CN102509414A - Smog detection method based on computer vision - Google Patents

Abstract

The invention discloses a smog detection method based on computer vision. The method comprises the following steps of: detecting motion regions of a scene; performing feature weighting summation on the motion regions to obtain initial attribute judgment of each motion region; and determining whether the motion regions in a video sequence belong to the same target with an interframe motion region correlation method, comprehensively analyzing the same target, and judging whether the target is smog. A relation among interframe motion regions is determined with the interframe motion region correlation method, and comprehensive judgment is performed on the attribute of a motion target. The method has the characteristics of low complexity and reduction in the false alarm rate, and smog occurring in a monitored scene can be found accurately in time.

Description

A kind of smog detection method based on computer vision

Technical field

The invention belongs to computer vision methods, be specifically related to smog detection method, can be applicable to the fire alarm monitoring based on computer vision.

Background technology

Traditional fire alarm system based on smoke detector is owing to obtaining application widely to characteristics such as the high sensitivity of smog and low costs aspect the fire prevention and control.But because its special principle, promptly detector must contact and could report to the police with certain density smog, makes it can't be applied to big space and open-air atmosphere.In addition, the extended discovery time of smog of the time that smog diffuses to alarm detector is unfavorable for the discovery early of fire.

Computer vision is mainly studied the method for the information of from view data, obtaining; In fire alarm system based on video monitoring; Can analyze video image content through computer vision methods; Acquisition does not produce chemical reaction to the preliminary understanding of guarded region scene and need not contact with smog, therefore can monitor large space and open-air area; Simultaneously, the scene image information data based on the fire alarm system of video monitoring can obtain to enrich can in time provide fire location, the preliminary judgement of intensity of a fire size, and the very first time provides condition of a fire information, reduces fire damage.

Smoke Detection belongs to the detection identification problem of specific objective in the computer vision field, and some researchists have proposed the detection algorithm based on the smog different characteristic.Smoke Detection algorithm in actual at present the use mainly contains following several kinds:

1) based on the Smoke Detection of colouring information

Colouring information is the important information of figure, through in color graphics, seeking the zone of particular color, can find potential target area, thereby realizes the detection of smog.Yet, utilize colouring information to carry out Smoke Detection and also exist some significantly not enough, for example receive the interference of similar color target; Can in addition, set up suitable color model to the smog of different colours, also be the critical limitation that restrict colors information is used in Smoke Detection.

2) based on the Smoke Detection of movable information

There is specific rule (cigarette is toward the eminence diffusion) in the motion of smog, through calculating the light stream in the scene, finds the light stream kinetic characteristic of target, can smog and the target that does not possess these kinetic characteristics be made a distinction.Yet, the accuracy of optical flow computation, the image-forming conditions of guarded region etc. all have very big influence to the accurate testing result of smog.

3) based on the Smoke Detection of wavelet analysis

Wavelet analysis method is as signal Processing, and especially the important tool in the Flame Image Process all has important application in a lot of problems of image processing field.Through scene image is carried out wavelet transformation, obtain the wavelet domain information of image, can analyze image simultaneously at frequency domain and spatial domain.Scholar's research is arranged in the image smog zone with the difference of non-smog zone in wavelet field; A series of smog detection methods have been studied based on wavelet transformation; Like wavelet field energy loss and the relation that keeps energy, the statistical law of wavelet coefficient etc., obtained effect preferably.But wavelet analysis method often only to the smog of specific modality, is difficult to satisfy the application demand of some specific occasions.

Though the researchist has proposed different Smoke Detection algorithms; But because the change in shape of smog is varied; Concentration, the gray difference of the smog that different comburants produce are very big, add that the background of detection has nothing in common with each other, and are difficult to find the characteristic that can be good at describing smog in the image at present.

Summary of the invention

The object of the present invention is to provide a kind of smog detection method based on computer vision; At first the characteristic through in every frame video, calculating the moving region is carried out initial analysis to the category attribute in zone; According to the relation in interframe movement zone the moving target attribute is comprehensively judged then; Can realize the real-time Smoke Detection in the indoor and outdoor large space scope, for the fire prevention and control in places such as bulk storage plant provide technical support.

A kind of smog detection method based on computer vision is specially:

Detect the moving region of t two field picture, the sequence number of moving region is designated as i;

Extract the more than one characteristic of i moving region of t frame;

Calculate the characteristic weighing and the attribute score that obtains this moving region of i moving region of t frame;

Calculate the distance of all moving regions of i moving region of t frame and t-1 two field picture;

Confirm the minimum value and value of all moving regions of i moving region of t frame and t-1 two field picture;

If the minimum value and value of all moving regions of i moving region of t frame and t-1 two field picture less than minimum threshold of distance, is then upgraded the attribute score of i moving region of t frame according to the attribute score of the corresponding moving region of this minimum value of t-1 two field picture;

If the minimum value and value of all moving regions of i moving region of t frame and t-1 two field picture is then calculated the distance of all moving regions of i moving region of t frame and t-2 two field picture more than or equal to distance threshold;

Confirm the minimum value and value of all moving regions of i moving region of t frame and t-2 two field picture;

If the minimum value and value of all moving regions of i moving region of t frame and t-2 two field picture less than minimum threshold of distance, is then upgraded the attribute score of i moving region of t frame according to the attribute score of the corresponding moving region of this minimum value of t-2 two field picture;

If the attribute score of i moving region of t frame surpasses alarm threshold value, then assert to have smog.

Further, if the minimum value and value of all moving regions of i moving region of t frame and t-2 two field picture is more than or equal to minimum threshold of distance, then the attribute score of i moving region of t frame remains unchanged.

Further; The attribute that the attribute score of the corresponding moving region of this minimum value of said foundation t-1 two field picture is upgraded i moving region of t frame gets step by step and is specially: the attribute score of i moving region of t frame

is the attribute score of this minimum value correspondence moving region of t-1 two field picture, 0.8≤a≤0.95.

Further; The attribute that the attribute score of the corresponding moving region of this minimum value of said foundation t-2 two field picture is upgraded i moving region of t frame gets step by step and is specially: the attribute score of i moving region of t frame

is the attribute score of this minimum value correspondence moving region of t-2 two field picture, 0.75≤b≤0.9.

Further, the characteristic of said moving region comprises: the maximum maximum ratio of gradient pixel and region area greatly in the average of moving region and variance, moving region of the maximum modified-image of gray scale in the ratio of image average, the preceding historical frames image that continues that reduces of image and gray scale that increases of gray scale in the historical frames image that continues before the average gray average in the historical frames that continues before in gray average, the moving region image is passed through in number of times, the moving region.

Further, being specially of all moving regions of said i moving region of calculating t frame and t-1 two field picture apart from step:

Movement in the i-th region and the frame t-1 j-th motion area respectively choose a

and

cubes;

Calculate the distance of j moving region of i moving region and t-1 two field picture $D_{i,j}^{t,t-1}=\frac{{\mathrm{\Σ}}^{L_i^t/2}{\mathrm{Rank}}_m^{\mathrm{Max}}\left({\mathrm{\Σ}}^{L_j^{t-1}/2}{\mathrm{Rank}}_n^{\mathrm{Max}}\left(d_{m,n}^{i,j}\right)\right)}{L_j^{t-1}\×L_i^t/4},$ Wherein,

$d_{m,n}^{i,j}=|{\mathrm{\μ}}_m^{t,i}-{\mathrm{\μ}}_n^{t-1,j}|/{\mathrm{\λ}}_{\mathrm{mean}}+|{\mathrm{\σ}}_m^{t,i}-{\mathrm{\σ}}_n^{t-1,j}|/{\mathrm{\λ}}_{\mathrm{variance}}+\sqrt{{({\mathrm{\θx}}_m^{t,i}-{\mathrm{\θx}}_n^{t-1,j})}^2+{({\mathrm{\θy}}_m^{t,i}-{\mathrm{\θy}}_n^{t-1,j})}^2}/{\mathrm{\λ}}_{\mathrm{location}}$

M representes m square of i moving region of t frame, and n representes n square of j moving region of t-1 two field picture,, λ _Mean, λ _Variance, λ _LocationBe respectively weight parameter,

With

Be respectively the average and the variance of l square in i moving region of t frame,

Be the position of l square in i moving region of t frame,

With

Be respectively in the t-1 frame average and the variance of l square in j the moving region j,

Be the position of l square in j moving region in the t-1 frame,

Expression is that independent variable is asked maximum with n

Individual With,

Expression is that independent variable is asked maximum with m

Individual φ (m) with.

Further, the moving region step of said detection t two field picture comprises:

Generate the sport foreground image step:

T two field picture F _tIn each pixel F _t(x, y), respectively with t-Δ t ₁, t-Δ t ₂, t-Δ t ₃The pixel of correspondence in the two field picture $F_{t-{\mathrm{\Δ\; t}}_1}(x,y),F_{t-{\mathrm{\Δ\; t}}_2}(x,y),F_{t-{\mathrm{\Δ\; t}}_3}(x,y)$ Subtract each other and take absolute value, obtain difference ${\mathrm{Diff}}_{t-{\mathrm{\Δ\; t}}_3}(x,y),{\mathrm{Diff}}_{t-{\mathrm{\Δ\; t}}_2}(x,y),{\mathrm{Diff}}_{t-{\mathrm{\Δ\; t}}_3}(x,y),$ The maximal value and the motion detection threshold Δ F that get difference make comparisons, if think then that greater than threshold value Δ F there is motion in this some place, then with this pixel

Be changed to 255, otherwise put 0, thereby obtain the sport foreground image;

Sport foreground image filtering step;

The connected component labeling step.

Further, Δ F value is between 10 to 30.

Technique effect of the present invention is embodied in: the present invention adopts interframe target association method to confirm the relation in interframe movement zone, and the moving target attribute is comprehensively judged.This method has the advantages that complexity is low, reduce false alarm rate, the smog that occurs in can discovery monitoring scene promptly and accurately.

Description of drawings

Fig. 1 is the inventive method overview flow chart;

Fig. 2 is the video scene sectional drawing that there is smog in two width of cloth;

Fig. 3 carries out the synoptic diagram as a result that the moving region is detected for Fig. 2 scene;

Fig. 4 is a Smoke Detection synoptic diagram as a result.

Embodiment

Describe the present invention below in conjunction with instantiation.

If whether have the smog zone in the scene that needs detection video sequence F to monitor, referring to Fig. 1, the present invention moves as follows:

(1) moving object detection

Preceding ζ two field picture in the preservation video sequence is to image sequence Image_list, and the ζ from video sequence+1 two field picture begins, and can begin to detect present image F _tIn moving target, moving object detection may further comprise the steps:

(1.1) generate the sport foreground image

Present image F _tIn each pixel F _t(x, y), respectively with t-Δ t ₁, t-Δ t ₂, t-Δ t ₃The pixel of correspondence in the two field picture $F_{t-{\mathrm{\Δ\; t}}_1}(x,y),F_{t-{\mathrm{\Δ\; t}}_2}(x,y),F_{t-{\mathrm{\Δ\; t}}_3}(x,y)$ Subtract each other and take absolute value, obtain difference ${\mathrm{Diff}}_{t-{\mathrm{\Δ\; t}}_3}(x,y),{\mathrm{Diff}}_{t-{\mathrm{\Δ\; t}}_2}(x,y),{\mathrm{Diff}}_{t-{\mathrm{\Δ\; t}}_3}(x,y),$ The maximal value and the motion detection threshold Δ F that get difference make comparisons, if think then that greater than threshold value there is motion in this some place, then with this pixel

Be changed to 255, otherwise put 0, thereby obtain the sport foreground image, that is:

${\hat{F}}_t(x,y)=\left\{\begin{array}{cc}255,& \begin{array}{cc}\mathrm{if}& \max ({\mathrm{diff}}_{{t-\mathrm{\Δt}}_1}(x,y),{\mathrm{diff}}_{{t-\mathrm{\Δt}}_2}(x,y),{\mathrm{diff}}_{{t-\mathrm{\Δt}}_3}(x,y))>\mathrm{\ΔF}\end{array}\\ 0,& \mathrm{else}\end{array}\right.$

Wherein the motion detection threshold of Δ F for setting artificially set based on the quality of video image, the smokescope that need detect etc., and general value is between 10 to 30.

(1.2) the sport foreground image is carried out filtering

In order to eliminate isolated noise point that exists in the above-mentioned foreground image that obtains

and the target area that is connected disconnection, select for use median filter that

carried out Filtering Processing in this instance.

Medium filtering is based on the theoretical a kind of nonlinear signal processing technology that can effectively suppress noise of sequencing statistical; The ultimate principle of medium filtering is to replace the color value of certain pixel in the image with the intermediate value after each pixel color value sorts in the neighborhood of this pixel; Let the color value of surrounding pixel more near actual value; Thereby eliminate isolated noise spot; The neighborhood of in the present embodiment

being selected for use when carrying out medium filtering is 8 neighborhoods of this pixel, and the intermediate value of gray-scale value of promptly choosing all pixels in 8 neighborhoods is as the filtered result of this pixel.So-called pixel (x, neighborhood y) are meant that this pixel has 4 levels and vertical neighbor, its coordinate be (x+1, y), (x-1, y); (x, y+1), (x, y-1), these four points are referred to as (x; Y) 4 neighborhoods, simultaneously (x, the neighbor at 4 diagonal angles y) has following coordinate: (x+1, x+1), (x+1; Y-1), (x-1, y+1), (x-1, y-1).8 points of all this are referred to as that (x, 8 neighborhoods y) are if (x y) is positioned at the border of image, and then some point in its 8 neighborhoods falls into the outside of image.

(1.3) connected component labeling:

Bianry image Through after the Filtering Processing, with pixel value wherein be 255 and the pixel that is arranged in the other side's 8 neighborhoods each other come out with same numeric indicia, all pixels that have identical numerical value in the image behind the mark then are under the jurisdiction of same connected domain, with all N that obtain _tIndividual connected domain is kept at object queue Objetc_list;

If N _t=0, then there is not moving target in the current scene, jump to step (4);

If N _t≠ 0, then there is moving target in the scene, continue execution in step (2);

(2) N that step (1) is obtained _tIndividual zone through the characteristic in the zoning, obtains the initial score in each zone

(2.1) gray average of image

in the zoning

is the brightness of image in the zone, can reflect the bright dark situation of image.Be defined as:

${\mathrm{Mean}}_i^t={\mathrm{\Σ}}_{(x,y)\∈I_i^t}{F}_t(x,y)/{\mathrm{Area}}_i^t;$

Wherein representes point (x; Y) belong in the scope of moving region i in the current t two field picture,

area of moving region i in the expression t two field picture;

(2.2) the average gray average in the past ζ two field picture is passed through number of times

and is passed through number of times for average mean in the zone in the zoning; The overall frequency information of motion in the reflecting regional is defined as:

${\mathrm{MCR}}_i^{t-\mathrm{\ζ},t}={\mathrm{\Σ}}_{(x,y)\∈I_i^t}{\mathrm{MCR}}^{t-\mathrm{\ζ},t}(x,y)/{\mathrm{Area}}_i^t;$

MCR wherein ^{T-ζ, t}(x, y) be point (x, y) average in the time range [t-ζ, t] is passed through number of times, promptly in the past in the ζ two field picture, adjacent two two field picture gray-scale values pass the gray average M of this some place in all ζ two field pictures ^{T-ζ, t}(x, number of times y).MCR ^{T-ζ, t}(x, computing method y) are: make MCR ^{T-ζ, t}(x, y)=0, ω=t-ζ ..., t-1:

If (F _ω(x, y)-M ^{T-ζ, t}(x, y)) * (M ^{T-ζ, t}(x, y)-F _ω+1(x, y))＜0; MCR then ^{T-ζ, t}(x, y)=MCR ^{T-ζ, t}(x, y)+1;

(2.3) calculate the maximum image TCincrease that increases of the interior gray scale of ζ two field picture in the past ^{T-ζ, t}, the gray scale maximum reduces image TCdecrease ^{T-ζ, t}With the maximum modified-image TCchange of gray scale ^{T-ζ, t}Thereby, calculate each regional statistical information ${\mathrm{TCquotient}}_i^{t-\mathrm{\ζ},t},{\mathrm{TCmean}}_i^{t-\mathrm{\ζ},t},{\mathrm{TC\; var\; iance}}_i^{t-\mathrm{\ζ},t};$

Maximum (minimizing/variation) image that increases of gray scale is meant that in the ζ two field picture, the adjacent two frame gray scales of each pixel increase the image that (minimizing/variation) maximum value is formed in the past;

TCincrease ^t-ζ，t(x，y)＝max _{q∈[t-ζ，t-1]}(F _q+1(x，y)-F _q(x，y))；

TCdecrease ^t-ζ，t(x，y)＝max _{q∈[t-ζ，t-1]}(F _q(x，y)-F _q+1(x，y))；

TCchange ^t-ζ，t(x，y)＝max _{q∈[t-ζ，t-1]}(|F _q(x，y)-F _q+1(x，y)|)；

If TCincrease ^{T-ζ, t}(x, y) (or TCdecrease ^{T-ζ, t}(x, y)) less than 0, and promptly (x y) locates gray-scale value continuous decrease (or increase) in the ζ two field picture in the past to point, is 0 with this disposal then.

refers to maximum image and the maximum ratio that reduces the image average of increasing in the moving region, reflected the situation of change of this regional luminance in the ζ two field picture time;

${\mathrm{TCquotient}}_i^{t-\mathrm{\ζ},t}={\mathrm{\Σ}}_{(x,y)\∈I_i^t}{\mathrm{TCincrease}}^{t-\mathrm{\ζ},t}(x,y)/{\mathrm{\Σ}}_{(x,y)\∈I_i^t}{\mathrm{TCdecrease}}^{t-\mathrm{\ζ},t}(x,y)$

Be maximum modified-image TCchange ^{T-ζ, t}Average in regional i and variance, reflected in the ζ two field picture time should the zone in the degree of irregularity of gray-value variation;

(2.4) ratio of big gradient pixel and region area in the zoning

In the regional i of

expression, gradient accounts for the ratio of whole region area greater than the pixel of threshold value Δ Grad:

${\mathrm{GRADquotient}}_i^t={\#}_{(x,y)\∈I_i^t}(F_t^{\mathrm{grad}}(x,y)>\mathrm{\ΔGrad})/{\mathrm{Area}}_i^t$

Wherein,

Be image F _tPoint (x, gradient y),

Be illustrated in the regional i scope, satisfy condition The number of pixel;

Wherein Δ Grad is a Grads threshold, can according to how much the preestablishing of the marginal information in the scene, and some feature-set self-adapting threshold that also can domain of dependence.In this example; Use this regional mean flow rate

as Grads threshold; Promptly brighter place can allow to exist significantly edge, and darker place tangible marginal information should not occur.

Compute gradient is used the sobel operator in this example, and the sobel operator is practice one of the most frequently used operator when counting the word gradient of falling into a trap.Through using template:

Respectively to image F _tCarry out convolution and obtain convolution results

And order

Try to achieve gradient image.

(2.5), calculate each regional initial attribute score

according to the provincial characteristics that step (2.1)-(2.4) are calculated

Be the number of a reflection target area i attribute, by

Obtain etc. feature calculation,

Be worth greatly more, think that then motion target area i possibly be smog more; But here Be not the probability that proper regional i is a smog,, therefore, claim because the nonnegativity that it does not satisfy probability function does not pass through normalization yet

Score for regional i; In this example

For

${\mathrm{TCquotient}}_i^{t-\mathrm{\ζ},t},{\mathrm{TCmean}}_i^{t-\mathrm{\ζ},t},{\mathrm{TC\; var\; iance}}_i^{t-\mathrm{\ζ},t},{\mathrm{GRADquotient}}_i^t$ Weighted sum promptly:

$p_i^t=W^T\×\mathrm{Feature}$

The T representing matrix transposition in the W upper right corner wherein, Feature are the vector that this provincial characteristics is formed: ${({1,{\mathrm{Meant}}_i^t,{\mathrm{MCR}}_i^{t-\mathrm{\ζ},t},\mathrm{TCquotient}}_i^{t-\mathrm{\ζ},t},{\mathrm{TCmean}}_i^{t-\mathrm{\ζ},t},{\mathrm{TC\; var\; iance}}_i^{t-\mathrm{\ζ},t},{\mathrm{GRADquotient}}_i^t)}^{T}W$ Weight vector (α for each characteristic ₀, α ₁, α ₂, α ₃, α ₅, α ₅, α ₆) ^TW can be obtained by sample training through the method for machine learning;

(3) interframe movement zone association; Confirm the relation of moving region on the time series, obtain each regional final score

(3.1) to all N in the present frame _tIndividual zone, picked at random in each zone

(in this example

0＜β＜1) individual 5 take advantage of 5 sizes square, calculate all

The statistical information of individual square and positional information, wherein the average of l square, variance and position are designated as respectively With

The statistical information and the positional information of all blockages that obtain sampling among the i of target area deposit in the data structure that target i is corresponding among the Objetc_list;

(3.2) to all N in the present frame _tIndividual target area is according among the i of target area

Individual square calculates the distance (being feature difference) of all moving regions in target area and the t-1 two field picture, and the frame pitch that forms all target areas in t two field picture and the t-1 two field picture is from matrix D MAT ^{T, t-1}(N _tRow N _T-1Row), DMAT ^{T, t-1}In the capable j column element of i be the distance between the j of target area in target area i and the t-1 two field picture in the t two field picture

The frame pitch of moving region is from matrix D MAT between two two field pictures ^{T, t-1}Form is following:

${\mathrm{DMAT}}^{t,t-1}=\left[\begin{array}{ccccc}{D}_{\mathrm{1,1}}^{t,t-1}& D_{\mathrm{1,2}}^{t,t-1}& ...& D_{1,N_{t-1}-1}^{t,t-1}& D_{1,N_{t-1}}^{t,t-1}\\ D_{\mathrm{2,1}}^{t,t-1}& D_{\mathrm{2,2}}^{t,t-1}& ...& D_{2,N_{t-1}-1}^{t,t-1}& D_{2,N_{t-1}}^{t,t-1}\\ .& .& & .& .\\ .& .& D_{i,j}^{t,t-1}& .& .\\ .& .& & .& .\\ D_{N_t-\mathrm{1,1}}^{t,t-1}& D_{N_t-\mathrm{1,2}}^{t,t-1}& ...& D_{N_t-1,N_{t-1}-1}^{t,t-1}& D_{N_t-1,N_{t-1}}^{t,t-1}\\ D_{N_t,1}^{t,t-1}& D_{N_t,2}^{t,t-1}& ...& D_{N_t,N_{t-1}-1}^{t,t-1}& D_{N_t,N_{t-1}}^{t,t-1}\end{array}\right]$

DMAT ^{T, t-1}In the capable j column element of i be the distance between the j of target area in target area i and the t-1 two field picture in the t two field picture Can obtain through the distance between the blockage that calculates stochastic sampling in two target areas:

At first defining the distance of any two blockages, is example with the distance

of the square n that samples among sampling square m among the i of t two field picture zone and the t-1 two field picture zone j:

$d_{m,n}^{i,j}=|{\mathrm{\μ}}_m^{t,i}-{\mathrm{\μ}}_n^{t-1,j}|/{\mathrm{\λ}}_{\mathrm{mean}}+|{\mathrm{\σ}}_m^{t,i}-{\mathrm{\σ}}_n^{t-1,j}|/{\mathrm{\λ}}_{\mathrm{variance}}+\sqrt{{({\mathrm{\θx}}_m^{t,i}-{\mathrm{\θx}}_n^{t-1,j})}^2+{({\mathrm{\θy}}_m^{t,i}-{\mathrm{\θy}}_n^{t-1,j})}^2}/{\mathrm{\λ}}_{\mathrm{location}}$

λ wherein _Mean, λ _Variance, λ _LocationBe the parameter that is provided with, in order to the weight of adjustment area, variance and distance; For the smog of motion, grey scale change is less, and the variance in each several part smog zone maybe be bigger, and motion slowly, the λ that chooses in this example of These characteristics based on smog _Mean, λ _Variance, λ _LocationBe respectively 40,600,4, in practical application, also can make suitable adjustment to three's weight according to concrete needs. the individual blockage that obtains for stochastic sampling among regional i and the j; Calculate their distances between any two; Can obtain moving region i, the interregional distance matrix of j

${\mathrm{dm\; at}}_{i,j}^{t,t-1}=\left[\begin{array}{ccccc}{d}_{\mathrm{1,1}}^{i,j}& d_{\mathrm{1,2}}^{i,j}& ...& d_{1,L_j^{t-1}-1}^{i,j}& d_{1,L_j^{t-1}}^{i,j}\\ d_{\mathrm{2,1}}^{i,j}& d_{\mathrm{2,2}}^{i,j}& ...& d_{2,L_j^{t-1}-1}^{i,j}& d_{2,L_j^{t-1}}^{i,j}\\ .& .& & .& .\\ .& .& d_{m,n}^{i,j}& .& .\\ .& .& & .& .\\ d_{L_i^t-\mathrm{1,1}}^{i,j}& d_{L_i^t-\mathrm{1,2}}^{i,j}& ...& d_{L_i^t-1,L_j^{t-1}-1}^{i,j}& d_{L_i^t-1,L_j^{t-1}}^{i,j}\\ d_{L_i^t,1}^{i,j}& d_{L_i^t,2}^{i,j}& ...& d_{L_i^t,L_j^{t-1}-1}^{i,j}& d_{L_i^t,L_j^{t-1}}^{i,j}\end{array}\right]$

Through interregional distance matrix, can obtain the distance of two moving regions:

$D_{i,j}^{t,t-1}=\frac{{\mathrm{\Σ}}^{L_i^t/2}{\mathrm{Rank}}_m^{\max }\left({\mathrm{\Σ}}^{L_j^{t-1}/2}{\mathrm{Rank}}_n^{\max }\left(d_{m,n}^{i,j}\right)\right)}{L_j^{t-1}\×L_i^t/4}$

Wherein

symbolic representation with n be independent variable ask maximum

individual

with; Promptly ask in the interregional distance matrix

maximum

individual element in each row with, obtain vectorial φ of one

row;

expression with m be independent variable ask maximum

individual φ (m) with; Promptly ask among the vectorial φ maximum

individual element with, renormalization obtains the distance

between two moving regions

(3.3) pass through DMAT ^{T, t-1}Ask in all object block of t-1 two field picture and the nearest distance of current goal piece i

And establish

Be in the t-1 two field picture with present frame t in the nearest pairing index in moving region of moving region i;

If

Think two target areas couplings, promptly corresponding same moving target then uses update coefficients a to upgrade score, promptly ${\tilde{p}}_i^t=p_i^t+a\×{\tilde{p}}_{J_i^{\mathrm{Min}}}^{t-1};$

Wherein ε is a minimum threshold of distance, with λ _Mean, λ _Variance, λ _LocationChoose relevantly, use in this example under 40,600,4 the situation, the ε value is 8; A is a update coefficients, and the size that score influenced before the expression target received is generally got between 0.8～0.95; After upgrading score, jump to step (3.6);

Otherwise thinking does not have the coupling of current moving region in the t-1 two field picture, need to the t-2 two field picture, seek the coupling target, continues execution in step (3.4).

Computing method be: through calculating

The frame pitch that obtains t two field picture and t-1 two field picture is from matrix D MAT ^{T, t-1}, to DMAT ^{T, t-1}In each the row i, ask all N in this row _T-1The minimum value of individual element can obtain the distance of nearest moving region in regional i and the t-1 two field picture

The columns at minimum value place for the index of the nearest moving region of correspondence

(3.4) if moving region i can't be related with moving region in the t-1 two field picture, according among the i of moving region The distance of all moving regions in individual square information calculations moving region and the t-2 two field picture is promptly calculated t two field picture and t-2 two field picture frame pitch from matrix D MAT ^{T, t-2}The value of all elements during i is capable (for the t-1 two field picture in the zone of zone association, need not to calculate again the distance of target area in itself and the t-2 two field picture), DMAT ^{T, t-2}In the capable k column element of i be the distance between the k of target area in target area i and the t-1 two field picture in the t two field picture

(3.5) Find t-2 frame all the target block and the current target block closest distance

and set

is t-2 frame image in the current frame t i recent movement moving region regional targets.

If

then thinks two target areas couplings; It is corresponding same moving target; Then score is upgraded with update coefficients b; Promptly

wherein b＜a be update coefficients; The size that score influenced before the expression target received is generally got between 0.75～0.9;

Otherwise; Thinking does not have the coupling of current moving region in the t-2 two field picture, and current moving region is emerging moving target

(3.6) whether judge

greater than alarm threshold value η; If

thinks that then the target area is a smog, report to the police;

Wherein, η chooses with update coefficients a, b and the sensitivity of reporting to the police is required relevant, and a, b equal respectively under 0.9,0.8 the situation in this example, and the η value is 3.5 to have obtained the detection effect than balance;

(4) accomplish the correspondence memory operation

(4.1) in the image sequence Image_list that preserves, discharge the t-ζ frame image information of being preserved, and preserve current t frame image information.

(4.2) information of moving region in the t-2 two field picture of being preserved among the release target area chained list Objetc_list;

(4.3) make t=t+1, continue execution in step (1);

Fig. 2 is the video scene sectional drawing that there is smog in two width of cloth; Fig. 3 carries out the synoptic diagram as a result that the moving region is detected for Fig. 2 scene; Fig. 4 is Smoke Detection figure as a result; Wherein black lines is the movement locus of associated region, and white portion is that final score

is negative zone; Light gray areas is to be identified as smog in the single-frame images; But holistic approach does not also reach the zone (comprising detected false-alarm and incipient smog in the single-frame images) of smog standard, the i.e. zone of on the time series; Dark grey partly is an alarm region.

Claims (8)

1. smog detection method based on computer vision is specially:

Detect the moving region of t two field picture, the sequence number of moving region is designated as i;

Extract the more than one characteristic of i moving region of t frame;

Calculate the characteristic weighing and the attribute score that obtains this moving region of i moving region of t frame;

Calculate the distance of all moving regions of i moving region of t frame and t-1 two field picture;

Confirm the minimum value and value of all moving regions of i moving region of t frame and t-1 two field picture;

If the minimum value and value of all moving regions of i moving region of t frame and t-1 two field picture less than minimum threshold of distance, is then upgraded the attribute score of i moving region of t frame according to the attribute score of the corresponding moving region of this minimum value of t-1 two field picture;

If the minimum value and value of all moving regions of i moving region of t frame and t-1 two field picture is then calculated the distance of all moving regions of i moving region of t frame and t-2 two field picture more than or equal to distance threshold;

Confirm the minimum value and value of all moving regions of i moving region of t frame and t-2 two field picture;

If the minimum value and value of all moving regions of i moving region of t frame and t-2 two field picture less than minimum threshold of distance, is then upgraded the attribute score of i moving region of t frame according to the attribute score of the corresponding moving region of this minimum value of t-2 two field picture;

If the attribute score of i moving region of t frame surpasses alarm threshold value, then assert to have smog.

2. smog detection method according to claim 1; It is characterized in that; If the minimum value and value of all moving regions of i moving region of t frame and t-2 two field picture is more than or equal to minimum threshold of distance, then the attribute score of i moving region of t frame remains unchanged.

3. smog detection method according to claim 1; It is characterized in that; The attribute that the attribute score of the corresponding moving region of this minimum value of said foundation t-1 two field picture is upgraded i moving region of t frame gets step by step and is specially: the attribute score of i moving region of t frame

is the attribute score of this minimum value correspondence moving region of t-1 two field picture, 0.8≤a≤0.95.

4. smog detection method according to claim 1; It is characterized in that; The attribute that the attribute score of the corresponding moving region of this minimum value of said foundation t-2 two field picture is upgraded i moving region of t frame gets step by step and is specially: the attribute score of i moving region of t frame

is the attribute score of this minimum value correspondence moving region of t-2 two field picture, 0.75≤b≤0.9.

5. smog detection method according to claim 1; It is characterized in that the characteristic of said moving region comprises: the maximum maximum ratio of gradient pixel and region area greatly in the average of moving region and variance, moving region of the maximum modified-image of gray scale in the ratio of image average, the preceding historical frames image that continues that reduces of image and gray scale that increases of gray scale in the historical frames image that continues before the average gray average in the historical frames that continues before in gray average, the moving region image is passed through in number of times, the moving region.

6. smog detection method according to claim 1 is characterized in that, being specially apart from step of all moving regions of said i moving region of calculating t frame and t-1 two field picture:

Movement in the i-th region and the frame t-1 j-th motion area respectively choose a

and

cubes;

Calculate the distance of j moving region of i moving region and t-1 two field picture

$D_{i,j}^{t,t-1}=\frac{{\mathrm{\Σ}}^{L_i^t/2}{\mathrm{Rank}}_m^{\mathrm{Max}}\left({\mathrm{\Σ}}^{L_j^{t-1}/2}{\mathrm{Rank}}_n^{\mathrm{Max}}\left(d_{m,n}^{i,j}\right)\right)}{L_j^{t-1}\×L_i^t/4},$ Wherein,

$d_{m,n}^{i,j}=|{\mathrm{\μ}}_m^{t,i}-{\mathrm{\μ}}_n^{t-1,j}|/{\mathrm{\λ}}_{\mathrm{mean}}+|{\mathrm{\σ}}_m^{t,i}-{\mathrm{\σ}}_n^{t-1,j}|/{\mathrm{\λ}}_{\mathrm{variance}}+\sqrt{{({\mathrm{\θx}}_m^{t,i}-{\mathrm{\θx}}_n^{t-1,j})}^2+{({\mathrm{\θy}}_m^{t,i}-{\mathrm{\θy}}_n^{t-1,j})}^2}/{\mathrm{\λ}}_{\mathrm{location}}$

M representes m square in i moving region of t frame, and n representes n square in j moving region of t-1 two field picture,, λ _Mean, λ _Variance, λ _LocationBe respectively weight parameter,

With

Be respectively the average and the variance of l square in i moving region of t frame,

Be the position of l square in i moving region of t frame,

With

Be respectively in the t-1 frame average and the variance of l square in j the moving region,

Be the position of l square in j moving region in the t-1 frame,

Expression is that independent variable is asked maximum with n

Individual With,

Expression is that independent variable is asked maximum with m

Individual φ (m) with.

7. smog detection method according to claim 1 is characterized in that, the moving region step of said detection t two field picture comprises:

Generate the sport foreground image step:

T two field picture F _tIn each pixel F _t(x, y), respectively with t-Δ t ₁, t-Δ t ₂, t-Δ t ₃The pixel of correspondence in the two field picture $F_{t-{\mathrm{\Δ\; t}}_1}(x,y),F_{t-{\mathrm{\Δ\; t}}_2}(x,y),F_{t-{\mathrm{\Δ\; t}}_3}(x,y)$ Subtract each other and take absolute value, obtain difference ${\mathrm{Diff}}_{t-{\mathrm{\Δ\; t}}_3}(x,y),{\mathrm{Diff}}_{t-{\mathrm{\Δ\; t}}_2}(x,y),{\mathrm{Diff}}_{t-{\mathrm{\Δ\; t}}_3}(x,y),$ The maximal value and the motion detection threshold Δ F that get difference make comparisons, if think then that greater than threshold value Δ F there is motion in this some place, then with this pixel Be changed to 255, otherwise put 0, thereby obtain the sport foreground image;

Sport foreground image filtering step;

The connected component labeling step.

8. smog detection method according to claim 7 is characterized in that, Δ F value is between 10 to 30.

Images