CN104200228A - Recognizing method and system for safety belt - Google Patents

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

A kind of securing band recognition methods and system

Technical field

The invention belongs to technical field of computer vision, more specifically, relate to a kind of securing band recognition methods and system based on many examples of margin maximization dictionary learning.

Background technology

Along with the development of global economy situation, people's living standard improves day by day, and the quantity that individual has motor vehicles is geometric growth situation, and automobile universal becomes current inexorable trend.What bring is the supervision and management work of automotive safety thereupon, and wearing of pilot harness is the work that primarily needs supervision and management, and the detection and Identification work of therefore studying pilot harness seems particularly urgent and important.To the research of securing band detection and Identification, be not also very perfect at present, securing band detection algorithm is before mainly to rely on the simple image processing algorithms such as rim detection, and accuracy rate is not also very high.The correlation techniques such as therefore wanting to obtain higher accuracy rate and detect faster effect must binding pattern identification, computer vision, first requirement can extract the car windshield approximate location of automobile in motion identify from complicated background, and whether by correlation techniques such as image pre-service, feature extraction, dictionary learning, training classifiers, detect driver has band securing band.

Yet securing band detects and also has a lot of difficult problems at present, for example, the impact of angle during image taking, light etc., the impact of driver's dressing color and securing band color contrast difference.These factors all can detect and bring adverse influence to securing band.

Summary of the invention

The object of the present invention is to provide a kind of securing band recognition methods based on many examples of margin maximization dictionary learning, the method identifying is simple and accuracy rate is higher.

For achieving the above object, according to one aspect of the present invention, provide a kind of securing band recognition methods based on many examples of margin maximization dictionary learning, comprised the steps:

(1) obtaining training data concentrates the proper vector of car bumper wind glass right half part image to represent;

(1.1) to the concentrated car windshield image of training image, first intercept the right half part image of car windshield, car windshield right half part is extracted to the image block of different scale;

Described car windshield image is from vehicle front, to be the image that car windshield is taken pilothouse, in its right half part image, includes driver.

Particularly, the right half part of the car windshield that intercepting is obtained, uses respectively the moving window of three kinds of scale size (for example 48*48,72*72,96*96) at the enterprising line slip intercepting of image, the final like this image block that obtains several different scales $D_N=\{D_{11},\·\·\·,D_{1N_1};D_{21},\·\·\·,D_{{2N}_2};D_{31},\·\·\·,D_{{3N}_3}\},$ N wherein ₁, N ₂, N ₃the number that represents respectively three kinds of different scale hypograph pieces, N=N ₁+ N ₂+ N ₃the number that represents total image block.

(1.2) the right half part image of car windshield is extracted to SIFT proper vector, and utilize default code book to encode to SIFT feature;

Described default code book can obtain by the SIFT proper vector that extraction obtains to training data is carried out to cluster.

(1.3) image block of the different scale obtaining in step (1.1) is zoomed to same scale size, and each image block after convergent-divergent is extracted respectively to proper vector, the proper vector of described image block is comprised of HOG proper vector, LBP proper vector, the SIFT proper vector of image block;

(1.3.1) the HOG proper vector of computed image piece:

Be specially, each pixel is asked to gradient (comprising size and Orientation), image is divided into cellule unit, add up the histogram of gradients of each cell unit, several cells unit (as 3*3) is formed to a region unit, the histogram of gradients of each region unit inner cell unit is stitched together as the HOG Feature Descriptor of this region unit, finally all proper vectors of zones of different piece is coupled together, the final HOG proper vector that obtains input picture piece represents;

(1.3.2) the LBP proper vector of computed image piece:

Be specially, first image being divided to size is the cellule unit of 16*16, calculates the LBP feature of each cell unit, and the LBP proper vector that obtains image block that finally the LBP histogram of each cellule unit is stitched together represents;

(1.3.3) the coding SIFT feature of computed image piece

Be specially, extract SIFT characteristic coordinates and be located at the feature in this image block, utilize word bag model (bag of words) and in conjunction with the code book of precognition, these features encoded, the histogram that coding is obtained is normalized the coding SIFT character representation sift that obtains image block, and the big or small C of code book is preset value.

(1.3.4) proper vector that obtains car windshield image finally represents

Be specially, first to each image block, by above-mentioned (1.3.1) and the HOG feature that (1.3.2) obtains in step and LBP feature respectively after normalization together with the coding SIFT merging features obtaining in step (1.3.3), as the character representation F of image block _i=[hog _i; lbp _i; Sift _i], 1≤i≤N, wherein i represents i image block, N presentation video piece sum; Together by the merging features of all image blocks finally $F={\left\{F_i\right\}}_{i=1}^N.$

(2) based on margin maximization multiclass example dictionary learning

(2.1) every training image training image being concentrated, utilize artificial mark that training dataset is divided into the sample image that has securing band and the sample image that there is no securing band, using the class image block sample in above-mentioned two class images as positive sample, another kind of image block is as negative sample;

(2.2), to positive sample and negative sample in step (2.1), based on margin maximization method, calculate the weight matrix of positive sample; Particularly, the image first aligning in sample carries out the label z that cluster obtains cluster _i, according to sample weight with to certainty ratio, extract positive sample, and all negative sample composition training datasets, by multiclass SVM, solve optimization problem and try to achieve weight matrix, by continuous iteration undated parameter, obtain optimum weight matrix W; Specific implementation step is as follows:

(2.2.1) align the image block sample F in sample _icarry out cluster, establish the middle calculation that K represents cluster, and according to cluster result to the given label z of each image block sample _i∈ 0,1 ..., K}, if F after cluster _ibelong to k class, z _i=k ∈ 1 ..., K}; If F _ibelong to negative class, z _i=0; For its label of image block sample in all negative samples, be z _i=0;

(2.2.2) definition weight matrix W=[w ₀, w ₁..., w _k], k ∈ 0,1 ..., K}, wherein w _kthe Clustering Model that represents k positive class, w ₀the Clustering Model that represents negative class, K represents cluster numbers; Definition sample weight p _ij, p _ijthat data in order to guarantee to extract from training data are positive sample, according to SVM decision function to proper vector F _ithe maximum difference belonging between positive class and negative class provides, definition p ^sfor extracting in proportion the positive sample in training data; P generally ^sfor fixed value, for example value is 0.7;

(2.2.3), under original state, establish the sample weight p of all positive classes _ijbe 1.First according to sample weight p _ijextract positive sample, then from the sample of selecting, extract p ^sratio, and all negative samples form training dataset D', wherein p ^srepresent to extract the ratio of sample, D' represents the new training dataset forming, and due to cluster label cicada, can solve following optimization problem by multiclass SVM and obtain W:

$\underset{W}{\min }\underset{k=0}{\overset{K}{\mathrm{\Σ}}}{\left|\right|w_k\left|\right|}^2+\mathrm{\λ}\underset{\mathrm{ij}}{\mathrm{\Σ}}\max (\mathrm{0,1}+w_{r_{\mathrm{ij}}}^T{x}_{\mathrm{ij}}-w_{z_{\mathrm{ij}}}^T{x}_{\mathrm{ij}}),$

Wherein r _ijexpression is to except this class, x _ijpeak response to all the other classes;

(2.2.4) upgrade sample weight and sample label, by step (2.2.2), obtain weight matrix and upgrade sample weight p _ij; According to upgrade the label of sample;

(2.2.5) repeat (2.2.3)～(2.2.4) process, circulation M time, wherein M is default cycle index, for example value is 5.

(2.3) positive sample and negative sample in step (2.1) are exchanged, the weight matrix that utilizes the positive sample obtaining after the method calculating exchange in step (2.2), is stitched together the weight matrix obtaining in the weight matrix obtaining and step (2.2) to obtain based on many examples of margin maximization dictionary here.

(3) sorter of training of safety band identification

(3.1) utilize the dictionary obtaining in step (2) to encode and obtain the coding characteristic vector of image the proper vector of the concentrated car bumper wind glass right half part of the training data obtaining in step (1); Specific coding mode is:

(3.1.1) image feature vector and weight matrix are multiplied each other, obtain the response mapping for each disaggregated model.

(3.1.2) to each response mapping, the response of statistics coordinate within image range, and carry out non-maximum the inhibition, obtain the vector of a H dimension;

(3.1.3) utilize space pyramid model, former training image is divided into 2*2 and 4*4 totally 20 pieces, for each piece, repeat (3.1.1) and computation process (3.1.2), finally obtain the proper vector that 20 H tie up;

(3.1.4) vector above-mentioned two steps being obtained is stitched together, and obtains the vector of 21*H dimension, is exactly the final proper vector of this image;

(3.2) utilize training data to concentrate the coding characteristic vector training svm classifier device of all images that obtain, can utilize packaged existing svm classifier device built-in function to be considered as a black box herein, input the proper vector of all training images, export a svm classifier device that can judge whether with securing band.

(4) identify car windshield image to be identified

(4.1) according to the method in step (1), image to be identified is extracted to proper vector;

(4.2) according to the method that training data is encoded in step (3.1), treating the proper vector that recognition image obtained encodes;

(4.3) utilize the svm classifier device training in step (3), image to be identified is classified, and the result of output safety band identification.

The invention discloses a kind of securing band recognition methods based on many examples of margin maximization dictionary learning.For the given width car windshield image of user, the present invention can intercept the car bumper wind glassy zone in image, and car windshield driver is partly carried out to securing band identification.

To the concentrated car windshield image of training data, obtain first the different images piece of presentation video, each image block is extracted respectively to feature, and be stitched together as the character representation of image, then the process of the feature of using image based on many examples of margin maximization dictionary learning, the dictionary obtaining with study is encoded to training data feature, finally with the features training svm classifier device after coding.

To car windshield image to be identified, first similar with the concentrated car windshield image processing method of training data, first, image block to each car windshield Image Acquisition different scale to be identified, each image block is extracted respectively to feature, and be stitched together as the character representation of image, then utilize the dictionary that training process obtains to encode to car windshield characteristics of image to be identified, finally with the sorter that training obtains, image to be identified is identified, obtained the recognition result of car windshield image to be identified.

According to another aspect of the present invention, a kind of securing band recognition system based on many examples of margin maximization dictionary learning is also provided, described system comprises car windshield image feature vector generation module, many examples of margin maximization dictionary learning module, sorter training module and car windshield picture recognition module to be identified, wherein:

Described car windshield right half part image feature vector generation module, for obtaining the proper vector of the car windshield right half part image of training image data centralization, represent, the representation module that specifically comprises different scale images piece acquisition module, car windshield right half part image SIFT characteristic extracting module, image block characteristics extraction module and characteristics of image, wherein:

Described different scale images piece acquisition module, for to the concentrated car windshield image of training image, first extracts the right half part of car windshield, and right half part car windshield is obtained to the image block under different scale;

Described car windshield SIFT characteristic extracting module, for car windshield image is extracted to SIFT feature, and the process of utilizing default code book to encode;

Described image block characteristics extraction module, extracts proper vector for the image block to the different scale acquiring.Specifically comprise image block convergent-divergent submodule, HOG proper vector calculating sub module, LBP proper vector calculating sub module and coding SIFT feature histogram calculating sub module.

Described image block convergent-divergent submodule, for the image block of different scale is zoomed to same yardstick, scales the images to three kinds of smallest dimension in yardstick conventionally;

Described HOG proper vector calculating sub module, the HOG proper vector of the image block acquiring for computed image piece convergent-divergent submodule;

Described LBP proper vector calculating sub module, the LBP proper vector of the image block acquiring for computed image piece convergent-divergent submodule;

Described coding SIFT feature histogram calculating sub module, the SIFT histogram feature that acquires image block for computed image piece convergent-divergent submodule is vectorial;

The character representation module of described image, for calculating the character representation of car windshield image, to every image, plays the merging features of all image blocks of presentation video the character representation that is used as image;

Described many examples of margin maximization dictionary learning module, be used for learning encoder dictionary, each class of concentrating for training data, trains respectively a multicategory classification device, obtain a disaggregated model, the weight in all sorter models is combined and obtains dictionary and represent;

Described sorter training module, for training of safety band recognition classifier, first carries out coding module to training data, and the disaggregated model that utilizes many examples of margin maximization dictionary learning module to obtain is encoded to training data; Finally utilize the coding characteristic vector training svm classifier device of all images that obtain in training image;

Described car windshield picture recognition module to be identified, be used for identifying image to be identified and whether have securing band, specifically comprise the extraction module of car windshield image block to be identified, the SIFT characteristic extracting 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:

Described car windshield image block extraction module to be identified, represents for obtaining the image block of car windshield right half part image;

The SIFT characteristic extracting module of described car windshield image, for obtaining the SIFT feature of car windshield right half part image, and encodes to it with known code book;

Described image block characteristics extraction module, for extracting the proper vector of presentation video piece;

Described image feature representation module, for being stitched together the image block characteristics of every piece image as the proper vector of image;

Described characteristics of image coding module, encodes to characteristics of image for the dictionary that utilizes the many learn-by-examples of margin maximization to obtain;

Described car windshield picture recognition module to be identified, for the securing band recognition classifier of utilizing sorter training module to train, classifies to image to be identified, and the recognition result of output safety band.

The above technical scheme of conceiving by the present invention, compared with prior art, the present invention has following technique effect:

Car windshield image be easily subject to the impact of the factors such as illumination, camera resolution and shooting angle and motorist's clothing and securing band color contrast and become be not easy identification, current algorithm is mainly to utilize image to process related algorithm to carry out the means such as rim detection, straight-line detection, these algorithms cannot adapt to the variation of external environment, and high identification cannot be all provided under various complex scenes.The inventive method adopts the various features presentation video that is stitched together, and makes the character representation of image have anti-rotation, the characteristics such as anti-noise, anti-illumination.Therefore, the inventive method can effectively overcome the impact of the factors such as illumination, shooting angle variation;

2. the inventive method has adopted many examples of margin maximization dictionary learning method, first each class image block that training data is concentrated passes through cluster, many examples of margin maximization dictionary learning obtains a disaggregated model, and the disaggregated model obtaining like this has the highest response to this similar image.Therefore, the inventive method can obtain higher recognition accuracy to car windshield;

3. method of the present invention is divided into 1*1 by image, 2*2, the image pyramid of 4*4; The feature of whole image is got up by the merging features of every sub regions; Therefore the feature of whole image comprises certain spatial positional information; The model obtaining after training, the weight that comprises every sub regions, according to the size of weight, this method can be selected representational region indirectly; Therefore the present invention does not need to carry out the concrete demarcation of concrete belt position, has reduced the work of artificial participation.

Accompanying drawing explanation

Fig. 1 is the securing band recognition methods process flow diagram that the present invention is based on many examples of margin maximization dictionary learning.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not intended to limit the present invention.In addition,, in each embodiment of described the present invention, involved technical characterictic just can not combine mutually as long as do not form each other conflict.

As shown in Figure 1, the securing band that the present invention is based on many examples of margin maximization dictionary learning detects with recognition methods and comprises the following steps:

(1) obtaining training data concentrates the proper vector of car bumper wind glass right half part image to represent

(1.1) to the concentrated car windshield image of training image, first intercept the right half part image of car windshield, car windshield right half part is extracted to the image block of different scale.

Described car windshield image is from vehicle front, to be the image that car windshield is taken pilothouse, in its right half part image, includes driver.

Particularly, the right side half image of the car windshield that intercepting is obtained, use respectively the moving window of three kinds of scale size (for example 48*48,72*72,96*96) in the enterprising line slip intercepting of image, moving window moving step length can be 1/3 length of window, picture material intercepting in moving window is each time preserved to the final like this image block that obtains several different scales $D_N=\{D_{11},\·\·\·,D_{1N_1};D_{21},\·\·\·,D_{{2N}_2};D_{31},\·\·\·,D_{{3N}_3}\},$ N wherein ₁, N ₂, N ₃the number that represents respectively three kinds of different scale hypograph pieces, N=N ₁+ N ₂+ N ₃the number that represents total image block.

(1.2) the right half part image of car windshield is extracted to yardstick invariant features conversion (Scale-invariant feature transform, SIFT) proper vector, and utilize default code book to encode to SIFT feature;

Described default code book can obtain by the SIFT proper vector that extraction obtains to training data is carried out to cluster.

(1.3) image block of the different scale obtaining in step (1.1) is zoomed to same scale size, and each image block after convergent-divergent is extracted respectively to proper vector, the proper vector of described image block is comprised of HOG proper vector, LBP proper vector, the SIFT proper vector of image block;

Then to each image block, the HOG that above-mentioned steps is obtained and LBP feature are stitched together as the proper vector of image block with SIFT histogram feature after normalization respectively, together by the merging features of all image blocks, obtain the proper vector of whole image finally;

Particularly, the image block of different scale can be zoomed in step (1.1) to a kind of in a plurality of yardsticks, generally zoom to smallest dimension.

(1.3.1) calculate histograms of oriented gradients (Histogram of oriented gradient, the HOG) proper vector of the image block after convergent-divergent:

Be specially, each pixel is asked to gradient (comprising size and Orientation), image is divided into cellule unit, add up the histogram of gradients of each cell unit, several cells unit (as 3*3) is formed to a region unit, the histogram of gradients of each region unit inner cell unit is stitched together as the HOG Feature Descriptor of this region unit, finally all proper vectors of zones of different piece is coupled together, the final HOG proper vector that obtains input picture piece represents;

In embodiments of the present invention, first by image gray processing, adopt Gamma correction method image to be carried out to the standardization of color space, object is the contrast that regulates image, reduce the shade of image local and the impact that illumination variation causes, can suppress the interference of noise simultaneously.Each pixel is asked to gradient (comprising size and Orientation), image is divided into cellule unit, add up the histogram of gradients of each cell unit, several cells unit (as 3*3) is formed to a region unit, the histogram of gradients of each region unit inner cell unit is stitched together as the HOG Feature Descriptor hog of this region unit _k=[h ₁, h ₂..., h _g], wherein g is the dimension of HOG proper vector, finally all proper vectors of zones of different piece is coupled together, the final HOG proper vector that obtains input picture piece represents hog=[hog ₁, hog ₂..., hog _nh], wherein nh represents the number of region unit.

(1.3.2) calculate local binary patterns (Local Binary Pattern, the LBP) proper vector of the image block after convergent-divergent;

Be specially, first image being divided to size is the cellule unit of 16*16, calculates the LBP feature of each cell unit, and the LBP proper vector that obtains image block that finally the LBP histogram of each cellule unit is stitched together represents;

In embodiments of the present invention, first image is divided into cellule unit (for example size is 16*16), for each pixel in each cellule unit, the gray-scale value of 8 adjacent pixels is compared with it, if surrounding pixel value is greater than center pixel value, the position of this pixel is marked as 1, otherwise be 0, like this, 8 points in 3*3 field are through relatively producing 8 bits, obtain the LBP value of this window center pixel, then calculate the histogram of each cellule unit, then this histogram is normalized, obtain like this LBP feature lbp of each cellule unit _j=[l ₁, l ₂..., l _l], wherein l represents the dimension of cellule element characteristic, the LBP proper vector that obtains image block that finally the LBP histogram of each cellule unit is stitched together represents lbp=[lbp ₁, lbp ₂..., lbp _nl], the number of cellule unit in nl presentation video wherein.

(1.3.3) calculate the coding SIFT feature of the image block after convergent-divergent;

Be specially, extract SIFT characteristic coordinates and be located at the feature in this image block, utilize word bag model (bag of words) and in conjunction with the code book of precognition, these features encoded, the histogram that coding is obtained is normalized the coding SIFT character representation sift that obtains image block, and the big or small C of code book is preset value.

(1.3.4) proper vector that obtains car windshield image finally represents

Be specially, above-mentioned steps (1.3.1) and the HOG feature obtaining in (1.3.2) and LBP feature are distinguished to normalization afterwards together with obtaining SIFT merging features in step (1.3.3), as the character representation F of image block _i=[hog _i; lbp _i; Sift _i], 1≤i≤N, wherein i represents i image block, N presentation video piece sum.Together by the merging features of all image blocks finally

(2) based on many examples of margin maximization dictionary learning

(2.1) training image can be divided into two class P={P ₊, P _-, P wherein ₊represent the car windshield image that comprises securing band, P _-represent not comprise the car windshield image of securing band.From above-mentioned steps (1), every piece image comprises several image blocks, and each image block can be F by the HOG feature of splicing, LBP feature and SIFT character representation _i.To two class images difference training classifiers, using the class image block sample in above-mentioned two class sample images as positive sample, another kind of image block sample is as negative sample.

(2.2), to positive sample and negative sample in step (2.1), based on margin maximization method, calculate the weight matrix of positive sample; Particularly, the image first aligning in sample carries out the label z that cluster obtains cluster _i, according to sample weight with to certainty ratio, extract positive sample, and all negative sample composition training datasets, by multiclass SVM, solve optimization problem and try to achieve weight matrix, by continuous iteration undated parameter, obtain optimum weight matrix W; Specific implementation step is as follows:

(2.2.1) align the image block sample feature F in sample _icarry out cluster, establish the middle calculation that K represents cluster, and according to cluster result to the given label z of each image block _i∈ 0,1 ..., K}, if F after cluster _ibelong to k class, z _i=k ∈ 1 ..., K}; If F _ibelong to negative class, z _i=0; For its label of image block sample in all negative samples, be z _i=0.

(2.2.2) definition weight matrix W=[w ₀, w ₁..., w _k], k ∈ 0,1 ..., K}, wherein w _kthe Clustering Model that represents k positive class, w ₀the Clustering Model that represents negative class, K represents cluster numbers; Definition sample weight p _ij, p _ijthat data in order to guarantee to extract from training data are positive sample, according to SVM decision function to proper vector F _ithe maximum difference belonging between positive class and negative class provides, definition p ^sfor extracting in proportion the positive sample in training data, generally p ^sfor fixed value, for example value is 0.7;

(2.2.3), under original state, establish the sample weight p of all positive classes _ijbe 1.First according to sample weight p _ijextract positive sample, then from the sample of selecting, extract p ^sratio, and all negative samples form training dataset D', wherein p ^srepresent to extract the ratio of sample, D' represents the new training dataset forming, and due to cluster label cicada, can solve following optimization problem by multiclass SVM and obtain W:

$\underset{W}{\min }\underset{k=0}{\overset{K}{\mathrm{\Σ}}}{\left|\right|w_k\left|\right|}^2+\mathrm{\λ}\underset{\mathrm{ij}}{\mathrm{\Σ}}\max (\mathrm{0,1}+w_{r_{\mathrm{ij}}}^T{x}_{\mathrm{ij}}-w_{z_{\mathrm{ij}}}^T{x}_{\mathrm{ij}}),$

Wherein r _ijexpression is to except this class, x _ijpeak response to all the other classes.

(2.2.4) upgrade sample weight and sample label, by step (2.2.2), obtain weight matrix and upgrade sample weight p _ij, according to upgrade the label of sample.

(2.2.5) repeat (2.2.3)～(2.2.4) process, circulation M time, wherein M is default cycle index, for example value is 5.

(2.3) positive sample and negative sample in step (2.1) are exchanged, the weight matrix that utilizes the positive sample obtaining after the method calculating exchange in step (2.2), is stitched together the weight matrix obtaining in the weight matrix obtaining and step (2.2) to obtain based on many examples of margin maximization dictionary here;

In this step, positive sample and negative sample in step (2.1) are exchanged, after exchange, positive sample in the past just becomes new negative sample, and negative sample in the past just becomes new positive sample, so calculate the weight matrix of the positive sample obtaining after exchanging here, in fact namely the weight matrix of the negative sample in calculation procedure (2.1).

(3) sorter of training of safety band identification

(3.1) utilize the dictionary obtaining in step (2) to encode and obtain the coding characteristic vector of image the proper vector of the concentrated car bumper wind glass right half part of the training data obtaining in step (1).Specific coding mode is:

(3.1.1) image feature vector and weight matrix are multiplied each other, obtain the response mapping for each disaggregated model.

(3.1.2) to each response mapping, the response of statistics coordinate within image range, and carry out non-maximum the inhibition, obtain the vector of a H dimension;

(3.1.3) utilize space pyramid model, former training image is divided into 2*2 and 4*4 totally 20 pieces, for each piece, repeat (3.1.1) and computation process (3.1.2), finally obtain the proper vector that 20 H tie up;

(3.1.4) vector above-mentioned two steps being obtained is stitched together, and obtains the vector of 21*H dimension, is exactly the final proper vector of this image;

(3.2) utilize the coding characteristic vector training svm classifier device of all images that obtain in training data.

Utilize training image training svm classifier device, can utilize packaged existing svm classifier device built-in function to be considered as a black box herein, input the proper vector of all images, export a svm classifier device that can judge whether with securing band.

(4) identify car windshield image to be identified

(4.1) according to the method in step (1), image to be identified is extracted to proper vector;

(4.2) according to the method that training data is encoded in step (3.1), treating the proper vector that recognition image obtained encodes;

(4.3) utilize the svm classifier device training in step (3), image to be identified is classified, and the result of output safety band identification.

Further, the present invention also provides a kind of securing band based on many examples of margin maximization dictionary learning to detect and recognition system, described system comprises car windshield image feature vector generation module, many examples of margin maximization dictionary learning module, sorter training module and car windshield picture recognition module to be identified, wherein:

Described car windshield right half part image feature vector generation module, for obtaining the proper vector of the car windshield right half part image of training image data centralization, represent, the representation module that specifically comprises different scale images piece acquisition module, car windshield right half part image SIFT characteristic extracting module, image block characteristics extraction module and characteristics of image, wherein:

Described different scale images piece acquisition module, for to the concentrated car windshield image of training image, first extracts the right half part of car windshield, and right half part car windshield is obtained to the image block under different scale;

Described car windshield SIFT characteristic extracting module, for car windshield image is extracted to SIFT feature, and the process of utilizing default code book to encode;

Described image block characteristics extraction module, extracts proper vector for the image block to the different scale acquiring.Specifically comprise image block convergent-divergent submodule, HOG proper vector calculating sub module, LBP proper vector calculating sub module and coding SIFT feature histogram calculating sub module.

Described image block convergent-divergent submodule, for the image block of different scale is zoomed to same yardstick, scales the images to three kinds of smallest dimension in yardstick conventionally;

Described HOG proper vector calculating sub module, the HOG proper vector of the image block acquiring for computed image piece convergent-divergent submodule;

Described LBP proper vector calculating sub module, the LBP proper vector of the image block acquiring for computed image piece convergent-divergent submodule;

Described coding SIFT feature histogram calculating sub module, the SIFT histogram feature that acquires image block for computed image piece convergent-divergent submodule is vectorial;

The character representation module of described image, for calculating the character representation of car windshield image, to every image, plays the merging features of all image blocks of presentation video the character representation that is used as image;

Described many examples of margin maximization dictionary learning module, be used for learning encoder dictionary, each class of concentrating for training data, trains respectively a multicategory classification device, obtain a disaggregated model, the weight in all sorter models is combined and obtains dictionary and represent;

Described sorter training module, for training of safety band recognition classifier, first carries out coding module to training data, and the disaggregated model that utilizes many examples of margin maximization dictionary learning module to obtain is encoded to training data; Finally utilize the coding characteristic vector training svm classifier device of all images that obtain in training image;

Described car windshield picture recognition module to be identified, be used for identifying image to be identified and whether have securing band, specifically comprise the extraction module of car windshield image block to be identified, the SIFT characteristic extracting 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:

Described car windshield image block extraction module to be identified, represents for obtaining the image block of car windshield right half part image;

The SIFT characteristic extracting module of described car windshield image, for obtaining the SIFT feature of car windshield right half part image, and utilizes word bag model in conjunction with the code book of precognition, SIFT feature to be encoded;

Described image block characteristics extraction module, for extracting the proper vector of presentation video piece;

Described image feature representation module, for being stitched together the image block characteristics of every piece image as the proper vector of image;

Described characteristics of image coding module, encodes to characteristics of image for the dictionary that utilizes the many learn-by-examples of margin maximization to obtain;

Described car windshield picture recognition module to be identified, for the securing band recognition classifier of utilizing sorter training module to train, classifies to image to be identified, and the recognition result of output safety band.

Particularly, the right half part of the car windshield that intercepting is obtained, use respectively the moving window of three kinds of scale size (for example 48*48,72*72,96*96) in the enterprising line slip intercepting of image, moving window moving step length is 1/3 length of window, picture material intercepting in moving window is each time preserved to the final like this image block that obtains several different scales $D_N=\{D_{11},\·\·\·,D_{1N_1};D_{21},\·\·\·,D_{{2N}_2};D_{31},\·\·\·,D_{{3N}_3}\},$ N wherein ₁, N ₂, N ₃the number that represents respectively three kinds of different scale hypograph pieces, N=N ₁+ N ₂+ N ₃the number that represents total image block.

Be specially, described HOG proper vector calculating sub module, for each pixel is asked to gradient (comprising size and Orientation), image is divided into cellule unit, add up the histogram of gradients of each cell unit, several cells unit (as 3*3) is formed to a region unit, the histogram of gradients of each region unit inner cell unit is stitched together as the HOG Feature Descriptor hog of this region unit _k=[h ₁, h ₂..., h _g], wherein g is the dimension of HOG proper vector, finally all proper vectors of zones of different piece is coupled together, the final HOG proper vector that obtains input picture piece represents hog=[hog ₁, hog ₂..., hog _nh], wherein nh represents the number of region unit.

Particularly, described LBP proper vector calculating sub module, is first divided into image the cellule unit of 16*16, calculates the histogram of each cellule unit, then this histogram is normalized, and obtains like this LBP feature lbp of each cellule unit _j=[l ₁, l ₂..., l _l], wherein l represents the dimension of cellule element characteristic, the LBP proper vector that obtains image block that finally the LBP histogram of each cellule unit is stitched together represents lbp=[lbp ₁, lbp ₂..., lbp _nl], the number of cellule unit in nl presentation video wherein.

Particularly, described SIFT coding histogram feature, for extracting SIFT characteristic coordinates, be located at the feature in this image block, utilize word bag model (bag of words) and in conjunction with the code book of precognition, these features encoded, the histogram that coding is obtained is normalized and obtains image block SIFT character representation sift, and the C size of code book is generally preset value.

Particularly, many examples of margin maximization dictionary learning module is used for carrying out following step, to generate encoder dictionary, first training data is divided into the sample that has securing band and the sample that there is no securing band according to artificial annotation results by data, using the class image block sample in above-mentioned two class sample images as positive sample, another kind of image block sample, as negative sample, aligns sample and negative sample and repeats following step:

(2.2.1) align the image block sample feature F in sample _icarry out cluster, establish the middle calculation that K represents cluster, and according to cluster result to the given label z of each image block _i∈ 0,1 ..., K}, if F after cluster _ibelong to k class, z _i=k ∈ 1 ..., K}; If F _ibelong to negative class, z _i=0; For its label of image block sample in all negative samples, be z _i=0;

(2.2.2) definition weight matrix W=[w ₀, w ₁..., w _k], k ∈ 0,1 ..., K}, wherein w _kthe Clustering Model that represents k positive class, w ₀the Clustering Model that represents negative class, K represents cluster numbers; Definition sample weight p _ij, p _ijthat data in order to guarantee to extract from training data are positive sample, according to SVM decision function to proper vector F _ithe maximum difference belonging between positive class and negative class provides, definition p ^sfor extracting in proportion the positive sample in training data;

(2.2.3), under original state, establish the sample weight p of all positive classes _ijbe 1, first according to sample weight p _ijextract positive sample, then from the sample of selecting, extract p ^sratio, and all negative samples form training dataset D', wherein p ^srepresent to extract the ratio of sample, D' represents the new training dataset forming, and solves following optimization problem obtain W by multiclass SVM:

$\underset{W}{\min }\underset{k=0}{\overset{K}{\mathrm{\Σ}}}{\left|\right|w_k\left|\right|}^2+\mathrm{\λ}\underset{\mathrm{ij}}{\mathrm{\Σ}}\max (\mathrm{0,1}+w_{r_{\mathrm{ij}}}^T{x}_{\mathrm{ij}}-w_{z_{\mathrm{ij}}}^T{x}_{\mathrm{ij}}),$

Wherein r _ijexpression is to except this class, x _ijpeak response to all the other classes;

(2.2.4) upgrade sample weight and sample label, with the weight matrix that step (2.2.2) obtains, upgrade sample weight p _ij; According to upgrade the label of sample;

(2.2.5) repeat (2.2.3)～(2.2.4) process, circulation M time, wherein M is default cycle index.For example value is 5.

Then above-mentioned positive sample and negative sample are exchanged, utilize said method to recalculate the weight matrix of the positive sample after exchange, two weight matrix that obtain are stitched together and are obtained based on many examples of margin maximization dictionary.

Claims (10)

1. the securing band recognition methods based on many examples of margin maximization dictionary learning, is characterized in that, described method comprises the steps:

(1) obtaining training data concentrates the proper vector of car bumper wind glass right half part image to represent:

(1.1) to the concentrated car windshield image of training image, first intercept the right half part image of car windshield, car windshield right half part is extracted to the image block of different scale; Wherein, described car windshield image is from vehicle front, to be the image that car windshield is taken pilothouse, in its right half part image, includes driver;

(1.2) the right half part image of car windshield is extracted to yardstick invariant features conversion (Scale-invariant feature transform, SIFT) proper vector, and utilize the code book of precognition to encode to SIFT feature;

(1.3) image block of the different scale obtaining in step (1.1) is zoomed to same scale size, and each image block after convergent-divergent is extracted respectively to proper vector, the proper vector of described image block is by the histograms of oriented gradients of image block (Histogram of oriented gradients, HOG) proper vector, local binary pattern (Local Binary Pattern, LBP) proper vector, SIFT proper vector form;

(2) based on many examples of margin maximization dictionary learning:

(2.1) every training image training image being concentrated, utilize artificial mark that training dataset is divided into the sample image that has securing band and the sample image that there is no securing band, using the class image block sample in above-mentioned two class images as positive sample, another kind of image block sample is as negative sample;

(2.2), to positive sample and negative sample in step (2.1), based on margin maximization method, obtain the weight matrix of positive sample; Particularly, the image first aligning in sample carries out the label z that cluster obtains cluster _i, according to sample weight with to certainty ratio, extract positive sample, and all negative sample composition training datasets, by multiclass SVM, solve optimization problem and try to achieve weight matrix, by continuous iteration undated parameter, obtain optimum weight matrix W;

(2.3) positive sample and negative sample in step (2.1) are exchanged, the weight matrix that utilizes the positive sample obtaining after the method calculating exchange in step (2.2), is stitched together the weight matrix obtaining in the weight matrix obtaining and step (2.2) to obtain based on many examples of margin maximization dictionary here;

(3) sorter of training of safety band identification:

(3.1) utilize the dictionary obtaining in step (2) to encode and obtain the coding characteristic vector of image the proper vector of the concentrated car bumper wind glass right half part of the training data obtaining in step (1);

(3.2) utilize the coding characteristic vector training svm classifier device of all images that obtain in training data;

(4) identify car windshield image to be identified

(4.1) according to the method in step (1), image to be identified is extracted to proper vector;

(4.2) according to the method in step (3.1), the proper vector having obtained is encoded;

(4.3) utilize the svm classifier device training in step (3), image to be identified is classified, and the result of output safety band identification.

2. the securing band recognition methods based on many examples of margin maximization dictionary learning according to claim 1, is characterized in that, described step (2.2) specifically comprises:

(2.2.1) align the image block sample feature F in sample _icarry out cluster, establish the middle calculation that K represents cluster, and according to cluster result to the given label z of each image block _i∈ 0,1 ..., K}, if F after cluster _ibelong to k class, z _i=k ∈ 1 ..., K}; If F _ibelong to negative class, z _i=0; For its label of image block sample in all negative samples, be z _i=0;

(2.2.2) definition weight matrix W=[w ₀, w ₁..., w _k], k ∈ 0,1 ..., K}, wherein w _kthe Clustering Model that represents k positive class, w ₀the Clustering Model that represents negative class, K represents cluster numbers; Definition sample weight p _ij, p _ijthat data in order to guarantee to extract from training data are positive sample, according to SVM decision function to proper vector F _ithe maximum difference belonging between positive class and negative class provides, definition p ^sfor extracting in proportion the positive sample in training data;

(2.2.3), under original state, establish the sample weight p of all positive classes _ijbe 1, first according to sample weight p _ijextract positive sample, then from the sample of selecting, extract p ^sratio, and all negative samples form training dataset D', wherein p ^srepresent to extract the ratio of sample, D' represents the new training dataset forming, and solves following optimization problem obtain W by multiclass SVM:

$\underset{W}{\min }\underset{k=0}{\overset{K}{\mathrm{\Σ}}}{\left|\right|w_k\left|\right|}^2+\mathrm{\λ}\underset{\mathrm{ij}}{\mathrm{\Σ}}\max (\mathrm{0,1}+w_{r_{\mathrm{ij}}}^T{x}_{\mathrm{ij}}-w_{z_{\mathrm{ij}}}^T{x}_{\mathrm{ij}}),$

Wherein r _ijexpression is to except the x of this class _ijpeak response to all the other classes;

(2.2.4) upgrade sample weight and sample label, with the weight matrix that step (2.2.2) obtains, upgrade sample weight p _ij; According to upgrade the label of sample;

(2.2.5) repeat (2.2.3)～(2.2.4) process, circulation M time, wherein M is default cycle index.

3. the securing band recognition methods based on many examples of margin maximization dictionary learning according to claim 1 and 2, is characterized in that, described step (3.1) is specially:

(3.1.1) by the image feature vector obtaining in step (1) with based on many examples of margin maximization dictionary, multiply each other, obtain the response mapping for each disaggregated model;

(3.1.2) to each response mapping, the response of statistics coordinate within image range, and carry out non-maximum the inhibition, obtain the vector of a H dimension;

(3.1.3) utilize space pyramid model, former training image is divided into 2*2 and 4*4 totally 20 pieces, for each piece, repeat (3.1.1) and computation process (3.1.2), finally obtain the proper vector that 20 H tie up;

(3.1.4) vector above-mentioned two steps being obtained is stitched together, and obtains the vector of 21*H dimension, is exactly the final proper vector of this image.

4. the securing band recognition methods based on many examples of margin maximization dictionary learning according to claim 1 and 2, it is characterized in that, described step (1.1) is specially: the right half part of the car windshield that intercepting is obtained, with the moving window of three kinds of scale size, in the enterprising line slip of image, intercept respectively, picture material intercepting in moving window is each time preserved to the final like this image block that obtains a plurality of different scales.

5. the securing band recognition methods based on many examples of margin maximization dictionary learning according to claim 1 and 2, is characterized in that, described step (1.3) is specially:

(1.3.1) calculate the HOG proper vector of the image block after convergent-divergent:;

(1.3.2) calculate the LBP proper vector of the image block after convergent-divergent;

(1.3.3) calculate the coding SIFT feature of the image block after convergent-divergent;

(1.3.4) by above-mentioned (1.3.1) and the HOG feature that (1.3.2) obtains in step and LBP feature respectively after normalization together with the coding SIFT merging features obtaining in step (1.3.3), as the character representation of image block.

6. the securing band recognition methods based on many examples of margin maximization dictionary learning according to claim 1 and 2, is characterized in that, described step (1.3.1) is specially:

Each pixel is asked to gradient, described gradient comprises size and Orientation, image is divided into cellule unit, add up the histogram of gradients of each cell unit, several cells unit is formed to a region unit, the histogram of gradients of each region unit inner cell unit is stitched together as the HOG Feature Descriptor of this region unit, finally all proper vectors of zones of different piece is coupled together, the final HOG proper vector that obtains input picture piece represents.

7. the securing band recognition methods based on many examples of margin maximization dictionary learning according to claim 1 and 2, is characterized in that, described step (1.3.2) is specially:

First image is divided into the cellule unit of 16*16, calculates the LBP feature of each cellule unit, the LBP proper vector that obtains image block that finally the LBP histogram of each cellule unit is stitched together represents.

8. the securing band recognition methods based on many examples of margin maximization dictionary learning according to claim 1 and 2, is characterized in that, described step (1.3.3) is specially:

Extract SIFT characteristic coordinates and be located at the feature in this image block, utilize word bag model and in conjunction with the code book of precognition, these features encoded, the histogram that coding is obtained is normalized the coding SIFT character representation sift that obtains image block, and the big or small C of code book is preset value.

9. the securing band recognition methods based on many examples of margin maximization dictionary learning according to claim 1 and 2, is characterized in that, described step (1.3.4) is specially:

Be specially, first to each image block, above-mentioned steps (1.3.1) and the HOG feature obtaining in (1.3.2) and LBP feature are distinguished to normalization afterwards together with obtaining SIFT merging features in step (1.3.3), as the character representation F of image block _i=[hog _i; lbp _i; Sift _i], 1≤i≤N, wherein i represents i image block, N presentation video piece sum; Together by the merging features of all image blocks finally

$F={\left\{F_i\right\}}_{i=1}^N.$

10. the securing band recognition system based on many examples of margin maximization dictionary learning, it is characterized in that, described system comprises car windshield image feature vector generation module, many examples of margin maximization dictionary learning module, training data is carried out to coding module, sorter training module and car windshield picture recognition module to be identified, wherein:

Described car windshield image feature vector generation module, for obtaining the proper vector of the car windshield image of training image data centralization, represent, the representation module that specifically comprises different scale images piece acquisition module, car windshield image SIFT characteristic extracting module, image block characteristics extraction module and characteristics of image, wherein:

Described different scale images piece acquisition module, for to the concentrated car windshield image of training image, first extracts the right half part of car windshield, and right half part car windshield is obtained to the image block under different scale;

Described car windshield SIFT characteristic extracting module, for car windshield image is extracted to SIFT feature, and utilizes word bag model in conjunction with the code book of precognition, SIFT feature to be encoded;

Described image block characteristics extraction module, extracts proper vector for the image block to the different scale acquiring; Specifically comprise image block convergent-divergent submodule, HOG proper vector calculating sub module, LBP proper vector calculating sub module and coding SIFT feature histogram calculating sub module;

Described image block convergent-divergent submodule, for the image block of different scale is zoomed to same yardstick, scales the images to three kinds of smallest dimension in yardstick conventionally;

Described HOG proper vector calculating sub module, the HOG proper vector of the image block acquiring for computed image piece convergent-divergent submodule;

Described LBP proper vector calculating sub module, the LBP proper vector of the image block acquiring for computed image piece convergent-divergent submodule;

Described coding SIFT feature histogram calculating sub module, the SIFT histogram feature that acquires image block for computed image piece convergent-divergent submodule is vectorial;

The character representation module of described image, for calculating the character representation of car windshield image, to every image, plays the merging features of all image blocks of presentation video the character representation that is used as image;

Described many examples of margin maximization dictionary learning module, be used for learning encoder dictionary, each class of concentrating for training data, trains respectively a multicategory classification device, obtain a disaggregated model, the weight in all sorter models is combined and obtains dictionary and represent;

Training data is carried out to coding module, for the disaggregated model that utilizes many examples of margin maximization dictionary learning module to obtain, training data is encoded;

Described sorter training module, for training of safety band recognition classifier, utilizes the coding characteristic vector training svm classifier device of all images that obtain in training image;

Described car windshield picture recognition module to be identified, be used for identifying image to be identified and whether have securing band, specifically comprise the extraction module of car windshield image block to be identified, the SIFT characteristic extracting 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:

Described car windshield image block extraction module to be identified, represents for obtaining the image block of car windshield right half part image;

The SIFT characteristic extracting module of described car windshield image, for obtaining the SIFT feature of car windshield right half part image, and encodes to it with known code book;

Described image block characteristics extraction module, for extracting the proper vector of presentation video piece;

Described image feature representation module, for being stitched together the image block characteristics of every piece image as the proper vector of image;

Described characteristics of image coding module, encodes to characteristics of image for the dictionary that utilizes many examples of margin maximization dictionary learning to obtain;

Described car windshield picture recognition module to be identified, for the securing band recognition classifier of utilizing sorter training module to train, classifies to image to be identified, and the recognition result of output safety band.