CN104200228B - Recognizing method and system for safety belt - Google Patents
Recognizing method and system for safety belt Download PDFInfo
- Publication number
- CN104200228B CN104200228B CN201410442594.1A CN201410442594A CN104200228B CN 104200228 B CN104200228 B CN 104200228B CN 201410442594 A CN201410442594 A CN 201410442594A CN 104200228 B CN104200228 B CN 104200228B
- Authority
- CN
- China
- Prior art keywords
- image
- image block
- sample
- feature
- training
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 59
- 239000013598 vector Substances 0.000 claims abstract description 135
- 238000012549 training Methods 0.000 claims abstract description 103
- 230000013016 learning Effects 0.000 claims abstract description 40
- 230000008569 process Effects 0.000 claims abstract description 13
- 238000000605 extraction Methods 0.000 claims description 29
- 239000011159 matrix material Substances 0.000 claims description 29
- 239000000284 extract Substances 0.000 claims description 23
- 239000011521 glass Substances 0.000 claims description 12
- 239000000203 mixture Substances 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 11
- 230000004044 response Effects 0.000 claims description 10
- 238000005457 optimization Methods 0.000 claims description 7
- 230000006870 function Effects 0.000 claims description 6
- 239000012141 concentrate Substances 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 5
- 238000013507 mapping Methods 0.000 claims description 3
- 239000000463 material Substances 0.000 claims description 3
- 230000001629 suppression Effects 0.000 claims description 3
- 238000001514 detection method Methods 0.000 abstract description 11
- 238000005286 illumination Methods 0.000 abstract description 4
- 238000012706 support-vector machine Methods 0.000 abstract 2
- 230000002411 adverse Effects 0.000 abstract 1
- 238000005070 sampling Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000013075 data extraction Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000012937 correction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a method for detecting and recognizing the tying of a safety belt of a driver based on interval-maximizing and multi-sample dictionary learning. The method comprises the following steps: treating a driver-containing data-concentrated part of a windshield of a left half car as an image to be processed for input; acquiring a feature vector representation of each image; performing dictionary learning process for a training data set by an interval-maximizing and multi-example manner; respectively training an SVM (Support Vector Machine) multi-classifier for each type in the training data set after clustering to obtain a classifying model; coding the training data through the acquired dictionary; training the classifier through the coded feature vectors; detecting and recognizing an image to be recognized through the trained classifier, so as to determine whether the safety belt is tied from the image to be recognized. According to the recognizing method and system for the safety belt, the safety belt can be simply and easily detected and recognized, the popularization capacity is high, the detection and recognition accuracy is high, and the speed is fast; in addition, the influences of illumination and noise and other adverse factors can be effectively overcome.
Description
Technical field
The invention belongs to technical field of computer vision, more particularly, to a kind of many example words of margin maximization are based on
The seat belt recognition methodss of allusion quotation study and system.
Background technology
As the continuous development of global economy situation, the living standard of people are increasingly improved, individual possesses motor vehicles
Quantity be in geometric growth situation, the popularization of automobile becomes current inexorable trend.What is brought therewith is the prison of automotive safety
Superintend and direct and management work, wearing for pilot harness is the work for primarily needing supervision and management, therefore study driver safety
The detection and identification work of band seems particularly urgent and important.The research for detecting to seat belt at present and recognizing also is not very complete
Kind, mainly by the simple image processing algorithm such as rim detection, accuracy rate is not also very to seat belt detection algorithm before
It is high.Therefore want to obtain higher accuracy rate and faster Detection results have to binding pattern identification, computer vision etc.
Correlation technique, it is desirable to first the car windshield approximate location of automobile in motion can be extracted and identified from complicated background
Come, detect whether driver has band peace by correlation techniques such as Image semantic classification, feature extraction, dictionary learning, training graders
Full band.
But also there are many difficult problems in seat belt detection at present, for example, the shadow of angle, light during image taking etc.
Ring, the impact of driver's dressing color and seat belt color contrast difference.These factors all can be come not to seat belt detection band
The impact of profit.
The content of the invention
It is an object of the invention to provide a kind of seat belt recognition methodss based on many example dictionary learnings of margin maximization,
The method identification process is simple and accuracy rate is higher.
For achieving the above object, according to one aspect of the present invention, there is provided one kind is based on many example words of margin maximization
The seat belt recognition methodss of allusion quotation study, comprise the steps：
(1) obtaining training data concentrates the characteristic vector of car bumper wind glass right half part image to represent；
(1.1) the car windshield image concentrated to training image, intercepts first the right half part image of car windshield,
The image block of different scale is extracted to car windshield right half part；
The car windshield image is the image shot to driver's cabin from vehicle frontal i.e. car windshield, and it right half
Include driver in parts of images.
Specifically, the right half part of car windshield for intercepting being obtained, respectively with three kinds of scale sizes (such as 48*48,
72*72,96*96) sliding window intercept in the enterprising line slip of image, so finally give the image block of several different scalesWherein N1、N2、N3Three kinds of different scale hypograph blocks are represented respectively
Number, N=N1+N2+N3Represent the number of total image block.
(1.2) to the right half part image zooming-out SIFT feature vector of car windshield, and using default code book pair
SIFT feature is encoded；
The default code book can be by obtaining to carrying out cluster to the SIFT feature vector that training data extraction is obtained.
(1.3) image block of the different scale obtained in step (1.1) is zoomed to into same scale size, and to scaling after
Each image block extract characteristic vector respectively, the characteristic vector of described image block be by image block HOG characteristic vectors,
LBP characteristic vectors, SIFT feature vector composition；
(1.3.1) the HOG characteristic vectors of image block are calculated：
Specifically, seeking each pixel gradient (including size and Orientation), minicell unit is divided an image into, counted
Several cell factories (such as 3*3) are constituted a region unit by the histogram of gradients of each cell factory, will be thin in each region unit
The histogram of gradients of born of the same parents' unit is stitched together the HOG Feature Descriptors as the region unit, finally by all of zones of different block
Characteristic vector is coupled together, and the final HOG characteristic vectors for obtaining input picture block are represented；
(1.3.2) the LBP characteristic vectors of image block are calculated：
Specifically, first by the minicell unit that image division size is 16*16, the LBP for calculating each cell factory is special
Levy, finally the LBP rectangular histograms of each minicell unit are stitched together obtains the LBP characteristic vectors of image block and represent；
(1.3.3) the coding SIFT feature of image block is calculated
Specifically, extracting SIFT feature coordinate is located at the feature in the image block, using bag of words (bag of
Words) and with reference to the code book predicted these features are encoded, the rectangular histogram that coding is obtained is normalized and is obtained image
The coding SIFT feature of block represents sift, and size C of code book is preset value.
(1.3.4) characteristic vector for obtaining car windshield image is finally represented
Specifically, first to each image block, by the HOG features obtained in above-mentioned (1.3.1) and (1.3.2) step and
It is stitched together with the coding SIFT feature that obtains in step (1.3.3) after the difference normalization of LBP features, as image block
Character representation Fi=[hogi；lbpi；sifti], 1≤i≤N, wherein i represent i-th image block, and N represents image block sum；Most
Afterwards by the merging features of all image blocks together
(2) based on margin maximization multiclass example dictionary learning
(2.1) the every training image concentrated to training image, being divided into training dataset using artificial mark has safety
The sample image of band and the sample image without seat belt, using the class image block sample in above-mentioned two classes image as positive sample
This, then another kind of image block is used as negative sample；
(2.2) to the positive sample and negative sample in step (2.1), based on margin maximization method, the power of positive sample is calculated
Weight matrix；Specifically, first to the image in positive sample carry out clustering the label z for obtaining clusteringi, according to sample weight and given ratio
Example extracts positive sample, and all negative samples constitute training dataset, solve optimization problem by multiclass SVM and try to achieve weight square
Battle array, by continuous iteration undated parameter the weight matrix W of optimum is obtained；Implement step as follows：
(2.2.1) to the image block sample F in positive sampleiClustered, if K represents the middle calculation of cluster, and according to poly-
Class result gives a label z to each image block samplei∈ { 0,1 ..., K }, if F after clusteriBelong to kth class, then zi=k
∈{1,…,K}；If FiBelong to negative class, then zi=0；For image block sample its label in all negative samples is zi=0；
(2.2.2) weight matrix W=[w are defined0,w1,…,wK], k ∈ { 0,1 ..., K }, wherein wkRepresent k-th positive class
Clustering Model, w0The Clustering Model of negative class is represented, K represents cluster numbers；Define sample weight pij, pijIt is to ensure from training
The data extracted in data are positive sample, according to SVM decision functions to characteristic vector FiBelong between positive class and negative class most
It is poor greatly to be given, i.e.,Define psTo extract the positive sample in training data in proportion；Generally ps
For fixed value, such as value is 0.7；
(2.2.3) under original state, if sample weight p of all positive classesijIt is 1.First according to sample weight pijExtract just
Sample, then extracts p from the sample for selectingsRatio, and all of negative sample composition training dataset D', wherein psExpression is taken out
The ratio of sampling example, D' represents the training dataset of new composition, because cluster labels there is known, can be solved by multiclass SVM
Following optimization problems are obtaining W：
WhereinrijRepresent in addition to this class, xijThe maximum of remaining class is rung
Should；
(2.2.4) sample weight and sample label are updated, obtains weight matrix to update sample weight with step (2.2.2)
pij；According toTo update the label of sample；
(2.2.5) repeat (2.2.3)～(2.2.4) processes, circulate M time, wherein M is default cycle-index, for example, take
It is worth for 5.
(2.3) positive sample and negative sample in step (2.1) is interchangeable, is calculated using the method in step (2.2)
The weight matrix of the positive sample obtained after exchange, the weight matrix that will be obtained in the weight matrix for obtaining here and step (2.2)
It is stitched together and obtains based on many example dictionaries of margin maximization.
(3) grader of training of safety band identification
(3.1) concentrate car bumper wind glass right the training data obtained in step (1) using the dictionary obtained in step (2)
The characteristic vector of half part carries out encoding the coding characteristic for obtaining image vector；Specific coding mode is：
(3.1.1) image feature vector is multiplied with weight matrix, is obtained for the response of each disaggregated model maps.
(3.1.2) to each response mapping, response of the coordinate within image range is counted, and carries out non-maximum suppression
System, obtains the vector of a H dimension；
(3.1.3) utilization space pyramid model, is divided into 2*2 and 4*4 totally 20 blocks, for each block by former training image
Repeat the calculating process of (3.1.1) and (3.1.2), finally obtain the characteristic vector of 20 H dimensions；
(3.1.4) vector that above-mentioned two step is obtained is stitched together, obtains the vector of 21*H dimensions, be exactly the image
Final characteristic vector；
(3.2) the coding characteristic vector training SVM classifier of all images for obtaining is concentrated using training data, herein may be used
To be considered as a black box using packaged existing SVM classifier built-in function, the characteristic vector of all training images is input into,
Output one is capable of deciding whether the SVM classifier with seat belt.
(4) car windshield image to be identified is recognized
(4.1) according to the method in step (1) to image zooming-out characteristic vector to be identified；
(4.2) images to be recognized has been obtained spy according to the method encoded to training data in step (3.1)
Levy vector to be encoded；
(4.3) using the SVM classifier trained in step (3), image to be identified is classified, and output safety
Result with identification.
The invention discloses a kind of seat belt recognition methodss based on many example dictionary learnings of margin maximization.For user
A given width car windshield image, the present invention can be intercepted to the car bumper wind glassy zone in image, to car bumper wind
Glass driver part carries out seat belt identification.
The car windshield image concentrated to training data, obtains first the different images block for representing image, to each figure
As block extracts respectively feature, and the character representation as image that is stitched together, then margin maximization is based on the feature of image
The process of many example dictionary learnings, the dictionary obtained with study is encoded to training data feature, finally with the spy after coding
Levy training SVM classifier.
To car windshield image to be identified, the car windshield image processing method class concentrated with training data first
Seemingly, first, the image block of different scale is obtained to each car windshield image to be identified, each image block is extracted respectively
Feature, and the character representation as image that is stitched together, the dictionary for then being obtained using training process is to car bumper wind to be identified
Glass characteristics of image is encoded, and the grader for finally being obtained with training is identified to image to be identified, obtains to be identified
The recognition result of car windshield image.
It is another aspect of this invention to provide that additionally providing a kind of seat belt based on many example dictionary learnings of margin maximization
Identifying system, the system includes car windshield image feature vector generation module, many example dictionary learnings of margin maximization
Module, classifier training module and car windshield picture recognition module to be identified, wherein：
The car windshield right half part image feature vector generation module, for obtaining training image data set in
The characteristic vector of car windshield right half part image represents, specifically includes different scale images block acquisition module, car bumper wind glass
The representation module of glass right half part image SIFT feature extraction module, image block characteristics extraction module and characteristics of image, wherein：
The different scale images block acquisition module, for the car windshield image concentrated to training image, carries first
The right half part of pick-up windshield, to right half part car windshield the image block under different scale is obtained；
The car windshield SIFT feature extraction module, for car windshield image zooming-out SIFT feature and sharp
The process encoded with default code book；
Described image block feature extraction module, for extracting characteristic vector to the image block of the different scale for acquiring.
Specifically include image block scaling submodule, HOG characteristic vector calculating sub modules, LBP characteristic vectors calculating sub module and coding
SIFT feature histogram calculation submodule.
Described image block scales submodule, for zooming to same yardstick to the image block of different scale, generally by image
Zoom to the smallest dimension in three kinds of yardsticks；
The HOG characteristic vectors calculating sub module, for calculating the image block that image block scaling submodule is acquired
HOG characteristic vectors；
The LBP characteristic vectors calculating sub module, for calculating the image block that image block scaling submodule is acquired
LBP characteristic vectors；
The coding SIFT feature histogram calculation submodule, for calculating image block scaling submodule image is acquired
The SIFT histogram features vector of block；
The character representation module of described image, for calculating the character representation of car windshield image, to every image, will
The merging features for representing all image blocks of image act the character representation for being used as image；
The margin maximization many example dictionary learning modules, for learning encoder dictionary, for training data is concentrated
Each class, is respectively trained a multi classifier, obtains a disaggregated model, and the weight in all of sorter model is merged
Get up to obtain dictionary to represent；
The classifier training module, for training of safety band recognition classifier, carries out coding mould to training data first
Block, the disaggregated model obtained using many example dictionary learning modules of margin maximization is encoded to training data；Finally utilize
The coding characteristic vector training SVM classifier of all images obtained in training image；
The car windshield picture recognition module to be identified, for recognizing whether images to be recognized has seat belt, specifically
Extraction module, the SIFT feature extraction module of car windshield image including car windshield image block to be identified, image block
Characteristic extracting module, image feature representation module, characteristics of image coding module, car windshield picture recognition module to be identified,
Wherein：
The car windshield image block extraction module to be identified, for obtaining the figure of car windshield right half part image
As block is represented；
The SIFT feature extraction module of the car windshield image, for obtaining car windshield right half part image
SIFT feature, and it is encoded with known code book；
Described image block feature extraction module, for extracting the characteristic vector for representing image block；
Described image character representation module, for the image block characteristics of every piece image to be stitched together as the spy of image
Levy vector；
Described image feature coding module, for the dictionary that obtained using margin maximization multi-instance learning to characteristics of image
Encoded；
The car windshield picture recognition module to be identified, for the seat belt trained using classifier training module
Recognition classifier, classifies to image to be identified, and the recognition result of output safety band.
By the contemplated above technical scheme of the present invention, compared with prior art, the present invention has following technique effect：
1. car windshield image is easily subject to illumination, camera resolution and shooting angle and motorist's clothing and safety
Impact with factors such as color contrasts and becoming is not easy identification, and current algorithm mainly uses image procossing related algorithm
The means such as rim detection, straight-line detection are carried out, these algorithms cannot adapt to the change of external environment, it is impossible in various complex scenes
Under high identification is all provided.The inventive method is stitched together expression image using various features so that the mark sheet of image
Show with characteristics such as anti-rotation, anti-noise, anti-light photographs.Therefore, the inventive method can effectively overcome illumination, shooting angle change etc.
The impact of factor；
2. the inventive method employs many example dictionary learning methods of margin maximization, each class concentrated to training data
First by cluster, many example dictionary learnings of margin maximization obtain a disaggregated model to image block, the classification mould for so obtaining
Type there is highest to respond the similar image.Therefore, to obtain higher identification to car windshield accurate for the inventive method
Rate；
3. the method for the present invention divides the image into the image pyramid for 1*1,2*2,4*4；The feature of whole image is by every
The merging features of sub-regions get up；Therefore the feature of whole image includes certain spatial positional information；Obtain after training
Model, the weight comprising every sub-regions, according to the size of weight, this method energy indirect selections go out representational region；
Therefore the present invention need not carry out the concrete demarcation of specific belt position, reduce the work of artificial participation.
Description of the drawings
Fig. 1 is seat belt recognition methodss flow chart of the present invention based on many example dictionary learnings of margin maximization.
Specific embodiment
In order that the objects, technical solutions and advantages of the present invention become more apparent, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, and
It is not used in the restriction present invention.As long as additionally, technical characteristic involved in invention described below each embodiment
Not constituting conflict each other just can be mutually combined.
As shown in figure 1, seat belt detection and recognition methodss bag of the present invention based on many example dictionary learnings of margin maximization
Include following steps：
(1) obtaining training data concentrates the characteristic vector of car bumper wind glass right half part image to represent
(1.1) the car windshield image concentrated to training image, intercepts first the right half part image of car windshield,
The image block of different scale is extracted to car windshield right half part.
The car windshield image is the image shot to driver's cabin from vehicle frontal i.e. car windshield, and it right half
Include driver in parts of images.
Specifically, the image of the right side half of car windshield for intercepting being obtained, respectively with three kinds of scale sizes (such as 48*48,
72*72,96*96) sliding window intercept in the enterprising line slip of image, sliding window moving step length can be 1/3 length of window
Degree, the picture material in sliding window each time is intercepted and is preserved, and so finally gives the image of several different scales
BlockWherein N1、N2、N3Three kinds of different scale hypographs are represented respectively
The number of block, N=N1+N2+N3Represent the number of total image block.
(1.2) (Scale-invariant is changed to the right half part image zooming-out scale invariant feature of car windshield
Feature transform, SIFT) characteristic vector, and SIFT feature is encoded using default code book；
The default code book can be by obtaining to carrying out cluster to the SIFT feature vector that training data extraction is obtained.
(1.3) image block of the different scale obtained in step (1.1) is zoomed to into same scale size, and to scaling after
Each image block extract characteristic vector respectively, the characteristic vector of described image block be by image block HOG characteristic vectors,
LBP characteristic vectors, SIFT feature vector composition；
Then it is straight with SIFT after HOG the and LBP features difference normalization that above-mentioned steps are obtained to each image block
Square figure merging features as the characteristic vector of image block, are obtained finally the merging features by of all image blocks together together
The characteristic vector of whole image；
Specifically, the image block of different scale can be zoomed to one kind in step (1.1) in multiple yardsticks, general contracting
It is put into smallest dimension.
(1.3.1) histograms of oriented gradients (the Histogram of oriented of the image block after scaling are calculated
Gradient, HOG) characteristic vector：
Specifically, seeking each pixel gradient (including size and Orientation), minicell unit is divided an image into, counted
Several cell factories (such as 3*3) are constituted a region unit by the histogram of gradients of each cell factory, will be thin in each region unit
The histogram of gradients of born of the same parents' unit is stitched together the HOG Feature Descriptors as the region unit, finally by all of zones of different block
Characteristic vector is coupled together, and the final HOG characteristic vectors for obtaining input picture block are represented；
In embodiments of the present invention, first image gray processing is carried out into color space using Gamma correction methods to image
Standardization, it is therefore an objective to adjust the contrast of image, reduces the impact caused by the shade and illumination variation of image local, while can
To suppress the interference of noise.Gradient (including size and Orientation) is sought each pixel, minicell unit is divided an image into, is united
The histogram of gradients of each cell factory is counted, several cell factories (such as 3*3) region unit is constituted into, by each region unit
The histogram of gradients of cell factory is stitched together the HOG Feature Descriptor hog as the region unitk=[h1,h2,…,hg], its
Middle g is the dimension of HOG characteristic vectors, finally couples together all characteristic vectors of zones of different block, obtains input picture block
Final HOG characteristic vectors represent hog=[hog1,hog2,…,hognh], wherein nh represents the number of region unit.
(1.3.2) local binary patterns (Local Binary Pattern, the LBP) feature of the image block after scaling is calculated
Vector；
Specifically, first by the minicell unit that image division size is 16*16, the LBP for calculating each cell factory is special
Levy, finally the LBP rectangular histograms of each minicell unit are stitched together obtains the LBP characteristic vectors of image block and represent；
In embodiments of the present invention, minicell unit (such as size is 16*16) is divided an image into first, for each
Each pixel in minicell unit, the gray value of 8 adjacent pixels is compared with it, if surrounding pixel values are more than
Center pixel value, then the position of the pixel be marked as 1, be otherwise 0, so, 8 point Jing in 3*3 fields compare and can produce
Raw 8 bits, that is, obtain the LBP values of the window center pixel, then calculates the rectangular histogram of each minicell unit, so
The rectangular histogram is normalized afterwards, so obtains LBP features lbp of each minicell unitj=[l1,l2,…,ll],
Wherein l represents the dimension of minicell element characteristic, and finally the LBP rectangular histograms of each minicell unit are stitched together obtains figure
As the LBP characteristic vectors of block represent lbp=[lbp1,lbp2,…,lbpnl], wherein nl represents the individual of image medium-small cells unit
Number.
(1.3.3) the coding SIFT feature of the image block after scaling is calculated；
Specifically, extracting SIFT feature coordinate is located at the feature in the image block, using bag of words (bag of
Words) and with reference to the code book predicted these features are encoded, the rectangular histogram that coding is obtained is normalized and is obtained image
The coding SIFT feature of block represents sift, and size C of code book is preset value.
(1.3.4) characteristic vector for obtaining car windshield image is finally represented
Specifically, by the HOG features obtained in above-mentioned steps (1.3.1) and (1.3.2) and the normalization of LBP features difference it
Afterwards with obtain SIFT feature and be stitched together in step (1.3.3), as the character representation F of image blocki=[hogi；lbpi；
sifti], 1≤i≤N, wherein i represent i-th image block, and N represents image block sum.Finally the feature of all image blocks is spelled
It is connected together
(2) based on many example dictionary learnings of margin maximization
(2.1) training image can be divided into two class P={ P+,P-, wherein P+Represent the car windshield figure comprising seat belt
Picture, P-Represent the car windshield image not comprising seat belt.From above-mentioned steps (1), per piece image several are included
Image block, each image block can be expressed as F by the HOG features, LBP features and SIFT feature splicedi.Two class images are distinguished
Training grader, using the class image block sample in above-mentioned two classes sample image as positive sample, then another kind of image block sample
As negative sample.
(2.2) to the positive sample and negative sample in step (2.1), based on margin maximization method, the power of positive sample is calculated
Weight matrix；Specifically, first to the image in positive sample carry out clustering the label z for obtaining clusteringi, according to sample weight and given ratio
Example extracts positive sample, and all negative samples constitute training dataset, solve optimization problem by multiclass SVM and try to achieve weight square
Battle array, by continuous iteration undated parameter the weight matrix W of optimum is obtained；Implement step as follows：
(2.2.1) to image block sample feature F in positive sampleiClustered, if K represents the middle calculation of cluster, and root
One label z is given to each image block according to cluster resulti∈ { 0,1 ..., K }, if F after clusteriBelong to kth class, then zi=k
∈{1,…,K}；If FiBelong to negative class, then zi=0；For image block sample its label in all negative samples is zi=0.
(2.2.2) weight matrix W=[w are defined0,w1,…,wK], k ∈ { 0,1 ..., K }, wherein wkRepresent k-th positive class
Clustering Model, w0The Clustering Model of negative class is represented, K represents cluster numbers；Define sample weight pij, pijIt is to ensure from training
The data extracted in data are positive sample, according to SVM decision functions to characteristic vector FiBelong between positive class and negative class most
It is poor greatly to be given, i.e.,Define psTo extract the positive sample in training data in proportion, generally ps
For fixed value, such as value is 0.7；
(2.2.3) under original state, if sample weight p of all positive classesijIt is 1.First according to sample weight pijExtract just
Sample, then extracts p from the sample for selectingsRatio, and all of negative sample composition training dataset D', wherein psExpression is taken out
The ratio of sampling example, D' represents the training dataset of new composition, because cluster labels there is known, can be solved by multiclass SVM
Following optimization problems are obtaining W：
WhereinrijRepresent in addition to this class, xijThe maximum of remaining class is rung
Should.
(2.2.4) sample weight and sample label are updated, obtains weight matrix to update sample weight with step (2.2.2)
pij, according toTo update the label of sample.
(2.2.5) repeat (2.2.3)～(2.2.4) processes, circulate M time, wherein M is default cycle-index, for example, take
It is worth for 5.
(2.3) positive sample and negative sample in step (2.1) is interchangeable, is calculated using the method in step (2.2)
The weight matrix of the positive sample obtained after exchange, the weight matrix that will be obtained in the weight matrix for obtaining here and step (2.2)
It is stitched together and obtains based on many example dictionaries of margin maximization；
In this step, the positive sample and negative sample in step (2.1) is interchangeable, after exchange, former positive sample
New negative sample is reformed into, and former negative sample reforms into new positive sample, so calculating the positive sample obtained after exchange here
This weight matrix, the weight matrix of negative sample actually namely in calculation procedure (2.1).
(3) grader of training of safety band identification
(3.1) concentrate car bumper wind glass right the training data obtained in step (1) using the dictionary obtained in step (2)
The characteristic vector of half part carries out encoding the coding characteristic for obtaining image vector.Specific coding mode is：
(3.1.1) image feature vector is multiplied with weight matrix, is obtained for the response of each disaggregated model maps.
(3.1.2) to each response mapping, response of the coordinate within image range is counted, and carries out non-maximum suppression
System, obtains the vector of a H dimension；
(3.1.3) utilization space pyramid model, is divided into 2*2 and 4*4 totally 20 blocks, for each block by former training image
Repeat the calculating process of (3.1.1) and (3.1.2), finally obtain the characteristic vector of 20 H dimensions；
(3.1.4) vector that above-mentioned two step is obtained is stitched together, obtains the vector of 21*H dimensions, be exactly the image
Final characteristic vector；
(3.2) using the coding characteristic vector training SVM classifier of all images obtained in training data.
SVM classifier is trained using training image, can be regarded using packaged existing SVM classifier built-in function herein
For a black box, the characteristic vector of all images is input into, output one is capable of deciding whether the svm classifier with seat belt
Device.
(4) car windshield image to be identified is recognized
(4.1) according to the method in step (1) to image zooming-out characteristic vector to be identified；
(4.2) images to be recognized has been obtained spy according to the method encoded to training data in step (3.1)
Levy vector to be encoded；
(4.3) using the SVM classifier trained in step (3), image to be identified is classified, and output safety
Result with identification.
Further, present invention also offers a kind of seat belt detection based on many example dictionary learnings of margin maximization with
Identifying system, the system includes car windshield image feature vector generation module, many example dictionary learnings of margin maximization
Module, classifier training module and car windshield picture recognition module to be identified, wherein：
The car windshield right half part image feature vector generation module, for obtaining training image data set in
The characteristic vector of car windshield right half part image represents, specifically includes different scale images block acquisition module, car bumper wind glass
The representation module of glass right half part image SIFT feature extraction module, image block characteristics extraction module and characteristics of image, wherein：
The different scale images block acquisition module, for the car windshield image concentrated to training image, carries first
The right half part of pick-up windshield, to right half part car windshield the image block under different scale is obtained；
The car windshield SIFT feature extraction module, for car windshield image zooming-out SIFT feature and sharp
The process encoded with default code book；
Described image block feature extraction module, for extracting characteristic vector to the image block of the different scale for acquiring.
Specifically include image block scaling submodule, HOG characteristic vector calculating sub modules, LBP characteristic vectors calculating sub module and coding
SIFT feature histogram calculation submodule.
Described image block scales submodule, for zooming to same yardstick to the image block of different scale, generally by image
Zoom to the smallest dimension in three kinds of yardsticks；
The HOG characteristic vectors calculating sub module, for calculating the image block that image block scaling submodule is acquired
HOG characteristic vectors；
The LBP characteristic vectors calculating sub module, for calculating the image block that image block scaling submodule is acquired
LBP characteristic vectors；
The coding SIFT feature histogram calculation submodule, for calculating image block scaling submodule image is acquired
The SIFT histogram features vector of block；
The character representation module of described image, for calculating the character representation of car windshield image, to every image, will
The merging features for representing all image blocks of image act the character representation for being used as image；
The margin maximization many example dictionary learning modules, for learning encoder dictionary, for training data is concentrated
Each class, is respectively trained a multi classifier, obtains a disaggregated model, and the weight in all of sorter model is merged
Get up to obtain dictionary to represent；
The classifier training module, for training of safety band recognition classifier, carries out coding mould to training data first
Block, the disaggregated model obtained using many example dictionary learning modules of margin maximization is encoded to training data；Finally utilize
The coding characteristic vector training SVM classifier of all images obtained in training image；
The car windshield picture recognition module to be identified, for recognizing whether images to be recognized has seat belt, specifically
Extraction module, the SIFT feature extraction module of car windshield image including car windshield image block to be identified, image block
Characteristic extracting module, image feature representation module, characteristics of image coding module, car windshield picture recognition module to be identified,
Wherein：
The car windshield image block extraction module to be identified, for obtaining the figure of car windshield right half part image
As block is represented；
The SIFT feature extraction module of the car windshield image, for obtaining car windshield right half part image
SIFT feature, and using bag of words SIFT feature is encoded with reference to the code book predicted；
Described image block feature extraction module, for extracting the characteristic vector for representing image block；
Described image character representation module, for the image block characteristics of every piece image to be stitched together as the spy of image
Levy vector；
Described image feature coding module, for the dictionary that obtained using margin maximization multi-instance learning to characteristics of image
Encoded；
The car windshield picture recognition module to be identified, for the seat belt trained using classifier training module
Recognition classifier, classifies to image to be identified, and the recognition result of output safety band.
Specifically, the right half part of car windshield for intercepting being obtained, respectively with three kinds of scale sizes (such as 48*48,
72*72,96*96) sliding window intercept in the enterprising line slip of image, sliding window moving step length for window 1/3 length, will
Each time the picture material in sliding window is intercepted and preserved, and so finally gives the image block of several different scalesWherein N1、N2、N3Three kinds of different scale hypograph blocks are represented respectively
Number, N=N1+N2+N3Represent the number of total image block.
Specifically, the HOG characteristic vectors calculating sub module, for seeking each pixel gradient (including size and side
To), minicell unit is divided an image into, the histogram of gradients of each cell factory is counted, by several cell factories (such as 3*3)
One region unit of composition, the histogram of gradients of each region unit inner cell unit is stitched together special as the HOG of the region unit
Levy the sub- hog of descriptionk=[h1,h2,…,hg], wherein g is the dimension of HOG characteristic vectors, finally by all spies of zones of different block
Levy vector to couple together, the final HOG characteristic vectors for obtaining input picture block represent hog=[hog1,hog2,…,hognh], its
Middle nh represents the number of region unit.
Specifically, the LBP characteristic vectors calculating sub module, divides an image into first the minicell unit of 16*16, meter
The rectangular histogram of each minicell unit is calculated, then the rectangular histogram is normalized, so obtain each minicell unit
LBP features lbpj=[l1,l2,…,ll], wherein l represents the dimension of minicell element characteristic, finally by each minicell list
The LBP rectangular histograms of unit are stitched together and obtain the LBP characteristic vectors of image block and represent lbp=[lbp1,lbp2,…,lbpnl], its
Middle nl represents the number of image medium-small cells unit.
Specifically, the SIFT encodes histogram feature, is located in the image block for extracting SIFT feature coordinate
These features are encoded by feature using bag of words (bag of words) and with reference to the code book predicted, and coding is obtained
Rectangular histogram be normalized and obtain image block SIFT feature and represent that sift, the C of code book are typically sized to preset value.
Specifically, many example dictionary learning modules of margin maximization are used to perform following step, first to generate encoder dictionary
Training data is first splitted data into the sample and sample without seat belt of seat belt according to artificial annotation results, will be above-mentioned
A class image block sample in two class sample images aligns sample as positive sample, then another kind of image block sample as negative sample
This and negative sample repeat following step：
(2.2.1) to image block sample feature F in positive sampleiClustered, if K represents the middle calculation of cluster, and root
One label z is given to each image block according to cluster resulti∈ { 0,1 ..., K }, if F after clusteriBelong to kth class, then zi=k
∈{1,…,K}；If FiBelong to negative class, then zi=0；For image block sample its label in all negative samples is zi=0；
(2.2.2) weight matrix W=[w are defined0,w1,…,wK], k ∈ { 0,1 ..., K }, wherein wkRepresent k-th positive class
Clustering Model, w0The Clustering Model of negative class is represented, K represents cluster numbers；Define sample weight pij, pijIt is to ensure from training
The data extracted in data are positive sample, according to SVM decision functions to characteristic vector FiBelong between positive class and negative class most
It is poor greatly to be given, i.e.,Define psTo extract the positive sample in training data in proportion；
(2.2.3) under original state, if sample weight p of all positive classesij1 is, first according to sample weight pijExtract just
Sample, then extracts p from the sample for selectingsRatio, and all of negative sample composition training dataset D', wherein psExpression is taken out
The ratio of sampling example, D' represents the training dataset of new composition, solves following optimization problems to obtain W by multiclass SVM：
WhereinrijRepresent in addition to this class, xijThe maximum of remaining class is rung
Should；
(2.2.4) sample weight and sample label are updated, the weight matrix obtained with step (2.2.2) is updating sample power
Weight pij；According toTo update the label of sample；
(2.2.5) repeat (2.2.3)～(2.2.4) processes, circulate M time, wherein M is default cycle-index.For example take
It is worth for 5.
Then above-mentioned positive sample and negative sample are interchangeable, using said method the positive sample after exchanging is recalculated
Weight matrix, obtain two weight matrix is stitched together and is obtained based on many example dictionaries of margin maximization.
Claims (9)
1. a kind of seat belt recognition methodss based on many example dictionary learnings of margin maximization, it is characterised in that methods described bag
Include following step：
(1) obtaining training data concentrates the characteristic vector of car bumper wind glass right half part image to represent：
(1.1) the car windshield image concentrated to training image, intercepts first the right half part image of car windshield, to car
Windshield right half part extracts the image block of different scale；Wherein, the car windshield image is from vehicle frontal i.e. car
The image shot to driver's cabin at windshield, in its right half part image driver is included；
(1.2) (Scale-invariant is changed to the right half part image zooming-out scale invariant feature of car windshield
Feature transform, SIFT) characteristic vector, and SIFT feature is encoded using the code book predicted；
(1.3) image block of the different scale obtained in step (1.1) is zoomed to into same scale size, and to scaling after it is every
One image block extracts respectively characteristic vector, and the characteristic vector of described image block is by the histograms of oriented gradients of image block
(Histogram of oriented gradients, HOG) characteristic vector, local binary pattern (Local Binary
Pattern, LBP) characteristic vector, SIFT feature vector composition；
(2) based on many example dictionary learnings of margin maximization：
(2.1) the every training image concentrated to training image, being divided into training dataset using artificial mark has seat belt
Sample image and the sample image without seat belt, using the class image block sample in above-mentioned two classes image as positive sample, then
Another kind of image block sample is used as negative sample；
(2.2) to the positive sample and negative sample in step (2.1), based on margin maximization method, the weight square of positive sample is obtained
Battle array；Specifically, first to the image in positive sample carry out clustering the label z for obtaining clusteringi, taken out according to sample weight and given ratio
Positive sample is taken, and all negative samples constitute training dataset, solve optimization problem by multiclass SVM and try to achieve weight matrix, lead to
Cross the weight matrix W that continuous iteration undated parameter obtains optimum；
(2.3) positive sample and negative sample in step (2.1) is interchangeable, is calculated using the method in step (2.2) and exchanged
The weight matrix of the positive sample for obtaining afterwards, by the weight matrix obtained in the weight matrix for obtaining here and step (2.2) splicing
Obtain together based on many example dictionaries of margin maximization；
(3) grader of training of safety band identification：
(3.1) car bumper wind glass right-hand part is concentrated to the training data obtained in step (1) using the dictionary obtained in step (2)
The characteristic vector divided carries out encoding the coding characteristic for obtaining image vector；
(3.2) using the coding characteristic vector training SVM classifier of all images obtained in training data；
(4) car windshield image to be identified is recognized
(4.1) according to the method in step (1) to image zooming-out characteristic vector to be identified；
(4.2) characteristic vector for having obtained is encoded according to the method in step (3.1)；
(4.3) using the SVM classifier trained in step (3), image to be identified is classified, and output safety band is known
Other result；
The step (2.2) specifically includes：
(2.2.1) to image block sample feature F in positive sampleiClustered, if K represents the middle calculation of cluster, and according to cluster
As a result a label z is given to each image blocki∈ { 0,1 ..., K }, if F after clusteriBelong to kth class, then zi=k ∈
{1,…,K}；If FiBelong to negative class, then zi=0；For image block sample its label in all negative samples is zi=0；
(2.2.2) weight matrix W=[w are defined0,w1,…,wK], k ∈ { 0,1 ..., K }, wherein wkRepresent the cluster of k-th positive class
Model, w0The Clustering Model of negative class is represented, K represents cluster numbers；Define sample weight pij, pijIt is to ensure from training data
The data of extraction are positive sample, according to SVM decision functions to characteristic vector FiBelong between positive class and negative class maximum difference to
Go out, i.e.,Define psTo extract the positive sample in training data in proportion；
(2.2.3) under original state, if sample weight p of all positive classesij1 is, first according to sample weight pijExtract positive sample
Example, then extracts p from the sample for selectingsRatio, and all of negative sample composition training dataset D', wherein psRepresent and extract
The ratio of sample, D' represents the training dataset of new composition, solves following optimization problems to obtain W by multiclass SVM：
WhereinrijRepresent to except the outer x of this classijPeak response to remaining class；
(2.2.4) sample weight and sample label are updated, the weight matrix obtained with step (2.2.2) is updating sample weight
pij；According toTo update the label of sample；
(2.2.5) repeat (2.2.3)～(2.2.4) processes, circulate M time, wherein M is default cycle-index.
2. the seat belt recognition methodss based on many example dictionary learnings of margin maximization according to claim 1, its feature
It is that the step (3.1) is specially：
(3.1.1) by the image feature vector obtained in step (1) be multiplied based on many example dictionaries of margin maximization, obtain right
Map in the response of each disaggregated model；
(3.1.2) to each response mapping, response of the coordinate within image range is counted, and carries out non-maximum suppression, obtained
To the vector of a H dimension；
(3.1.3) utilization space pyramid model, by former training image 2*2 and 4*4 totally 20 blocks are divided into, and are repeated for each block
(3.1.1) with the calculating process of (3.1.2), the characteristic vector of 20 H dimensions is finally obtained；
(3.1.4) vector that above-mentioned two step is obtained is stitched together, obtains the vector of 21*H dimensions, be exactly the final of the image
Characteristic vector.
3. the seat belt recognition methodss based on many example dictionary learnings of margin maximization according to claim 1 and 2, it is special
Levy and be, the step (1.1) is specially：The right half part of the car windshield obtained to intercepting is big with three kinds of yardsticks respectively
Little sliding window is intercepted in the enterprising line slip of image, the picture material in sliding window each time is intercepted and is preserved, this
Sample finally gives the image block of multiple different scales.
4. the seat belt recognition methodss based on many example dictionary learnings of margin maximization according to claim 1 and 2, it is special
Levy and be, the step (1.3) is specially：
(1.3.1) the HOG characteristic vectors of the image block after scaling are calculated；
(1.3.2) the LBP characteristic vectors of the image block after scaling are calculated；
(1.3.3) the coding SIFT feature of the image block after scaling is calculated；
(1.3.4) by the HOG features obtained in above-mentioned (1.3.1) and (1.3.2) step and LBP features difference normalization after with
The coding SIFT feature obtained in step (1.3.3) is stitched together, used as the character representation of image block.
5. the seat belt recognition methodss based on many example dictionary learnings of margin maximization according to claim 4, its feature
It is that the step (1.3.1) is specially：
Ask each pixel gradient, the gradient to include size and Orientation, divide an image into minicell unit, count each
Several cell factories are constituted a region unit, by the ladder of each region unit inner cell unit by the histogram of gradients of cell factory
The HOG Feature Descriptors that degree rectangular histogram is stitched together as the region unit, finally connect all characteristic vectors of zones of different block
Pick up and, the final HOG characteristic vectors for obtaining input picture block are represented.
6. the seat belt recognition methodss based on many example dictionary learnings of margin maximization according to claim 4, its feature
It is that the step (1.3.2) is specially：
The minicell unit of 16*16 is divided an image into first, the LBP features of each minicell unit is calculated, finally by each
The LBP rectangular histograms of minicell unit are stitched together and obtain the LBP characteristic vectors of image block and represent.
7. the seat belt recognition methodss based on many example dictionary learnings of margin maximization according to claim 5, its feature
It is that the step (1.3.3) is specially：
Extract SIFT feature coordinate and be located at the feature in the image block, using bag of words and with reference to the code book predicted to these
Feature is encoded, and the rectangular histogram that obtains of coding is normalized and is obtained the coding SIFT feature of image block and is represented sift, code
This size C is preset value.
8. the seat belt recognition methodss based on many example dictionary learnings of margin maximization according to claim 4, its feature
It is that the step (1.3.4) is specially：
Specifically, first to each image block, by the HOG features obtained in above-mentioned steps (1.3.1) and (1.3.2) and LBP
After feature difference normalization and obtain SIFT feature in step (1.3.3) to be stitched together, as the character representation of image block
Fi=[hogi；lbpi；sifti], 1≤i≤N, wherein i represent i-th image block, and N represents image block sum；Finally will be all
The merging features of image block are together
9. a kind of seat belt identifying system based on many example dictionary learnings of margin maximization, it is characterised in that the system bag
Include car windshield image feature vector generation module, many example dictionary learning modules of margin maximization, training data is carried out
Coding module, classifier training module and car windshield picture recognition module to be identified, wherein：
The car windshield image feature vector generation module, for obtaining training image data set in car windshield figure
The characteristic vector of picture represents, specifically includes different scale images block acquisition module, car windshield image SIFT feature and extract mould
The representation module of block, image block characteristics extraction module and characteristics of image, wherein：
The different scale images block acquisition module, for the car windshield image concentrated to training image, extracts first car
The right half part of windshield, to right half part car windshield the image block under different scale is obtained；
The car windshield SIFT feature extraction module, for car windshield image zooming-out SIFT feature, and utilizes word
Bag model combines the code book of precognition and SIFT feature is encoded；
Described image block feature extraction module, for extracting characteristic vector to the image block of the different scale for acquiring；Specifically
Submodule is scaled including image block, HOG characteristic vector calculating sub modules, LBP characteristic vectors calculating sub module is special with coding SIFT
Levy histogram calculation submodule；
Described image block scales submodule, for zooming to same yardstick to the image block of different scale, generally by image scaling
To the smallest dimension in three kinds of yardsticks；
The HOG characteristic vectors calculating sub module, the HOG for calculating the image block that image block scaling submodule is acquired is special
Levy vector；
The LBP characteristic vectors calculating sub module, the LBP for calculating the image block that image block scaling submodule is acquired is special
Levy vector；
The coding SIFT feature histogram calculation submodule, for calculating image block scaling submodule image block is acquired
SIFT histogram features vector；
The character representation module of described image, for calculating the character representation of car windshield image, to every image, would indicate that
The merging features of all image blocks of image act the character representation for being used as image；
The many example dictionary learning modules of the margin maximization, it is each for what training data was concentrated for learning encoder dictionary
Class, is respectively trained a multi classifier, obtains a disaggregated model, and the weight in all of sorter model is combined
Obtain dictionary to represent；
Coding module is carried out to training data, for the disaggregated model obtained using many example dictionary learning modules of margin maximization
Training data is encoded；
The classifier training module, for training of safety band recognition classifier, using all images obtained in training image
Coding characteristic vector training SVM classifier；
The car windshield picture recognition module to be identified, for recognizing whether images to be recognized has seat belt, specifically includes
The extraction module of car windshield image block to be identified, the SIFT feature extraction module of car windshield image, image block characteristics
Extraction module, image feature representation module, characteristics of image coding module, car windshield picture recognition module to be identified, wherein：
The car windshield image block extraction module to be identified, for obtaining the image block of car windshield right half part image
Represent；
The SIFT feature extraction module of the car windshield image, for obtaining the SIFT of car windshield right half part image
Feature, and it is encoded with known code book；
Described image block feature extraction module, for extracting the characteristic vector for representing image block；
Described image character representation module, for the image block characteristics of every piece image to be stitched together as the feature of image
Vector；
Described image feature coding module, for the dictionary that obtained using many example dictionary learnings of margin maximization to characteristics of image
Encoded；
The car windshield picture recognition module to be identified, for the seat belt identification trained using classifier training module
Grader, classifies to image to be identified, and the recognition result of output safety band.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410442594.1A CN104200228B (en) | 2014-09-02 | 2014-09-02 | Recognizing method and system for safety belt |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410442594.1A CN104200228B (en) | 2014-09-02 | 2014-09-02 | Recognizing method and system for safety belt |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104200228A CN104200228A (en) | 2014-12-10 |
| CN104200228B true CN104200228B (en) | 2017-05-10 |
Family
ID=52085517
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410442594.1A Expired - Fee Related CN104200228B (en) | 2014-09-02 | 2014-09-02 | Recognizing method and system for safety belt |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104200228B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106295601B (en) * | 2016-08-18 | 2019-02-26 | 合肥工业大学 | A kind of improved Safe belt detection method |
| CN106909894B (en) * | 2017-02-14 | 2018-08-14 | 北京深瞐科技有限公司 | Vehicle brand type identifier method and system |
| CN107529659B (en) * | 2017-07-14 | 2018-12-11 | 深圳云天励飞技术有限公司 | Seatbelt wearing detection method, device and electronic equipment |
| CN107577994A (en) * | 2017-08-17 | 2018-01-12 | 南京邮电大学 | A kind of pedestrian based on deep learning, the identification of vehicle auxiliary product and search method |
| CN108416289A (en) * | 2018-03-06 | 2018-08-17 | 陕西中联电科电子有限公司 | A kind of working at height personnel safety band wears detection device and detection method for early warning |
| CN109492772B (en) * | 2018-11-28 | 2020-06-23 | 北京百度网讯科技有限公司 | Method and device for generating information |
| CN110163287A (en) * | 2019-05-24 | 2019-08-23 | 三亚中科遥感研究所 | A kind of mesoscale eddy detection method and device |
| CN111553893A (en) * | 2020-04-24 | 2020-08-18 | 成都飞机工业（集团）有限责任公司 | Method for identifying automatic wiring and cutting identifier of airplane wire harness |
| CN112016502B (en) * | 2020-09-04 | 2023-12-26 | 平安国际智慧城市科技股份有限公司 | Safety belt detection method, safety belt detection device, computer equipment and storage medium |
| CN112149567A (en) * | 2020-09-23 | 2020-12-29 | 中国科学院长春光学精密机械与物理研究所 | Offshore target detection method, device and equipment |
| CN115797340A (en) * | 2023-02-03 | 2023-03-14 | 西南石油大学 | Industrial surface defect detection method based on multi-instance learning |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103745226A (en) * | 2013-12-31 | 2014-04-23 | 国家电网公司 | Dressing safety detection method for worker on working site of electric power facility |
| CN103955704A (en) * | 2014-04-26 | 2014-07-30 | 合肥工业大学 | Safety belt detection method based on Adaboost |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2574958B1 (en) * | 2011-09-28 | 2017-02-22 | Honda Research Institute Europe GmbH | Road-terrain detection method and system for driver assistance systems |
-
2014
- 2014-09-02 CN CN201410442594.1A patent/CN104200228B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103745226A (en) * | 2013-12-31 | 2014-04-23 | 国家电网公司 | Dressing safety detection method for worker on working site of electric power facility |
| CN103955704A (en) * | 2014-04-26 | 2014-07-30 | 合肥工业大学 | Safety belt detection method based on Adaboost |
Non-Patent Citations (2)
| Title |
|---|
| "图像处理与机器学习在未系安全带驾车检测中的应用";吴法;《中国博士学位论文全文数据库 信息科技辑》;20140815(第08期);I138-12 * |
| "最大间隔方法及其在图像检索中的应用";胡洋;《中国博士学位论文全文数据库 信息科技辑》;20091015(第10期);I138-38 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104200228A (en) | 2014-12-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104200228B (en) | Recognizing method and system for safety belt | |
| CN106599773B (en) | Deep learning image identification method and system for intelligent driving and terminal equipment | |
| Zang et al. | Vehicle license plate recognition using visual attention model and deep learning | |
| CN108154102B (en) | Road traffic sign identification method | |
| WO2019169816A1 (en) | Deep neural network for fine recognition of vehicle attributes, and training method thereof | |
| CN103514456B (en) | Image classification method and device based on compressed sensing multi-core learning | |
| CN109583482B (en) | Infrared human body target image identification method based on multi-feature fusion and multi-kernel transfer learning | |
| Puranic et al. | Vehicle number plate recognition system: a literature review and implementation using template matching | |
| CN109086792A (en) | Based on the fine granularity image classification method for detecting and identifying the network architecture | |
| CN110321967B (en) | Image classification improvement method based on convolutional neural network | |
| CN105069481B (en) | Natural scene multiple labeling sorting technique based on spatial pyramid sparse coding | |
| CN105956560A (en) | Vehicle model identification method based on pooling multi-scale depth convolution characteristics | |
| CN110175615B (en) | Model training method, domain-adaptive visual position identification method and device | |
| Khalid et al. | Automatic measurement of the traffic sign with digital segmentation and recognition<? show [AQ ID= Q1]?> | |
| CN110689043A (en) | Vehicle fine granularity identification method and device based on multiple attention mechanism | |
| JP2016062610A (en) | Feature model creation method and feature model creation device | |
| CN106257496B (en) | Mass network text and non-textual image classification method | |
| CN105989334B (en) | Road detection method based on monocular vision | |
| Kuang et al. | Feature selection based on tensor decomposition and object proposal for night-time multiclass vehicle detection | |
| CN103366181A (en) | Method and device for identifying scene integrated by multi-feature vision codebook | |
| CN106570439B (en) | Vehicle detection method and device | |
| CN111046732A (en) | Pedestrian re-identification method based on multi-granularity semantic analysis and storage medium | |
| CN108230330B (en) | Method for quickly segmenting highway pavement and positioning camera | |
| CN116188999B (en) | Small target detection method based on visible light and infrared image data fusion | |
| Sugiharto et al. | Traffic sign detection based on HOG and PHOG using binary SVM and k-NN |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170510 |