CN105868734A - Power transmission line large-scale construction vehicle recognition method based on BOW image representation model - Google Patents

Abstract

The invention discloses a power transmission line large-scale construction breakage-prevention large-scale construction vehicle recognition method based on a BOW image representation model. The method comprises the steps that median filtering is performed on a current image frame obtained by a camera; a Gaussian mixture modeling method is used for background modeling and foreground recognition; an intra-large-region texture-based method is used for eliminating shadows of a moving object; BOW model characteristics of each foreground region are extracted; the extracted characteristics are fed into a multi-classification SVM learned in advance for vehicle type recognition. According to the power transmission line large-scale construction breakage-prevention large-scale construction vehicle recognition method, the defect that other color-based methods can only detect vehicles of a certain specific color can be overcome, types of various large-scale construction vehicles can be detected, and the accuracy is high.

Description

Transmission line of electricity Large Construction vehicle identification method based on BOW characterization image model

Technical field

The invention belongs to transmission line of electricity external force damage prevention field, relate to a kind of transmission line of electricity Large Construction vehicle identification method based on BOW characterization image model.

Background technology

Whether the algorithm that on-line monitoring device for transmission line based on intelligent video analysis is used is identified being divided into two classes to detection target by it: a class only detects intrusion target and carries out unifying identifier, testing result is submitted to user and manually differentiates decision-making；The another kind of intrusion target that first detects, then uses the technology such as artificial intelligence, machine learning, pattern recognition that intrusion target is identified classification, the result of point good class is submitted to user.First kind method is normally only used for the situation that foreground moving object is few, and when prospect moving object is more, real effective target can be flooded by its a large amount of invalid targets detected.And technology is required higher by Equations of The Second Kind method.

First kind algorithm difficult point is that moving target detects.Moving object detection is divided can be divided three classes by principle: optical flow method, frame differential method and background subtraction method.Optical flow method is based on the dividing method scape relief method detecting the estimation of optical flow field.Optical flow method is that its amount of calculation is huge, and algorithm is the most sufficiently complex, poor real based on the dividing method detecting the estimation of optical flow field, to transmission line of electricity big machinery intrusion detection inapplicable.Frame differential method uses the Differential Detection between image sequence to determine that moving target, algorithm the most easily realize, but when target travel slowly or speed quickly time, the target of extraction can not be close to its true shape.Background subtraction method is the moving object detection algorithm of current main-stream, and it is that video image and background reference image are carried out difference, and foreground point is i.e. moving target.Its processing speed is fast, very strong to the adaptability of different scenes, can relatively accurately extract the true shape of target.

First Equations of The Second Kind algorithm is also required to carry out moving object detection, and then intrusion target is identified classification, and its difficult point essentially consists in target recognition, i.e. judges whether moving target is the Large Construction vehicle such as crane, cement pump truck.Document [realization of the intelligent early-warning function that anti-big machinery external force is destroyed in transmission line of electricity] uses background subtraction method to carry out background detection, and uses color as the characteristics of image of big machinery, it is modeled.Carry out arm detection according to arm region area, eccentricity and compactness, use HOUGH transformation calculations arm to stretch angle, send alarm according to the distance between transmission line of electricity.Document [Intelligent Measurement of moving target and identification in transmission line of electricity monitoring system] uses feature based on color to identify big machinery in foreground target, then location crane wheel, use area, eccentricity and 3 indexs of compactness as the input of grader characteristic vector, identify arm, and and then hazard recognition behavior.Method in this two documents all uses color characteristic to carry out target recognition, it is believed that the color of crane is yellow.In actual applications, when the crane of non-yellow occur, its effect is the most very poor.

Summary of the invention

For the problem overcoming prior art to exist, it is an object of the invention to provide a kind of transmission line of electricity Large Construction vehicle identification method based on BOW characterization image model, the method can detect the classification of various Large Construction vehicle, owing to using the feature insensitive to color, it can overcome the shortcoming that can only detect certain color vehicle specific that other method based on color exists, and accuracy is high.

The purpose of the present invention is achieved through the following technical solutions:

A kind of transmission line of electricity Large Construction vehicle identification method based on BOW characterization image model, comprises the following steps:

(1) current image frame obtaining photographic head carries out medium filtering；

(2) method of Gaussian modeling is used to carry out background modeling and prospect identification；

(3) in using big region, method based on texture eliminates the shade of moving target；

(4) the BOW aspect of model of each foreground area is extracted；

(5) the many classification SVM support vector machine feature feeding of extraction succeeded in school in advance carry out vehicle cab recognition.

Further, the described current image frame obtaining photographic head carries out median filtering step and is:

First image is become RGB color from YUV color space conversion, converts thereof into gray-scale map further, then for from the second row to row second from the bottom, from each pixel f of secondary series to row second from the bottom, (x, y), by 8 pixel f (x-1 around it and it, y-1), and f (x-1, y), f (x-1, y+1), f (x, y-1), f (x, y+1), f (x+1, y-1), f (x+1, y), f (x+1, y+1) is arranged from small to large by gray value, take the value of that number middle as f (x, value y).

The described method using Gaussian modeling carries out the step of background modeling and prospect identification:

(1) background modeling:

Setting mixed Gauss model to be made up of 5 Gaussian functions, first in model, the average of each Gaussian function, variance, weights are both configured to 0, i.e. initialization model matrix parameter.

30 frames in video are used to be used for training mixed Gauss model.For each pixel, set up its mixed Gauss model.When first pixel is come, fixing initial mean value, variance, and each weights are set for it and are disposed as 0.2.

During non-first frame training, when tail pixel value, with the average ratio of the most existing Gaussian function relatively, if the value of this pixel and the equal value difference of its model are in the variance of 3 times, then it is assumed that this point belongs to this Gaussian function.Now it is updated by equation below:

${\widehat{\mathrm{\π}}}_m\←{\widehat{\mathrm{\π}}}_m+\mathrm{\α}(o_m^{\left(t\right)}-{\widehat{\mathrm{\π}}}_m),$

Wherein, It is 1.

When the difference of this pixel point value and average is not in its variance of 3 times, takeIt is 0.

When after the frame number 30 arriving training, carry out the different pixels point adaptive selection of mixed Gauss model number.First divided by variance, each Gauss is sorted from big to small with weights, then choose B, foremost Gauss so that it is meet, wherein

$B=\underset{b}{\arg \min }({\mathrm{\Σ}}_{m=1}^b{\widehat{\mathrm{\π}}}_m>(1-c_f\left)\right),$

Then accumulation weight is normalized to 1.

(2) prospect identification: for the dynamic picture frame of complexity, superposition to multiple Gauss distribution of the different weights that each pixel is learnt by upper step models, when the value of current pixel point and the mean bias of modeling are in 2.5 σ, then it is assumed that this pixel belongs to background, otherwise belongs to prospect.

In described big region, the step of method based on texture elimination moving target shade is:

(41) use the preselected shadows pixels of following formula (x, y):

SP value is to represent when 1 that current pixel is shadows pixels, is to represent when 0 not to be.In above formula, α=0.8, β=0.94.I_t(x, y) V is t picture frame (x, y) value of the V component of position, B_t(x, y) V is t background image (x, y) value of the V component of position using Gaussian modeling to obtain；I_t(x, y) H is t picture frame (x, y) value of the H component of position, B_t(x, y) H is t moment background image (x, y) value of the H component of position using Gaussian modeling to obtain；I_t(x, y) S is t picture frame (x, y) value of the S component of position, B_t(x, y) S is background image (x, y) value of the S component of position using Gaussian modeling t to obtain.FPM^TLIt it is foreground image areas flag bit.

(42) all connected domains of extraction are concentrated from preselected shadows pixels.

(43) for each connected domain, calculate each pixel p=(x, y) gradient direction at place and gradient-norm:

$|\▿p|=\sqrt{\▿x^2+\▿y^2}$

${\mathrm{\θ}}_p=arctan2(\▿y/\▿x)$

WhereinFor vertical gradient, i.e. | f (x, y)-f (x+1, y) |,For horizontal direction gradient, i.e. | f (x, y)-f (x, y+1) |, (x, is y) that image is at point (x, y) gray value at place to f, f (x, y+1) is the image gray value at point (x, y+1) place；Wherein θ_pIn the range of [-π, π], the most only retain | θ_p| value more than a certain threshold tau_mThose pixels.

(44) for each pixel p remaining=(x, y), the gradient direction calculating present image and background image is poor

${\mathrm{\Δ\θ}}_p=arccos\frac{{\▿}_x^F{\▿}_x^B+{\▿}_y^F{\▿}_y^B}{{\left\{({\▿}_x^{F^2}+{\▿}_y^{F^2})({\▿}_x^{B^2}+{\▿}_y^{B^2})\right\}}^{\frac{1}{2}}}$

Represent in present image at pixel p the Grad of horizontal direction,Represent the Grad of horizontal direction at pixel p in background image,Represent the Grad of vertical direction at pixel p in present image,Represent the Grad of vertical direction at p in background image.

(45) present image and the gradient direction dependency of background image are estimated with following formula

$c=\left\{\underset{p=1}{\overset{n}{\Σ}}H({\mathrm{\τ}}_a-{\mathrm{\Δ\θ}}_p)\right\}/n$

Wherein n is the selected pixels number of candidate shadow region, and H () is unit progression function, and i.e. if differential seat angle is less than or equal to threshold tau a, then the value of function is 1, is otherwise 0.C is the ratio of pixel count similar with gradient direction in background image in present image in region.As c compares threshold tau_cGreatly, then it is assumed that candidate region is shadow region.

The step of the described BOW aspect of model extracting each foreground area is:

(51) local shape factor

Divide image with the uniform grid (grid) that the length of side is 8 pixels, the block of 4 grid protocol extracts SIFT and describes son.Each piece is the rectangle of 16 × 16 pixels, comprises 4 × 4 buckets, and each bucket is 4 × 4 pixels.Calculating the gradient information in 8 directions in each bucket, the most each piece is characterized by 4 × 4 × 8=128 dimensional vector.The step-length that block moves is 8 pixels.

(52) study visual dictionary

The Dense SIFT local feature using KMEANS method to go out all image zooming-out in training storehouse clusters.Each cluster centre is i.e. a visual vocabulary in dictionary, and all visual vocabularies form a visual dictionary.

(53) characteristic quantification coding and spatial pyramid

After obtaining visual dictionary, then each feature of image can be mapped on certain word in dictionary, then each visual word occurrence number on this image in statistics dictionary, thus obtains the histogram vectors of the phenogram picture of one 300 dimension.

Use multiple dimensioned method of partition, image is divided into three layers of pyramid of 1 × 1,2 × 2,4 × 4 space separatings, adds up the feature of each sub-block respectively, finally the merging features of all pieces is got up, form complete feature.So feature of piece image is 6300 dimensions.

(54) train classification models predicting

Image in all training sets is used above 1)-3) method in step extracts feature, uses SVM support vector machine to be trained the characteristic vector of training set, it is thus achieved that the disaggregated model of Large Construction vehicle；Kernel function form such as following formula:

$k(x,y)=\underset{k=1}{\overset{\mathrm{\π}}{\Σ}}min(x_k,y_k)$

The step carrying out vehicle cab recognition in the described many classification SVM support vector machine feature feeding of extraction succeeded in school in advance is:

For the foreground area after each removal shade detected, use the Dense SIFT feature of aforementioned extraction image, and carry out feature after quantization is quantified according to visual dictionary, the feature in region is obtained again through pyramid coding, send in the SVM classifier trained and classify, as it is categorized as crane, excavator or cement pump truck, then it is assumed that be to have potential outer broken possible Large Construction vehicle, otherwise it is assumed that not outer broken probability.

Compared with prior art, the present invention identifies transmission line of electricity Large Construction vehicle based on BOW characterization image model, use the feature insensitive to color, overcome the shortcoming that can only detect certain color vehicle specific that in prior art, method based on color exists, can detect the classification of various Large Construction vehicles, accuracy is high.

Accompanying drawing explanation

Fig. 1 is Large Construction vehicle identification method flow chart based on Computer Vision；

Detailed description of the invention

In order to be better understood from technical scheme, below in conjunction with accompanying drawing 1, the invention will be further described.

Fig. 1 describes a kind of transmission line of electricity external force damage prevention Large Construction vehicle identification method based on Computer Vision, comprises the following steps:

(1) current image frame obtaining photographic head carries out medium filtering；Step is as follows:

First image is become RGB color from YUV color space conversion, converts thereof into gray-scale map further, then for from the second row to row second from the bottom, from each pixel f of secondary series to row second from the bottom, (x, y), by 8 pixel f (x-1 around it and it, y-1), and f (x-1, y), f (x-1, y+1), f (x, y-1), f (x, y+1), f (x+1, y-1), f (x+1, y), f (x+1, y+1) is arranged from small to large by gray value, take the value of that number middle as f (x, value y).

(2) method of Gaussian modeling is used to carry out background modeling and prospect identification；Step is as follows:

1) background modeling:

Setting mixed Gauss model to be made up of 5 Gaussian functions, first in model, the average of each Gaussian function, variance, weights are both configured to 0, i.e. initialization model matrix parameter.

30 frames in video are used to be used for training mixed Gauss model.For each pixel, set up its mixed Gauss model.When first pixel is come, fixing initial mean value, variance, and each weights are set for it and are disposed as 0.2.

During non-first frame training, when tail pixel value, with the average ratio of the most existing Gaussian function relatively, if the value of this pixel and the equal value difference of its model are in the variance of 3 times, then it is assumed that this point belongs to this Gaussian function.Now it is updated by equation below:

${\widehat{\mathrm{\π}}}_m\←{\widehat{\mathrm{\π}}}_m+\mathrm{\α}(o_m^{\left(t\right)}-{\widehat{\mathrm{\π}}}_m),$

Wherein, It is 1.

When the difference of this pixel point value and average is not in its variance of 3 times, takeIt is 0.

When after the frame number 30 arriving training, carry out the different pixels point adaptive selection of mixed Gauss model number.

First divided by variance, each Gauss is sorted from big to small with weights, then choose B, foremost Gauss so that it is meet, wherein

$B=\underset{b}{\arg \min }({\mathrm{\Σ}}_{m=1}^b{\widehat{\mathrm{\π}}}_m>(1-c_f\left)\right),$

Then accumulation weight is normalized to 1.

2) prospect identification: for the dynamic picture frame of complexity, superposition to multiple Gauss distribution of the different weights that each pixel is learnt by upper step models, when the value of current pixel point and the mean bias of modeling are in 2.5 σ, then it is assumed that this pixel belongs to background, otherwise belongs to prospect.

(3) in using big region, method based on texture eliminates the shade of moving target；Step is as follows:

41) use the preselected shadows pixels of following formula (x, y):

SP value is to represent when 1 that current pixel is shadows pixels, is to represent when 0 not to be.In above formula, α=0.8, β=0.94.I_t(x, y) V is t picture frame (x, y) value of the V component of position, B_t(x, y) V is t background image (x, y) value of the V component of position using Gaussian modeling to obtain；I_t(x, y) H is t picture frame (x, y) value of the H component of position, B_t(x, y) H is t background image (x, y) value of the H component of position using Gaussian modeling to obtain；I_t(x, y) S is t picture frame (x, y) value of the S component of position, B_t(x, y) S is background image (x, y) value of the S component of position using Gaussian modeling t to obtain.FPM^TLIt it is foreground image areas flag bit.

42) all connected domains of extraction are concentrated from preselected shadows pixels.

43) for each connected domain, calculate each pixel p=(x, y) gradient direction at place and gradient-norm:

$|\▿p|=\sqrt{\▿x^2+\▿y^2}$

${\mathrm{\θ}}_p=arctan2(\▿y/\▿x)$

WhereinFor vertical gradient, i.e. | f (x, y)-f (x+1, y) |,For horizontal direction gradient, i.e. | f (x, y)-f (x, y+1) |, (x, is y) that image is at point (x, y) gray value at place to f, f (x, y+1) is the image gray value at point (x, y+1) place；Wherein θ_pIn the range of [-π, π], the most only retain | θ_p| value more than a certain threshold tau_mThose pixels.

44) for each pixel p remaining=(x, y), the gradient direction calculating present image and background image is poor

${\mathrm{\Δ\θ}}_p=arccos\frac{{\▿}_x^F{\▿}_x^B+{\▿}_y^F{\▿}_y^B}{{\left\{({\▿}_x^{F^2}+{\▿}_y^{F^2})({\▿}_x^{B^2}+{\▿}_y^{B^2})\right\}}^{\frac{1}{2}}}$

Represent in present image at pixel p the Grad of horizontal direction,Represent the Grad of horizontal direction at pixel p in background image,Represent the Grad of vertical direction at pixel p in present image,Represent the Grad of vertical direction at p in background image.

45) present image and the gradient direction dependency of background image are estimated with following formula

$c=\left\{\underset{p=1}{\overset{n}{\Σ}}H({\mathrm{\τ}}_a-{\mathrm{\Δ\θ}}_p)\right\}/n$

Wherein n is the selected pixels number of candidate shadow region, and H () is unit progression function, and i.e. if differential seat angle is less than or equal to threshold tau a, then the value of function is 1, is otherwise 0.C is the ratio of pixel count similar with gradient direction in background image in present image in region.As c is bigger than threshold tau c, then it is assumed that candidate region is shadow region.

(4) the BOW aspect of model of each foreground area is extracted；Step is as follows:

51) local shape factor

Divide image with the uniform grid (grid) that the length of side is 8 pixels, the block of 4 grid protocol extracts SIFT and describes son.Each piece is the rectangle of 16 × 16 pixels, comprises 4 × 4 buckets, and each bucket is 4 × 4 pixels.Calculating the gradient information in 8 directions in each bucket, the most each piece is characterized by 4 × 4 × 8=128 dimensional vector.The step-length that block moves is 8 pixels.

52) study visual dictionary

The Dense SIFT local feature using KMEANS method to go out all image zooming-out in training storehouse clusters.Each cluster centre is i.e. a visual vocabulary in dictionary, and all visual vocabularies form a visual dictionary.

53) characteristic quantification coding and spatial pyramid

After obtaining visual dictionary, then each feature of image can be mapped on certain word in dictionary, then each visual word occurrence number on this image in statistics dictionary, thus obtains the histogram vectors of the phenogram picture of one 300 dimension.

Use multiple dimensioned method of partition, image is divided into three layers of pyramid of 1 × 1,2 × 2,4 × 4 space separatings, adds up the feature of each sub-block respectively, finally the merging features of all pieces is got up, form complete feature.So feature of piece image is 6300 dimensions.

54) train classification models predicting

Image in all training sets is used above 1)-3) method in step extracts feature, uses SVM support vector machine to be trained the characteristic vector of training set, it is thus achieved that the disaggregated model of Large Construction vehicle；Kernel function form such as following formula:

$k(x,y)=\underset{k=1}{\overset{\mathrm{\π}}{\Σ}}min(x_k,y_k)$

(5) the many classification SVM support vector machine feature feeding of extraction succeeded in school in advance carry out vehicle cab recognition.Step is:

For the foreground area after each removal shade detected, use the Dense SIFT feature of aforementioned extraction image, and carry out feature after quantization is quantified according to visual dictionary, the feature in region is obtained again through pyramid coding, send in the SVM classifier trained and classify, as it is categorized as crane, excavator or cement pump truck, then it is assumed that be to have potential outer broken possible Large Construction vehicle, otherwise it is assumed that not outer broken probability.

Embodiment

The image obtained for monitoring camera, is handled as follows:

(1) current image frame obtaining monitoring camera carries out medium filtering.

(2) use the method for Gaussian modeling to carry out background modeling and prospect identification, then image is carried out shadow Detection elimination.

(3) the BOW aspect of model of each foreground area is extracted.

(4) the many classification SVM support vector machine feature feeding of extraction succeeded in school in advance carry out vehicle cab recognition, result marks all Large Construction vehicles.

Claims (6)

1. a transmission line of electricity Large Construction vehicle identification method based on BOW characterization image model, it is special Levy and be to comprise the following steps:

(1) current image frame obtaining photographic head carries out medium filtering；

(2) method of Gaussian modeling is used to carry out background modeling and prospect identification；

(3) in using big region, method based on texture eliminates the shade of moving target；

(4) the BOW aspect of model of each foreground area is extracted；

(5) the many classification SVM support vector machine feature feeding of extraction succeeded in school in advance carry out vehicle Identify.

Transmission line of electricity Large Construction vehicle based on BOW characterization image model the most according to claim 1 Recognition methods, it is characterised in that: the described current image frame obtaining photographic head carries out median filtering step and is:

First image is become RGB color from YUV color space conversion, convert thereof into ash further Degree figure, then for from the second row to row second from the bottom, from each pixel of secondary series to row second from the bottom (x, y), by its gray value f (x, the gray value of 8 pixels y) and around it (x-1, y), f (x-1, y+1), f (x, y-1), f (x, y+1), (x+1, y), f (x+1, y+1) is by ash for f (x+1, y-1), f for f (x-1, y-1), f Angle value arranges from small to large, takes the value of that number middle as pixel (x, gray value f (x, value y) y).

Transmission line of electricity Large Construction vehicle based on BOW characterization image model the most according to claim 1 Recognition methods, it is characterised in that: the described method using Gaussian modeling carries out background modeling and prospect is known Other step is:

(1) background modeling:

Set mixed Gauss model to be made up of 5 Gaussian functions, first the average of each Gaussian function in model, Variance, weights are both configured to 0, i.e. initialization model matrix parameter；

30 frames in video are used to be used for training mixed Gauss model；For each pixel, set up it and mix Close Gauss model, when first pixel is come, fixing initial mean value, variance, and each weights are set for it It is disposed as 0.2；

During non-first frame training, when tail pixel value, with the average of the most existing Gaussian function Relatively, if the value of this pixel and the equal value difference of its model are in the variance of 3 times, then it is assumed that this point belongs to this height This function, is now updated by equation below:

${\widehat{\mathrm{\π}}}_m\←{\widehat{\mathrm{\π}}}_m+\mathrm{\α}(o_m^{\left(t\right)}-{\widehat{\mathrm{\π}}}_m),$

Wherein,It is 1；

When the difference of this pixel point value and average is not in its variance of 3 times, takeIt is 0；

When after the frame number 30 arriving training, carry out the different pixels point adaptive selection of mixed Gauss model number； First divided by variance, each Gauss is sorted from big to small with weights, then chooses B, foremost Gauss, It is made to meet, wherein

$B=\underset{b}{\arg \min }({\mathrm{\Σ}}_{m=1}^b{\widehat{\mathrm{\π}}}_m>(1-c_f\left)\right),$

Then accumulation weight is normalized to 1；

2) prospect identification: for the dynamic picture frame of complexity, to each pixel by the different power of upper step study The superposition of multiple Gauss distribution of value models, when the value of current pixel point and the mean bias of modeling are at 2.5 σ Time interior, then it is assumed that this pixel belongs to background, otherwise belong to prospect.

Transmission line of electricity Large Construction vehicle based on BOW characterization image model the most according to claim 1 Recognition methods, it is characterised in that: in described big region, method based on texture eliminates the step of moving target shade Suddenly it is:

(41) use the preselected shadows pixels of following formula (x, y):

SP value is to represent when 1 that current pixel is shadows pixels, is to represent when 0 not to be.In above formula, α=0.8, β=0.94； I_t(x, y) V is t picture frame (x, y) value of the V component of position, B_t(x, y) V is to use mixed Gaussian T background image (x, y) value of the V component of position that modeling obtains；I_t(x, y) H is t image Frame (x, y) value of the H component of position, B_t(x, y) H is the t background using Gaussian modeling to obtain Image (x, y) value of the H component of position；I_t(x, y) S is that (x, y) S of position divides t picture frame The value of amount, B_t(x, y) S is t background image (x, y) S of position using Gaussian modeling to obtain The value of component.FPM^TLIt it is foreground image areas flag bit；

(42) all connected domains of extraction are concentrated from preselected shadows pixels；

(43) for each connected domain, calculate each pixel p=(x, y) gradient direction at place and gradient-norm:

$|\▿p|=\sqrt{{\▿x}^2+{\▿y}^2}$

${\mathrm{\θ}}_p=\arctan 2(\▿y/\▿x)$

WhereinFor horizontal direction gradient, i.e. | f (x, y)-f (x, y+1) |,For vertical gradient, i.e. | f (x, y)-f (x+1, y) |, (x is y) that at point, (x, y) gray value at place, f (x, y+1) is that image is at point to image to f The gray value at (x, y+1) place；Wherein θ_pIn the range of [-π, π], the most only retain | θ_p| value be more than A certain threshold tau_mThose pixels；

(44) for each pixel p remaining=(x, y), the gradient direction calculating present image and background image is poor

${\mathrm{\Δ\θ}}_p=\arccos \frac{{\▿}_x^F{\▿}_x^B+{\▿}_y^F{\▿}_y^B}{{\left\{({\▿}_x^{F^2}+{\▿}_y^{F^2})({\▿}_x^{B^2}+{\▿}_y^{B^2})\right\}}^{\frac{1}{2}}}$

Represent in present image at pixel p the Grad of horizontal direction,Represent in background image The Grad of horizontal direction at pixel p,Represent the Grad of vertical direction at pixel p in present image,Represent the Grad of vertical direction at p in background image；

(45) present image and the gradient direction dependency of background image are estimated with following formula

$c=\left\{\underset{p=1}{\overset{n}{\Σ}}H({\mathrm{\τ}}_a-{\mathrm{\Δ\θ}}_p)\right\}/n$

Wherein n is the selected pixels number of candidate shadow region, and H () is unit progression function, i.e. Such as differential seat angle Δ θ_pLess than or equal to threshold tau a, then the value of function is 1, is otherwise 0；C is currently in region The ratio of pixel count similar with gradient direction in background image in image, as c is bigger than threshold tau c, then it is assumed that Candidate region is shadow region.

Transmission line of electricity Large Construction vehicle based on BOW characterization image model the most according to claim 1 Recognition methods, it is characterised in that: the step of the described BOW aspect of model extracting each foreground area is:

(51) local shape factor

Divide image with the uniform grid that the length of side is 8 pixels, the block block of 4 grid protocol extracts SIFT Son is described；Each piece is the rectangle of 16 × 16 pixels, comprises 4 × 4 buckets, and each bucket is 4 × 4 pixels； Calculating the gradient information in 8 directions in each bucket, the most each piece is characterized by 4 × 4 × 8=128 dimensional vector； The step-length that block moves is 8 pixels；

(52) study visual dictionary

The Dense SIFT local feature using KMEANS method to go out all image zooming-out in training storehouse is carried out Cluster, each cluster centre is i.e. a visual vocabulary in dictionary, and all visual vocabularies form a visual word Allusion quotation；

(53) characteristic quantification coding and spatial pyramid

After obtaining visual dictionary, then each feature of image can be mapped on certain word in dictionary, then unites Meter dictionary in each visual word occurrence number on this image, thus obtain one 300 dimension phenogram as Histogram vectors；

Use multiple dimensioned method of partition, image is divided into three layers of gold of 1 × 1,2 × 2,4 × 4 space separatings Word tower, adds up the feature of each sub-block respectively, is finally got up by the merging features of all pieces, form complete spy Levy.So feature of piece image is 6300 dimensions；

(54) train classification models predicting

Image in all training sets is used above 51) to 53) method in step extracts feature, uses SVM The characteristic vector of training set is trained by support vector machine, it is thus achieved that the disaggregated model of Large Construction vehicle；Core letter Number form formula such as following formula:

$k(x,y)=\underset{k=1}{\overset{\mathrm{\π}}{\Σ}}min(x_k,y_k).$

Transmission line of electricity Large Construction vehicle based on BOW characterization image model the most according to claim 1 Recognition methods, it is characterised in that: described props up many classification SVM that the feature feeding of extraction succeeds in school in advance Hold and vector machine carries out the step of vehicle cab recognition be:

For the foreground area after each removal shade detected, use the Dense of aforementioned extraction image SIFT feature, and carry out feature after quantization is quantified according to visual dictionary, then obtain district through pyramid coding The feature in territory, sends in the SVM classifier trained and classifies, as it is categorized as crane, excavator Or cement pump truck, then it is assumed that it is to have potential outer broken possible Large Construction vehicle, otherwise it is assumed that outer broken possible Property.