A1 training sample) is inputted, whereinRepresent that sample belongs to negative sample
This,Expression sample is positive sample, and n is sample size.
A2) initialization sample weight, if positive sample, then sets sample weights；If negative sample,
Then weight is set to, whereinWithThe number of positive negative sample is represented respectively.
a3）The optimal Weak Classifier of training；Including following training step：
b1）Sample weights are normalized, and utilize formula：Sample weights are normalized；
b2）Using Haar-like features according to Pan and Zoom form calculus and all subcharacters are produced in sample window；
b3）By each class subcharacterIt is trained to a Weak Classifier, parameterRepresentative sample,Represent
Subcharacter,Majorization inequality direction is used to refer to,It is Weak Classifier threshold value, to subcharacterWhen training Weak Classifier,
First the corresponding characteristic value of all training samples is ranked up, by way of traversal, successively using each characteristic value as threshold value,
Error in classification when using this threshold value is calculated, finally compares the corresponding error amount of each threshold value, selects optimal threshold, so that
Train a Weak Classifier；
b4）Calculate the weighting fault rate of Weak Classifier：
；
b5）By comparing weighting fault rate, optimal Weak Classifier is selected from all Weak Classifiers,
It has the weighting fault rate of minimum；
b6）According to estimate of situation of this optimal Weak Classifier to sample, to readjust the weight of each sample：, whereinRepresent that sample is correctly classified,Represent by mistake classification,。
a4）Weak Classifier is cascaded into strong classifier
。
a5）Classification and Identification is done to sample using strong classifier, and calculates weighting fault rate, if error rate is pre- more than reaching
If standard and frequency of training be less than default value, then using the sample of classification error as additional sample proceed train；
Otherwise deconditioning, obtains Adaboost graders.
3）Two graders based on SVM；Construction decision function is calculated using the color characteristic and textural characteristics of image, is obtained
To two graders；Calculating comprises the following steps：
C1 SVM input feature value) is calculated；In feature selecting, the color characteristic and textural characteristics of image are used；Wherein,
Color characteristic uses the second moment with variance characteristic and the third moment with gradient characteristic as color moment characteristics；Textural characteristics
Using the energy based on gray level co-occurrence matrixes, entropy, the moment of inertia, related 4 parametric textures, wherein gray level co-occurrence matrixes use 0 °,
45 °, 90 °, 135 ° of four angles finally calculate the average and variance of four parametric textures；Color characteristic and textural characteristics are melted
It is combined into 14 dimensional feature vectors comprising color characteristic and textural characteristics；
C2) according to step c1) choose each sample of feature extraction characteristic vector, composing training sample set, wherein,,；
C3) construct and solve convex quadratic programming problem：
It must solve：；Wherein K（）This function stand kernel function；What is represented is Lagrangian system
Number, is the coefficient introduced in extreme-value problem transfer process, is also the value for needing to solve in SVMs is solved,
It is to illustrate that they are to solve for obtained value with No. *；
C4) basisCalculate：；It is that SVMs needs another of solution
Coefficient, is to illustrate that they are to solve for obtained value with No. *；
C5 decision function) is constructed, obtain two graders.
4）Instrument is positioned：The probability of instrument is oriented while reducing the probability for navigating to non-instrument in order to improve.Use tool
The Adaboost graders for having high detection rate carry out one-level detection, and the multiple regions similar with instrument can be gone out with Primary Location.I.e.
One-level detection improves the probability for orienting instrument.Then secondary positioning is done using SVM classifier, makes up Adaboost graders
Deficiency, and then orient from candidate region instrument.
Instrument positioning flow is as shown in figure 3, it is the picture to be positioned containing instrument that picture is inputted in Fig. 3.Detection window
Size is dynamic change, and initial window size is, after the detection for often terminating a window, window
Size increases by 5 pixels.Represent the central pixel point of detection window, detection window since the upper left corner of image, according to
It is secondary to enter line slip detection in picture, both horizontally and vertically enter line slip by 2 pixels every time.
Specific position fixing process is as follows：
D1 picture to be detected) is inputted（Dimension of picture is）, and image is pre-processed；Turn including cromogram at gray-scale map
Reason, medium filtering and gaussian filtering process；
D2 detection window size) is initialized；Window selection uses width and height identical size, and by detection window width
It is initialized as 20 pixels；
D3) it is using sizeDetection window, since initial position, with two pixels of horizontal direction and Vertical Square
Translated to two pixels for unit on whole pictures, often translate and can once obtain a new detection window；
D4) for a detection window, corresponding all Haar-like subcharacters under the window are calculated；
D5 classification judgement) is carried out to the image-region covered under detection window using Adaboost graders, if being judged as instrument
Table, then record position and the size of the window；
D6) adjustment window size is：If met, then into step d7), otherwise enter step
Rapid d3)；
D7 detection window) is merged；The size of detection window is different when being positioned because of instrument, can be detected near same position many
Individual region, therefore the window of analogous location is merged, it is to avoid repeat；
D8 all candidate's instrument regions) are extracted, and size normalization is carried out to it；
D9 classification judgement, output result) are carried out to candidate region using SVM classifier；
D10) output meter is in the position of image.
Claims (6)
1. a kind of instrument localization method suitable for Intelligent Mobile Robot, it is characterised in that：Comprise the following steps：
1）Collect sample：Sample includes positive sample and negative sample, and positive sample is the image of the only dial plate containing instrument, and negative sample is not wrap
Background image containing instrument；
2）Adaboost classifier trainings：Utilize step 1）The sample of collection is trained, selection of the training process including feature,
The training of Weak Classifier, the training of strong classifier；The sample used during training is constant, but sample weight according to the effect of classification
And constantly change, so it is considered as constantly being trained grader using new sample；Finally classifier stage is associated in
Together, Decision Classfication device is formed；
3）Two graders based on SVM；Construction decision function is calculated using the color characteristic and textural characteristics of image, two are obtained
Grader；
4）Instrument is positioned：One-level detection is carried out first by the Adaboost graders with high detection rate, can be with Primary Location
Go out the multiple regions similar with instrument；That is one-level detection improves the probability for orienting instrument；Then two are done using SVM classifier
Secondary positioning, makes up the deficiency of Adaboost graders, and then orients from candidate region instrument.
2. a kind of instrument localization method suitable for Intelligent Mobile Robot as claimed in claim 1, it is characterised in that：Institute
State step 4）Described instrument localization method specifically includes following steps：
D1 picture to be detected) is inputted（Dimension of picture is）, and image is pre-processed；Turn including cromogram at gray-scale map
Reason, medium filtering and gaussian filtering process；
D2 detection window size) is initialized；Window selection uses width and height identical size, and by detection window width
It is initialized as 20 pixels；
D3) it is using sizeDetection window, since initial position, with two pixels of horizontal direction and vertical direction
Two pixels are translated for unit on whole pictures, are often translated and can once obtain a new detection window；
D4) for a detection window, corresponding all Haar-like subcharacters under the window are calculated；
D5 classification judgement) is carried out to the image-region covered under detection window using Adaboost graders, if being judged as instrument
Table, then record position and the size of the window；
D6) adjustment window size is：If met, then into step d7), otherwise enter step
Rapid d3)；
D7 detection window) is merged；The size of detection window is different when being positioned because of instrument, can be detected near same position many
Individual region, therefore the window of analogous location is merged, it is to avoid repeat；
D8 all candidate's instrument regions) are extracted, and size normalization is carried out to it；
D9 classification judgement, output result) are carried out to candidate region using SVM classifier；
D10) output meter is in the position of image.
3. a kind of instrument localization method suitable for Intelligent Mobile Robot as claimed in claim 1, it is characterised in that：Institute
State step 1）Described in positive sample need to carry out dimension normalization, negative sample need not be normalized, but its size than
Normalized positive sample is big, and positive and negative samples are both needed to be labeled；The ratio of positive and negative samples is 1：2.5~1：Between 3.
4. a kind of instrument localization method suitable for Intelligent Mobile Robot as claimed in claim 1, it is characterised in that：Institute
State step 2）The specific training algorithm of described Adaboost classifier trainings comprises the following steps：
A1 training sample) is inputted, whereinRepresent that sample belongs to negative sample
This,Expression sample is positive sample, and n is sample size；
A2) initialization sample weight, if positive sample, then sets sample weights；If negative sample, then weigh
Reset and be set to, whereinWithThe number of positive negative sample is represented respectively；
a3）The optimal Weak Classifier of training；
a4）Weak Classifier is cascaded into strong classifier
；
a5）Classification and Identification is done to sample using strong classifier, and calculates weighting fault rate, if error rate is default more than reaching
Standard and frequency of training are less than default value, then proceed the sample of classification error as additional sample to train；Otherwise
Deconditioning, obtains Adaboost graders.
5. a kind of instrument localization method suitable for Intelligent Mobile Robot as claimed in claim 4, it is characterised in that：Institute
State step a3）Described in training optimal Weak Classifier comprise the following steps：
b1）Sample weights are normalized, and utilize formula：Sample weights are normalized；
b2）Using Haar-like features according to Pan and Zoom form calculus and all subcharacters are produced in sample window；
b3）By each class subcharacterIt is trained to a Weak Classifier, parameterRepresentative sample,Represent
Subcharacter,Majorization inequality direction is used to refer to,It is Weak Classifier threshold value, to subcharacterWhen training Weak Classifier,
First the corresponding characteristic value of all training samples is ranked up, by way of traversal, successively using each characteristic value as threshold value,
Error in classification when using this threshold value is calculated, finally compares the corresponding error amount of each threshold value, selects optimal threshold, so that
Train a Weak Classifier；
b4）Calculate the weighting fault rate of Weak Classifier：
；
b5）By comparing weighting fault rate, optimal Weak Classifier is selected from all Weak Classifiers,
It has the weighting fault rate of minimum；
b6）According to estimate of situation of this optimal Weak Classifier to sample, to readjust the weight of each sample：, whereinRepresent that sample is correctly classified,Represent by mistake classification,。
6. a kind of instrument localization method suitable for Intelligent Mobile Robot as claimed in claim 1, it is characterised in that：Institute
State step 3）Described in the calculating of two graders based on SVM comprise the following steps：
C1 SVM input feature value) is calculated；In feature selecting, the color characteristic and textural characteristics of image are used；Wherein,
Color characteristic uses the second moment with variance characteristic and the third moment with gradient characteristic as color moment characteristics；Textural characteristics
Using the energy based on gray level co-occurrence matrixes, entropy, the moment of inertia, related 4 parametric textures, wherein gray level co-occurrence matrixes use 0 °,
45 °, 90 °, 135 ° of four angles finally calculate the average and variance of four parametric textures；Color characteristic and textural characteristics are melted
It is combined into 14 dimensional feature vectors comprising color characteristic and textural characteristics；
C2) according to step c1) choose each sample of feature extraction characteristic vector, composing training sample set, wherein,,；
C3) construct and solve convex quadratic programming problem：
It must solve：；
C4) basisCalculate：；
C5 decision function) is constructed, obtain two graders.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710239499.5A CN106951930A (en) | 2017-04-13 | 2017-04-13 | A kind of instrument localization method suitable for Intelligent Mobile Robot |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710239499.5A CN106951930A (en) | 2017-04-13 | 2017-04-13 | A kind of instrument localization method suitable for Intelligent Mobile Robot |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106951930A true CN106951930A (en) | 2017-07-14 |
Family
ID=59475916
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710239499.5A Pending CN106951930A (en) | 2017-04-13 | 2017-04-13 | A kind of instrument localization method suitable for Intelligent Mobile Robot |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106951930A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108422432A (en) * | 2018-05-07 | 2018-08-21 | 哈工大（张家口）工业技术研究院 | A kind of crusing robot and the underground pipe gallery information management system with the robot |
| CN108582107A (en) * | 2018-05-07 | 2018-09-28 | 哈工大（张家口）工业技术研究院 | A kind of pipe gallery information system based on technology of Internet of things |
| CN108614482A (en) * | 2018-05-07 | 2018-10-02 | 哈工大（张家口）工业技术研究院 | A kind of underground pipe gallery information system based on BIM, GIS and IOT |
| CN109255336A (en) * | 2018-09-29 | 2019-01-22 | 南京理工大学 | Arrester recognition methods based on crusing robot |
| CN109344766A (en) * | 2018-09-29 | 2019-02-15 | 南京理工大学 | Slide block type breaker recognition methods based on crusing robot |
| CN109344768A (en) * | 2018-09-29 | 2019-02-15 | 南京理工大学 | Pointer breaker recognition methods based on crusing robot |
| CN109360289A (en) * | 2018-09-29 | 2019-02-19 | 南京理工大学 | Merge the electric power meter detection method of crusing robot location information |
| CN109359646A (en) * | 2018-09-29 | 2019-02-19 | 南京理工大学 | Liquid level type Meter recognition method based on crusing robot |
| CN109389165A (en) * | 2018-09-29 | 2019-02-26 | 南京理工大学 | Oil level gauge for transformer recognition methods based on crusing robot |
| CN109447061A (en) * | 2018-09-29 | 2019-03-08 | 南京理工大学 | Reactor oil level indicator recognition methods based on crusing robot |
| CN109447949A (en) * | 2018-09-29 | 2019-03-08 | 南京理工大学 | Insulated terminal defect identification method based on crusing robot |
| CN109446916A (en) * | 2018-09-29 | 2019-03-08 | 南京理工大学 | Discharge counter recognition methods based on crusing robot |
| CN109447062A (en) * | 2018-09-29 | 2019-03-08 | 南京理工大学 | Pointer-type gauges recognition methods based on crusing robot |
| CN109709541A (en) * | 2018-12-26 | 2019-05-03 | 杭州奥腾电子股份有限公司 | A kind of vehicle environment perception emerging system target erroneous detection processing method |
| CN111523592A (en) * | 2020-04-21 | 2020-08-11 | 易拍全球(北京)科贸有限公司 | Historical relic artwork field image similarity measurement algorithm based on deep learning |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102426653A (en) * | 2011-10-28 | 2012-04-25 | 西安电子科技大学 | Static human body detection method based on second generation Bandelet transformation and star type model |
| CN105005768A (en) * | 2015-07-06 | 2015-10-28 | 河海大学 | Dynamic percentage sample cutting AdaBoost human face detection algorithm |
| CN105184244A (en) * | 2015-08-28 | 2015-12-23 | 深圳英飞拓科技股份有限公司 | Video face detection method and apparatus |
| CN105447511A (en) * | 2015-11-15 | 2016-03-30 | 安徽清新互联信息科技有限公司 | SVM object detection method based on Adaboost Haar-Like features |
| CN105844285A (en) * | 2016-02-25 | 2016-08-10 | 中国农业大学 | Cucumber disease identification method and apparatus based on image information |
| CN105975979A (en) * | 2016-04-22 | 2016-09-28 | 浙江大学 | Instrument detection method based on machine vision |
| CN106503615A (en) * | 2016-09-20 | 2017-03-15 | 北京工业大学 | Indoor human body detecting and tracking and identification system based on multisensor |
-
2017
- 2017-04-13 CN CN201710239499.5A patent/CN106951930A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102426653A (en) * | 2011-10-28 | 2012-04-25 | 西安电子科技大学 | Static human body detection method based on second generation Bandelet transformation and star type model |
| CN105005768A (en) * | 2015-07-06 | 2015-10-28 | 河海大学 | Dynamic percentage sample cutting AdaBoost human face detection algorithm |
| CN105184244A (en) * | 2015-08-28 | 2015-12-23 | 深圳英飞拓科技股份有限公司 | Video face detection method and apparatus |
| CN105447511A (en) * | 2015-11-15 | 2016-03-30 | 安徽清新互联信息科技有限公司 | SVM object detection method based on Adaboost Haar-Like features |
| CN105844285A (en) * | 2016-02-25 | 2016-08-10 | 中国农业大学 | Cucumber disease identification method and apparatus based on image information |
| CN105975979A (en) * | 2016-04-22 | 2016-09-28 | 浙江大学 | Instrument detection method based on machine vision |
| CN106503615A (en) * | 2016-09-20 | 2017-03-15 | 北京工业大学 | Indoor human body detecting and tracking and identification system based on multisensor |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108422432A (en) * | 2018-05-07 | 2018-08-21 | 哈工大（张家口）工业技术研究院 | A kind of crusing robot and the underground pipe gallery information management system with the robot |
| CN108582107A (en) * | 2018-05-07 | 2018-09-28 | 哈工大（张家口）工业技术研究院 | A kind of pipe gallery information system based on technology of Internet of things |
| CN108614482A (en) * | 2018-05-07 | 2018-10-02 | 哈工大（张家口）工业技术研究院 | A kind of underground pipe gallery information system based on BIM, GIS and IOT |
| CN109389165A (en) * | 2018-09-29 | 2019-02-26 | 南京理工大学 | Oil level gauge for transformer recognition methods based on crusing robot |
| CN109447949A (en) * | 2018-09-29 | 2019-03-08 | 南京理工大学 | Insulated terminal defect identification method based on crusing robot |
| CN109344768A (en) * | 2018-09-29 | 2019-02-15 | 南京理工大学 | Pointer breaker recognition methods based on crusing robot |
| CN109360289A (en) * | 2018-09-29 | 2019-02-19 | 南京理工大学 | Merge the electric power meter detection method of crusing robot location information |
| CN109359646A (en) * | 2018-09-29 | 2019-02-19 | 南京理工大学 | Liquid level type Meter recognition method based on crusing robot |
| CN109255336A (en) * | 2018-09-29 | 2019-01-22 | 南京理工大学 | Arrester recognition methods based on crusing robot |
| CN109447061A (en) * | 2018-09-29 | 2019-03-08 | 南京理工大学 | Reactor oil level indicator recognition methods based on crusing robot |
| CN109344766A (en) * | 2018-09-29 | 2019-02-15 | 南京理工大学 | Slide block type breaker recognition methods based on crusing robot |
| CN109446916A (en) * | 2018-09-29 | 2019-03-08 | 南京理工大学 | Discharge counter recognition methods based on crusing robot |
| CN109447062A (en) * | 2018-09-29 | 2019-03-08 | 南京理工大学 | Pointer-type gauges recognition methods based on crusing robot |
| CN109360289B (en) * | 2018-09-29 | 2021-09-28 | 南京理工大学 | Power meter detection method fusing inspection robot positioning information |
| CN109709541A (en) * | 2018-12-26 | 2019-05-03 | 杭州奥腾电子股份有限公司 | A kind of vehicle environment perception emerging system target erroneous detection processing method |
| CN111523592A (en) * | 2020-04-21 | 2020-08-11 | 易拍全球(北京)科贸有限公司 | Historical relic artwork field image similarity measurement algorithm based on deep learning |
| CN111523592B (en) * | 2020-04-21 | 2023-05-09 | 易元数字(北京)科技集团有限公司 | Image similarity measurement method based on deep learning in cultural relic artwork field |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106951930A (en) | A kind of instrument localization method suitable for Intelligent Mobile Robot | |
| Hu et al. | Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries | |
| Fu et al. | Hyperspectral image super-resolution with optimized RGB guidance | |
| JP7523711B2 (en) | Image processing device and image processing method | |
| CN110660023B (en) | Video stitching method based on image semantic segmentation | |
| CN108596026B (en) | Cross-view gait recognition device and training method based on double-flow generation countermeasure network | |
| CN106875381A (en) | A kind of phone housing defect inspection method based on deep learning | |
| CN111461134A (en) | Low-resolution license plate recognition method based on generation countermeasure network | |
| CN106874826A (en) | Face key point-tracking method and device | |
| CN106203391A (en) | Face identification method based on intelligent glasses | |
| CN113159043B (en) | Feature point matching method and system based on semantic information | |
| CN108280411A (en) | A kind of pedestrian's searching method with spatial alternation ability | |
| Zeng et al. | An evaluation system for building footprint extraction from remotely sensed data | |
| CN109635661B (en) | Far-field wireless charging receiving target detection method based on convolutional neural network | |
| CN108334955A (en) | Copy of ID Card detection method based on Faster-RCNN | |
| CN112288758B (en) | Infrared and visible light image registration method for power equipment | |
| CN109308715A (en) | A kind of optical imagery method for registering combined based on point feature and line feature | |
| CN111242144A (en) | Method and device for detecting abnormality of power grid equipment | |
| CN112634125A (en) | Automatic face replacement method based on off-line face database | |
| CN107589093A (en) | A kind of ature of coal on-line checking analysis method based on regression analysis | |
| CN114241372A (en) | Target identification method applied to sector-scan splicing | |
| CN111461162A (en) | Zero-sample target detection model and establishing method thereof | |
| CN109523590A (en) | A kind of 3D rendering depth information visual comfort appraisal procedure based on sample | |
| Jingyi et al. | Notice of violation of IEEE publication principles: ground-based cloud image recognition system based on multi-CNN and feature screening and fusion | |
| CN110910497A (en) | Method and system for realizing augmented reality map |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170714 |
|
| WD01 | Invention patent application deemed withdrawn after publication |