CN105427276A - Camera detection method based on image local edge characteristics - Google Patents

Abstract

The invention belongs to the image processing field, more specifically relating to a method of determining whether a camera is shielded through image local edge characteristics, and provides a camera detection method based on image local edge characteristics, comprising: performing gray scale transformation on an RGB image I shot by a camera to obtain an image L; successively utilizing a gauss filter and a non-linear diffusion filter to process the image L to construct a non-linear multi-scale space; calculating the Hessian matrix response image LiHessian of each image Li in the space; utilizing a non-maximum value inhibition algorithm of a 3*3*3 neighborhood to position LiHessian local edge characteristic points; and finally counting the number of the characteristic points to compare with a threshold, and accurately determining a camera shielding case. The method can adapt to illumination change, accurately determining whether the camera is shielded under various illumination conditions, meanwhile employ a gray value mean square deviation to eliminate the false alarm caused by network and equipment faults, and reduce a rate of false alarm.

Description

A kind of camera detection method based on image local edge feature

Technical field

The invention belongs to image processing field, be specifically related to utilize image local edge feature to judge the method whether camera is blocked.

Background technology

Along with the development of society and the progress of science and technology, supervisory system is used in the every field that people produce, live widely.Camera is in supervisory system foremost, and be easily subject to extraneous interference, modal is exactly that camera is blocked.Although prior art can judge this situation preferably; but due to the limitation of existing algorithm and the complicacy of field condition; especially the impact of illumination variation, often can make equipment occur deviation for the judgement of image, causes the generation of reporting and failing to report phenomenon by mistake.Therefore must improve adaptability and the reliability of intelligent analysis process, accurately could reflect the truth of scene.

Summary of the invention

The object of this invention is to provide a kind of camera detection method based on image local edge feature, the change of illumination can be adapted to, can judge whether camera is blocked accurately at various illumination condition, the wrong report that network failure and equipment failure cause can be got rid of simultaneously, reduce rate of false alarm.

For achieving the above object, the technical solution adopted in the present invention is: a kind of camera occlusion detection method based on image local edge feature, is characterized in that: comprise the following steps:

A, to camera shooting RGB image I carry out greyscale transformation, obtain gray level image L, and calculate the mean square deviation ω of each grey scale pixel value in gray level image L;

B, judge whether the value of ω equals 0, if ω=0, then show network failure or equipment failure, terminate program; If ω > 0, then carry out the extraction of image local Edge Feature Points, comprise the following steps:

B1, adopt Gaussian filter to carry out filtering process to gray level image L, construct the multiscale space with N width image, described multiscale space is made up of O group image, often organizes image and has S sublayer, N=O × S; The scale parameter σ of every tomographic image _imark respectively by sequence number o and s, scale parameter σ _icalculate according to formula (1):

o ∈ [0 ..., O-1], s ∈ [0 ..., S-1], i ∈ [0 ..., N], σ ₀be the initial baseline value of scale parameter, be defaulted as 1.6;

B2, employing Nonlinear diffusion filtering device carry out DIFFUSION TREATMENT to the every tomographic image in multiscale space, generate non-linear multiscale space, wherein the image L of the bottom ⁰=L _σ, L _σwith being of a size of 9 × 9, standard deviation be 1.6 gaussian kernel and described gray level image L carry out convolution after obtain; Remainder layer image generates evolution graph as L according to following formula (2) ⁱ:

$L^{i+1}={(I-\mathrm{\τ}\underset{l=1}{\overset{m}{\Σ}}{A}_l(L^i\left)\right)}^{-1}{L}^i,$

Wherein A _lrepresent image L ⁱconductance matrix on each dimension l, I is original image, and τ is time step, τ=t _i+1-t _i, t _ifor the scale parameter in units of the time, i ∈ [0 ..., N];

B3, according to the every secondary evolution graph generated in b2 step as L ⁱcalculate the Hessian matrix response image L of its correspondence ⁱ _hessian, according to formula (3):

L ⁱ _hessian=σ ² _{i, norm}(L _xx ⁱl _yy ⁱ-L _xy ⁱl _xy ⁱ), wherein, L ⁱ _xxfor evolution graph is as L ⁱsecond order in x direction is reciprocal, L ⁱ _yyfor evolution graph is as L ⁱsecond order in y direction is reciprocal, L ⁱ _xyfor evolution graph is as L ⁱsecond-order mixed partial derivative, σ _{i, norm}for L ⁱthe round values of corresponding yardstick;

The non-maximal value Restrainable algorithms of b4, employing 3 × 3 × 3 neighborhoods, locates each Hessian response image L ⁱ _hessianin local edge unique point;

The quantity n of the local edge unique point that c, statistics are extracted, by n and threshold value n ₀compare, threshold value n ₀for the local feature of image captured when camera does not block is counted; If n≤n ₀, then send camera and to be blocked alarm.

Beneficial effect of the present invention: carry out numerical value contrast by the quantity extracting picture local edge unique point, and without the need to carrying out the contrast of plurality of pictures, greatly reduce calculated amount, improve computing velocity; And these local edge features possess good yardstick and rotational invariance, the situation such as change, visual angle change conversion, image scaling for illumination also keeps certain unchangeability, overcome some defects that prior art exists, the particularly wrong report that causes of illumination variation, local edge feature keeps good unchangeability to illumination variation, same scene, the characteristic number difference extracted under different light is very little, therefore there is good adaptability, can judge whether camera is blocked accurately at various illumination condition.The prerequisite extracting picture local edge unique point due to the present invention is mean square deviation ω > 0, so just must eliminate the abnormal conditions such as network failure or equipment failure, substantially increase reliability, reduce rate of false alarm.

Accompanying drawing explanation

Fig. 1 is FB(flow block) of the present invention.

Embodiment

When camera is not blocked, the picture clear-cut of shooting; And camera is when being blocked, shelter from camera lens very close to, the picture taken is caused to become very fuzzy, there is no edge contour clearly, therefore before and after blocking, the quantity of the local edge unique point that picture is detected differs greatly, and when the quantity of picture local edge unique point is lower than the threshold value of setting, then shows that camera is blocked.

The RGB image I of camera shooting is of a size of 1280x720, threshold value n ₀be set to 100.

As shown in Figure 1, a kind of camera occlusion detection method based on image local edge feature, comprises the following steps:

A, to camera shooting RGB image I carry out greyscale transformation, obtain gray level image L, and calculate the mean square deviation ω of each grey scale pixel value in gray level image L.

B, judge whether the value of ω equals 0, if ω=0, then show network failure or equipment failure, terminate program; If ω > 0, then carry out the extraction of image local Edge Feature Points, comprise the following steps:

B1, adopt Gaussian filter to carry out filtering process to gray level image L, construct the multiscale space with N width image, described multiscale space is made up of O group image, often organizes image and has S sublayer, N=O × S; The scale parameter σ of every tomographic image _imark respectively by sequence number o and s, scale parameter σ _icalculate according to formula (1):

o ∈ [0 ..., O-1], s ∈ [0 ..., S-1], i ∈ [0 ..., N], σ ₀be the initial baseline value of scale parameter, be defaulted as 1.6;

B2, employing Nonlinear diffusion filtering device carry out DIFFUSION TREATMENT to the every tomographic image in multiscale space, generate non-linear multiscale space, wherein the image L of the bottom ⁰=L _σ, L _σwith being of a size of 9 × 9, standard deviation be 1.6 gaussian kernel and described gray level image L carry out convolution after obtain; Remainder layer image generates evolution graph as L according to following formula (2) ⁱ:

$L^{i+1}={(I-\mathrm{\τ}\underset{l=1}{\overset{m}{\Σ}}{A}_l(L^i\left)\right)}^{-1}{L}^i,$

Wherein A _lrepresent image L ⁱconductance matrix on each dimension l, I is original image, and τ is time step, τ=t _i+1-t _i, t _ifor the scale parameter in units of the time, i ∈ [0 ..., N];

B3, according to the every secondary evolution graph generated in b2 step as L ⁱcalculate the Hessian matrix response image L of its correspondence ⁱ _hessian, according to formula (3):

L ⁱ _hessian=σ ² _{i, norm}(L _xx ⁱl _yy ⁱ-L _xy ⁱl _xy ⁱ), wherein, L ⁱ _xxfor evolution graph is as L ⁱsecond order in x direction is reciprocal, L ⁱ _yyfor evolution graph is as L ⁱsecond order in y direction is reciprocal, L ⁱ _xyfor evolution graph is as L ⁱsecond-order mixed partial derivative, σ _{i, norm}for L ⁱthe round values of corresponding yardstick;

The non-maximal value Restrainable algorithms of b4, employing 3 × 3 × 3 neighborhoods, locates each Hessian response image L ⁱ _hessianin local edge unique point.

The quantity n of the local edge unique point that c, statistics are extracted, by n and threshold value n ₀compare, threshold value n ₀for the local feature of image captured when camera does not block is counted; If n≤n ₀, then send camera and to be blocked alarm.

Further, step b4 specifically comprises the following steps:

B41, traversal response image L ⁱ _hessiain each response, if be less than default extreme value respthresh=0.001, then continue judge next response;

B42, current response diagram L ⁱ _hessiain response and 8 consecutive point of its same yardstick, and and neighbouring yardstick corresponding 9 × 2 points---totally 26 points compare, to guarantee all maximum point to be detected at metric space and two dimensional image space, this maximum point is local edge unique point.

Further, matrix A _lbe triple diagonal matrix and diagonal dominance.

Further, described formula (2) gets by carrying out discretize to nonlinear diffusion equations, and described nonlinear diffusion equations is as following formula (4):

$\frac{\∂L}{\∂t}=div\left(c\right(x,y,t)\×\▿L),$

Wherein, div and represent divergence and gradient respectively; T is the scale parameter in units of the time; Propagation function when c (x, y, t) is called that the evolution time is t, can make to spread the partial structurtes being adaptive to image, and propagation function c (x, y, t) adopts following formula (5) to calculate:

$c(x,y,t)=g\left(\right|\▿L_{\mathrm{\σ}}\left(x,y,t\right)\left|\right)=\frac{1}{1+\frac{\▿L_{\mathrm{\σ}x}{(x,y,t)}^2+\▿L_{\mathrm{\σ}y}{(x,y,t)}^2}{k^2}},$

Wherein, image L respectively _σhorizontal and vertical gradient; Parameter k controls other contrast factor of expansion stage, and the value of parameter k is gradient image the value of histogram on 70% hundredths.

Further, the computation process of described parameter k comprises the following steps:

B21, the Scharr operator using 3x3 and L _σcarry out convolution, obtain L respectively _σgradient image L in the horizontal direction _{σ x}with the gradient image L of vertical direction _{σ y};

B22, calculating L _σin the gradient magnitude at each pixel place, according to formula (6):

$M(x,y)=\sqrt{L_{\mathrm{\σ}x}{(x,y)}^2+L_{\mathrm{\σ}y}{(x,y)}^2},$ Wherein maximum gradient magnitude is M _max,

B23, gradient magnitude histogram is divided into nbins=300 dimension (bin), according to following formula (7), the gradient magnitude of each pixel is assigned to corresponding dimension nbin:

Add up the number that each gradient magnitude is not the pixel of 0, be designated as nps, then the value of histogram on 70% hundredths is: nthresh=nps*0.7;

B24, from position 0, each bin in traversal histogram, and to add up the value in each bin, is kept in variable nelements.As nelements >=nthresh, record the position kperc of now bin, the final expression formula of k is as following formula (8):

$k=\left\{\begin{array}{c}0.03,nelements<nthresh\\ M_{max}\&times;\frac{kperc}{nbins},\end{array}\right..$

The present invention in same scene, when illumination variation is larger, can stable detection to image local Edge Feature Points, it is high that the camera caused various situation blocks discrimination, can reach about 99%; The present invention, without the need to being algorithm configuration parameter according to different scenes, only using default parameters, therefore can better be applied to the cloud supervisory system of unified plan; The present invention accurately can judge that camera is blocked or the pictorial information caused due to network failure is lost.

Claims (6)

1., based on a camera occlusion detection method for image local edge feature, it is characterized in that: comprise the following steps:

A, to camera shooting RGB image I carry out greyscale transformation, obtain gray level image L, and calculate the mean square deviation ω of each grey scale pixel value in gray level image L;

B, judge whether the value of ω equals 0, if ω=0, then show network failure or equipment failure, terminate program; If ω ≠ 0, then carry out the extraction of image local Edge Feature Points, comprise the following steps:

B1, adopt Gaussian filter to carry out filtering process to gray level image L, construct the multiscale space with N width image, described multiscale space is made up of O group image, often organizes image and has S sublayer, N=O × S; The scale parameter σ of every tomographic image _imark respectively by sequence number o and s, scale parameter σ _icalculate according to formula (1):

o ∈ [0 ..., O-1], s ∈ [0 ..., S-1], i ∈ [0 ..., N], σ ₀be the initial baseline value of scale parameter, be defaulted as 1.6;

B2, employing Nonlinear diffusion filtering device carry out DIFFUSION TREATMENT to the every tomographic image in multiscale space, generate non-linear multiscale space, wherein the image L of the bottom ⁰=L _σ, L _σwith being of a size of 9 × 9, standard deviation be 1.6 gaussian kernel and described gray level image L carry out convolution after obtain; Remainder layer image generates evolution graph as L according to following formula (2) ⁱ:

$L^{i+1}={(I-\mathrm{\&tau;}\underset{l=1}{\overset{m}{\&Sigma;}}{A}_l\left(L^i\right))}^{-1}{L}^i,$

Wherein A _lrepresent image L ⁱconductance matrix on each dimension l, I is original image, and τ is time step, τ=t _i+1-t _i, t _ifor the scale parameter in units of the time,

$t_i=\frac{1}{2}{{\mathrm{\&sigma;}}_i}^2,i\&Element;\&lsqb;0,...,N\&rsqb;;$

B3, according to the every secondary evolution graph generated in b2 step as L ⁱcalculate the Hessian matrix response image L of its correspondence ⁱ _hessian, according to formula (3):

L ⁱ _hessian=σ ² _{i, norm}(L _xx ⁱl _yy ⁱ-L _xy ⁱl _xy ⁱ), wherein, L ⁱ _xxfor evolution graph is as L ⁱsecond order in x direction is reciprocal, L ⁱ _yyfor evolution graph is as L ⁱsecond order in y direction is reciprocal, L ⁱ _xyfor evolution graph is as L ⁱsecond-order mixed partial derivative, σ _{i, norm}for L ⁱthe round values of corresponding yardstick;

The non-maximal value Restrainable algorithms of b4, employing 3 × 3 × 3 neighborhoods, locates each Hessian response image L ⁱ _hessianin local edge unique point;

The quantity n of the local edge unique point that c, statistics are extracted, by n and threshold value n ₀compare, threshold value n ₀for the local feature of image captured when camera does not block is counted; If n≤n ₀, then send camera and to be blocked alarm.

2. the camera occlusion detection method based on image local edge feature according to claim 1, is characterized in that: step b4 specifically comprises the following steps:

B41, traversal response image L ⁱ _hessiain each response, if be less than default extreme value respthresh=0.001, then continue judge next response;

B42, current response diagram L ⁱ _hessiain response and 8 consecutive point of its same yardstick, and and neighbouring yardstick corresponding 9 × 2 points---totally 26 points compare, to guarantee all maximum point to be detected at metric space and two dimensional image space, this maximum point is local edge unique point.

3. the camera occlusion detection method based on image local edge feature according to claim 1, is characterized in that: RGB image I is of a size of 1280x720, described threshold value n ₀be set to 100.

4. the camera occlusion detection method based on image local edge feature according to claim 1, is characterized in that: matrix A _lbe triple diagonal matrix and diagonal dominance.

5. want the camera occlusion detection method based on image local edge feature described in 1 according to right, it is characterized in that: described formula (2) gets by carrying out discretize to nonlinear diffusion equations, and described nonlinear diffusion equations is as following formula (4):

$\frac{\&part;L}{\&part;t}=div\left(c\right(x,y,t)\&times;\&dtri;L),$

Wherein, div and represent divergence and gradient respectively; T is the scale parameter in units of the time; Propagation function when c (x, y, t) is called that the evolution time is t, can make to spread the partial structurtes being adaptive to image, and propagation function c (x, y, t) adopts following formula (5) to calculate:

$c(x,y,t)=g\left(\right|\&dtri;L_{\mathrm{\&sigma;}}\left(x,y,t\right)\left|\right)=\frac{1}{1+\frac{\&dtri;L_{\mathrm{\&alpha;}x}{(x,y,t)}^2+\&dtri;L_{\mathrm{\&alpha;}y}{(x,y,t)}^2}{k^2}},$

Wherein, image L respectively _σhorizontal and vertical gradient; Parameter k controls other contrast factor of expansion stage, and the value of parameter k is gradient image the value of histogram on 70% hundredths.

6. want the camera occlusion detection method based on image local edge feature described in 5 according to right, it is characterized in that: the computation process of described parameter k comprises the following steps:

B21, the Scharr operator using 3x3 and L _σcarry out convolution, obtain L respectively _σgradient image L in the horizontal direction _{σ x}with the gradient image L of vertical direction _{σ y};

B22, calculating L _σin the gradient magnitude at each pixel place, according to formula (6):

$M(x,y)=\sqrt{L_{\mathrm{\&alpha;}x}{(x,y)}^2+L_{\mathrm{\&alpha;}y}{(x,y)}^2},$ Wherein maximum gradient magnitude is M _max,

B23, gradient magnitude histogram is divided into nbins=300 dimension (bin), according to following formula (7), the gradient magnitude of each pixel is assigned to corresponding dimension nbin:

Add up the number that each gradient magnitude is not the pixel of 0, be designated as nps, then the value of histogram on 70% hundredths is: nthresh=nps*0.7;

B24, from position 0, each bin in traversal histogram, and to add up the value in each bin, is kept in variable nelements.As nelements>=nthresh, record the position kperc of now bin, the final expression formula of k is as following formula (8):

$k=\left\{\begin{array}{c}0.03,nelements<nthresh\\ M_{max}\&times;\frac{kperc}{nbins},\end{array}\right..$