CN104463253A - Fire fighting access safety detection method based on self-adaptation background study - Google Patents

Abstract

The invention discloses a fire fighting access safety detection method based on the self-adaptation background study. First N frames of images of a fire fighting access monitoring video stream serve as training samples to obtain a Gaussian mixture model after training. A background image is obtained according to the Gaussian mixture model. Background shearing is carried out on a detected image through the background image to extract the foreground area of the detected image. Whether the foreground area is the true foreground or other fake foreground caused by interference is judged through the HOG feature vector of the images, in the foreground area, of the foreground image and the detected image, and therefore whether abnormal conditions exist in the detected image is detected. If M continuous frames have abnormal conditions, a fire fighting access has a safety problem. The fire fighting access safety detection method based on the self-adaptation background study realizes accurate detection of safety of the fire fighting access through the built Gaussian mixture model of the background and based on foreground target detection of the HOG feature and the self-adaptation background strategy.

Description

Based on the passageway for fire apparatus safety detection method of adaptive background study

Technical field

The invention belongs to technical field of computer vision, more specifically say, relate to a kind of passageway for fire apparatus safety detection method based on adaptive background study.

Background technology

Passageway for fire apparatus refers to when various dangerous situation occurs, and implements the passage succoured and trapped personnel is evacuated for fire fighter.When there being the emergency conditioies such as fire to occur, fire fighter, fire-fighting vehicle and fire fighting equipment will enter from passageway for fire apparatus.Passageway for fire apparatus safety detection mainly comprises two aspects: (1) passageway for fire apparatus occlusion detection, (2) normally closed fire-proof door abnormal start-up detect.If passageway for fire apparatus is taken by motor vehicle or other obturators and fire fighter, fire-fighting vehicle and fire fighting equipment just may be caused not to reach the spot in time and incur loss through delay the rescue of fire, cause huge economic loss and casualties.Fire-proof door be a kind of there is resilience function close door, its effect is meet fire stability, integrality and thermal insulation within a certain period of time, except the effect with common door, has fire prevention, every cigarette, the specific function stopping high temperature.When being only in closed condition, after breaking out of fire, effectively could stop the invasion and attack of dense smoke raging fire, for evacuating personnel and fire-fighting rescue gain time.Therefore, passageway for fire apparatus occlusion detection and normally closed fire-proof door abnormal start-up detect and seem particularly important.

Traditional passageway for fire apparatus safety detection mainly relies on hand inspection, whether passageway for fire apparatus gets clogged, normally closed fire-proof door whether abnormal start-up to specify special staff to check, this kind of method is simple, do not need to rely on any equipment, cost is low, but the shortcoming of the method one is the potential safety hazard that can not exist in Timeliness coverage passageway for fire apparatus; Two is rely on staff significantly, and subjectivity is strong.Also have a kind of method to rely on collection video to be then connected to Control Room to be responsible for checking by special staff, although this method is more convenient than first method, reduce the burden of staff, the carrying out that can concentrate is monitored, but still very large to the degree of dependence of personnel, and do not reach the object in real time, automatically detected.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, a kind of passageway for fire apparatus safety detection method based on adaptive background study is proposed, by setting up the mixed Gauss model of background, foreground target based on HOG feature detects and adaptive RTS threshold adjustment strategy, realizes the whether safe accurate detection in passageway for fire apparatus.

For achieving the above object, the present invention is based on the passageway for fire apparatus safety detection method of adaptive background study, comprise the following steps:

S1: using N two field picture before the monitoring video flow of passageway for fire apparatus as training sample, training obtains the background mixed Gauss model mixed by three Gauss models, and the concrete steps of training comprise:

S1.1: mixed Gauss model carries out initialization, wherein Gauss model 1 carries out initialization according to the 1st two field picture, and expression formula is:

${\mathrm{\μ}}_{j,1}^1=x_{j,1}$

${\mathrm{\σ}}_{j,1}^1={\mathrm{\σ}}_{\mathrm{init}}$

${\mathrm{\ω}}_{j,1}^1={\mathrm{\ω}}_{\mathrm{init}}$

Wherein, represent the average of pixel j after the 1st two field picture training in Gauss model 1, the span of j is j=1,2 ..., M, M are pixel quantity in image, x _{j, 1}represent the pixel value of pixel j in the 1st two field picture, represent the variance of pixel j after the 1st two field picture training in Gauss model 1, σ _initfor the initial variance preset, represent the weight of pixel j after the 1st two field picture training in Gauss model 1, ω _initfor the initial weight preset;

The initialized expression formula of Gauss model 2,3 is:

${\mathrm{\μ}}_j^2={\mathrm{\μ}}_j^3=-1$

${\mathrm{\σ}}_{j,1}^2={\mathrm{\σ}}_{j,1}^3={\mathrm{\σ}}_{\mathrm{init}}$

${\mathrm{\ω}}_{j,1}^2={\mathrm{\ω}}_{j,1}^3=0$

represent the average of pixel j after the 1st two field picture training in Gauss model 2,3 respectively, represent the variance of pixel j after the 1st two field picture training in Gauss model 2,3, represent the weight of pixel j after the 1st two field picture training in Gauss model 2,3;

S1.2: adopt the 2nd, 3 successively ... N two field picture is trained mixed Gauss model, obtains mixed Gauss model, and the method for t+1 two field picture training is:

By three Gauss models of pixel j obtained through front t two field picture, according to value descending sort, i=1,2,3, adopt the pixel value x of pixel j in t+1 two field picture _{j, t+1}mate each Gauss model successively, matching process is: if wherein δ be greater than 0 constant, then think coupling, otherwise do not mate;

Once certain Gauss model the match is successful, then this Gauss model adopts pixel value x _{j, t+1}upgrade, other Gauss models remain unchanged, and more new formula is:

${\mathrm{\ω}}_{j,t+1}^i=(1-\mathrm{\α}){\mathrm{\ω}}_{j,t}^i+\mathrm{\α}$

${\mathrm{\μ}}_{j,t+1}^i=(1-\mathrm{\α}){\mathrm{\μ}}_{j,t}^i+\mathrm{\α}*x_{j,t+1}$

${\left({\mathrm{\σ}}_{j,t+1}^i\right)}^2=(1-\mathrm{\α}){\left({\mathrm{\σ}}_{j,t}^i\right)}^2+\mathrm{\α}*{(x_{j,t+1}-{\mathrm{\μ}}_{j,t}^i)}^2$

Wherein, α is default model learning rate, and span is 0 < α < 1, be illustrated respectively in the weight of i-th Gauss model in the mixed Gauss model of t frame and the rear pixel j of t+1 two field picture training, be illustrated respectively in the average of i-th Gauss model in the mixed Gauss model of t frame and the rear pixel j of t+1 two field picture training, be illustrated respectively in the variance of i-th Gauss model in the mixed Gauss model of t frame and the rear pixel j of t+1 two field picture training;

If pixel value x _{j, t+1}all do not mate with three Gauss models, then by last Gauss model according to pixel value x _{j, t+1}carry out replacement to upgrade, more new formula is:

${\mathrm{\&mu;}}_{j,t+1}^i=x_{j,t+1}$

${\mathrm{\&sigma;}}_{j,t+1}^i={\mathrm{\&sigma;}}_{\mathrm{init}}$

${\mathrm{\&omega;}}_{j,t+1}^i={\mathrm{\&omega;}}_{\mathrm{init}}$

After renewal completes, by the weight of three Gauss models be normalized, order remember that the average of the mixed Gauss model of having trained is variance is weight is

S2: to the detected image in fire-fighting channel monitoring video flowing, adopt background subtraction method to obtain the prospect bianry image of this two field picture according to mixed Gauss model, concrete grammar is:

S2.1: three of pixel j Gauss models are pressed carry out descending sort, get front B _jthe distribution as a setting of individual Gauss model, B _jcomputing formula be:

$B_j=\arg \underset{b}{\min }(\underset{i^{\&prime;}=1}{\overset{b}{\mathrm{\&Sigma;}}}{w}_j^{i^{\&prime;}}>T)$

T represents default threshold value;

S2.2: by the pixel value x ' of pixel j in detected image _jto B _jindividual Gauss model mates, if the match is successful any one Gauss model, then judge that pixel j is as background pixel, is set to 0 by the value of bianry image corresponding pixel points, otherwise judge that pixel j is as foreground pixel, is set to 1 by the value of bianry image corresponding pixel points;

S3: utilize HOG feature to judge the authenticity of foreground blocks in prospect bianry image, concrete grammar is:

S3.1: carry out connected domain analysis to the prospect bianry image that step S2 obtains, is classified as a class by the region being linked to be block and extracts the positional information on four borders, obtains the boundary rectangle coordinate of every block connected region;

S3.2: by the pixel value of the pixel average of Gauss model maximum for weight in three of each pixel Gauss models pixel as a setting, form a width background image;

S3.3: for each connected region, from the gray level image of detected image and background image, corresponding gray level image block is extracted respectively according to its boundary rectangle coordinate, extract the HOG proper vector of two gray level image blocks respectively, calculate the Euclidean distance D of two proper vectors, if D is greater than the threshold value T pre-set _d, then judge that this connected region is as real foreground blocks, otherwise be pseudo-foreground blocks;

S4: do not have connected domain or each connected region to be pseudo-foreground blocks if analyze the result obtained in step S3, then do not have abnormal conditions in this frame detected image, if having at least a connected region to be real foreground blocks, then judge that this frame has abnormal conditions; If continuous N frame all exists abnormal conditions, M is default frame number, then passageway for fire apparatus exists safety problem, otherwise passageway for fire apparatus safety.

The present invention is based on the passageway for fire apparatus safety detection method of adaptive background study, first the front N two field picture of passageway for fire apparatus monitoring video flow is utilized to obtain mixed Gauss model as training sample training, then background image is obtained according to mixed Gauss model, background image is adopted to carry out to detected image the foreground area that the background method of wiping out extracts detected image, by foreground image and detected image, the HOG proper vector of the image in foreground area judges that this foreground area is the pseudo-prospect that real prospect or other interference cause, thus judge whether there are abnormal conditions in detected image, if all there are abnormal conditions in continuous N frame, then there is safety problem in passageway for fire apparatus.The present invention has following beneficial effect:

(1) the present invention sets up the background mixed Gauss model of monitoring scene adaptively by training sample, and the background image of acquisition more tallies with the actual situation, and the testing result based on this background image is more accurate;

(2) adopt HOG proper vector to judge the authenticity of foreground target, effectively can get rid of the interference that illumination variation etc. causes;

(3) real-time online updating can also be carried out by detected image to the mixed Gauss model of background, and the strategy only upgrading background is adopted when upgrading, so both necessary renewal was carried out to scene, and turn avoid the overlong time that foreground target occurs and melt mistake for background gradually.

Accompanying drawing explanation

Fig. 1 is the embodiment process flow diagram of the passageway for fire apparatus safety detection method that the present invention is based on adaptive background study;

Fig. 2 is the process flow diagram of the mixed Gauss model training of background in Fig. 1;

Fig. 3 is the process flow diagram that in Fig. 1, prospect bianry image generates;

Fig. 4 is the process flow diagram that in Fig. 1, foreground blocks authenticity judges;

Fig. 5 is the normal detected image of scene 1;

Fig. 6 is the background image obtained according to background mixed Gauss model in the moment shown in Fig. 5;

Fig. 7 is the prospect bianry image that scene 1 obtains;

Fig. 8 is the detected image of scene 2 illuminance abrupt variation;

Fig. 9 is the background image that the background mixture model of moment shown in Fig. 8 obtains;

Figure 10 is the prospect bianry image that scene 2 obtains;

Figure 11 is the gray level image of detected image and the background image extracted after scene 2 connected domain analysis, and wherein Figure 11 (a) is the gray level image of detected image, and Figure 11 (b) is the gray level image of background image;

Figure 12 is the histogram of the HOG proper vector of scene 2;

Figure 13 is the detected image of scene 3 channel block;

Figure 14 is the background image that the background mixture model of moment shown in Figure 13 obtains;

Figure 15 is the prospect bianry image that scene 2 obtains;

Figure 16 is the gray level image of detected image and the background image extracted after scene 3 connected domain analysis, and wherein Figure 16 (a) is the gray level image of detected image, and Figure 16 (b) is the gray level image of background image;

Figure 17 is the histogram of the HOG proper vector of scene 3;

Figure 18 is the scene schematic diagram of illuminance abrupt variation scene, and wherein Figure 18 (a) is detected image, and Figure 18 (b) is background image, and Figure 18 (c) is testing result figure;

Figure 19 is the scene schematic diagram that fire-proof door opens scene 1, and wherein Figure 19 (a) is detected image, and Figure 19 (b) is background image, and Figure 19 (c) is testing result figure;

Figure 20 is the scene schematic diagram that fire-proof door opens scene 2, and wherein Figure 20 (a) is detected image, and Figure 20 (b) is background image, and Figure 20 (c) is testing result figure;

Figure 21 is the scene schematic diagram that scene 1 is blocked in passageway for fire apparatus, and wherein Figure 21 (a) is detected image, and Figure 21 (b) is background image, and Figure 21 (c) is testing result figure;

Figure 22 is the scene schematic diagram that scene 2 is blocked in passageway for fire apparatus, and wherein Figure 22 (a) is detected image, and Figure 22 (b) is background image, and Figure 22 (c) is testing result figure;

Figure 23 is the scene schematic diagram of outdoor fire fighting passage scene 2, and wherein Figure 23 (a) is detected image, and Figure 23 (b) is background image, and Figure 23 (c) is testing result figure.

Embodiment

Below in conjunction with accompanying drawing, the specific embodiment of the present invention is described, so that those skilled in the art understands the present invention better.Requiring particular attention is that, in the following description, when perhaps the detailed description of known function and design can desalinate main contents of the present invention, these are described in and will be left in the basket here.

Embodiment

Fig. 1 is the embodiment process flow diagram of the passageway for fire apparatus safety detection method that the present invention is based on adaptive background study.As shown in Figure 1, the passageway for fire apparatus safety detection method that the present invention is based on adaptive background study comprises the following steps:

S101: the mixed Gauss model training of background:

Using N two field picture before the monitoring video flow of passageway for fire apparatus as training sample, training obtains the mixed Gauss model of each pixel of background mixed by three Gauss models.

Fig. 2 is the process flow diagram of mixed Gauss model training in Fig. 1.As shown in Figure 2, the concrete steps of mixed Gauss model training comprise:

S201: mixed Gauss model carries out initialization:

First initialization is carried out to mixed Gauss model.The background model of each pixel in the present invention in image is mixed by three Gauss models.Wherein Gauss model 1 carries out initialization according to t=1 two field picture, and expression formula is:

${\mathrm{\&mu;}}_{j,1}^1=x_{j,1}$

${\mathrm{\&sigma;}}_{j,1}^1={\mathrm{\&sigma;}}_{\mathrm{init}}$

${\mathrm{\&omega;}}_{j,1}^1={\mathrm{\&omega;}}_{\mathrm{init}}$

Wherein, represent the average of pixel j after the 1st two field picture training in Gauss model 1, the span of j is j=1,2 ..., M, M are pixel quantity in image, x _{j, 1}represent the pixel value of pixel j in the 1st two field picture, represent the variance of pixel j after the 1st two field picture training in Gauss model 1, σ _initfor the initial variance preset, represent the weight of pixel j after the 1st two field picture training in Gauss model 1, ω _initfor the initial weight preset.In the present embodiment, initial variance σ _initbe set to 30, initial weight ω _initbe 0.05.

The initialized expression formula of Gauss model 2,3 is:

${\mathrm{\&mu;}}_j^2={\mathrm{\&mu;}}_j^3=-1$

${\mathrm{\&sigma;}}_{j,1}^2={\mathrm{\&sigma;}}_{j,1}^3={\mathrm{\&sigma;}}_{\mathrm{init}}$

${\mathrm{\&omega;}}_{j,1}^2={\mathrm{\&omega;}}_{j,1}^3=0$

represent the average of pixel j after the 1st two field picture training in Gauss model 2,3 respectively, represent the variance of pixel j after the 1st two field picture training in Gauss model 2,3, represent the weight of pixel j after the 1st two field picture training in Gauss model 2,3.

S202: Gauss model sorts:

By three Gauss models of pixel j obtained through front t two field picture, according to value descending sort.

S203: make i=1.

S204: adopt t+1 two field picture to mate Gauss model i:

Adopt the pixel value x of pixel j in t+1 two field picture _{j, t+1}mate Gauss model i, matching process is: if then think coupling, otherwise do not mate.δ be greater than 0 constant, δ=2.5 are set usually.

S205: judge that whether coupling is successful, enter step S206, otherwise enter step S207.

S206: upgrade Gauss model i, enter step S210, more new formula is:

${\mathrm{\&omega;}}_{j,t+1}^i=(1-\mathrm{\&alpha;}){\mathrm{\&omega;}}_{j,t}^i+\mathrm{\&alpha;}$

${\mathrm{\&mu;}}_{j,t+1}^i=(1-\mathrm{\&alpha;}){\mathrm{\&mu;}}_{j,t}^i+\mathrm{\&alpha;}*x_{j,t+1}$

${\left({\mathrm{\&sigma;}}_{j,t+1}^i\right)}^2=(1-\mathrm{\&alpha;}){\left({\mathrm{\&sigma;}}_{j,t}^i\right)}^2+\mathrm{\&alpha;}*{(x_{j,t+1}-{\mathrm{\&mu;}}_{j,t}^i)}^T(x_{j,t+1}-{\mathrm{\&mu;}}_{j,t}^i)$

Wherein, α is default model learning rate, and span is 0 < α < 1, be illustrated respectively in the weight of i-th Gauss model in the mixed Gauss model of t frame and the rear pixel j of t+1 two field picture training, be illustrated respectively in the average of i-th Gauss model in the mixed Gauss model of t frame and the rear pixel j of t+1 two field picture training, be illustrated respectively in the variance of i-th Gauss model in the mixed Gauss model of t frame and the rear pixel j of t+1 two field picture training.

S207: judge whether i=3, if not, enter step S208, otherwise enter step S209.

S208: make i=i+1, returns step S204.

S209: replace and upgrade last Gauss model, enter step 210:

More new formula is:

${\mathrm{\&mu;}}_{j,t+1}^i=x_{j,t+1}$

${\mathrm{\&sigma;}}_{j,t+1}^i={\mathrm{\&sigma;}}_{\mathrm{init}}$

${\mathrm{\&omega;}}_{j,t+1}^i={\mathrm{\&omega;}}_{\mathrm{init}}$

S210: weight normalization:

After each renewal completes, by the weight of three Gauss models be normalized, order

S211: judge whether t=N-1, if not, enter step S212, otherwise enter step S213.

S212: make t=t+1, returns step S202.

S213: trained, obtains mixed Gauss model, remembers that the average of the mixed Gauss model of having trained is variance is weight is

In actual applications, if monitor video image is gray level image, then pixel value is gray-scale value, each pixel only needs a mixed Gauss model, if be coloured image, so has RGB tri-passages, there are three mixed Gauss models in so each pixel, respectively corresponding RGB tri-passages.

S102: the prospect bianry image of detected image generates:

After the mixed Gauss model establishing background, need to carry out foreground target detection to detection picture frame.To the detected image in fire-fighting channel monitoring video flowing, background subtraction method is adopted to obtain the prospect bianry image of this two field picture according to mixed Gauss model.

Fig. 3 is the process flow diagram that prospect bianry image generates.As shown in Figure 3, prospect bianry image generates and comprises the following steps:

S301: the B of comparative selection _jindividual Gauss model:

The mixed Gauss model of pixel j actually depict the probability distribution of pixel value in time domain, in order to determine in mixed Gauss model, which Gauss model is produced by background, need by three of pixel j Gauss models by carry out descending sort, get front B _jthe distribution as a setting of individual Gauss model, B _jcomputing formula be:

$B_j=\arg \underset{b}{\min }(\underset{i^{\&prime;}=1}{\overset{b}{\mathrm{\&Sigma;}}}{w}_j^{i^{\&prime;}}>T)$

T represents default threshold value, and its implication is the ratio of background pixel point, and value is 0.8 in the present embodiment.

S302: to B _jindividual Gauss model mates.

By the pixel value x ' of pixel j in detected image _jto B _jindividual Gauss model mates, and matching process is identical with step S204.

S303: the match is successful to have judged whether Gauss model, if there is any one, the match is successful, enters step S304, otherwise enter step S305.

S304: judge that pixel j is as background pixel, is set to 0 by the value of bianry image corresponding pixel points.

S305: judge that pixel j is as foreground pixel, is set to 1 by the value of bianry image corresponding pixel points.

In coloured image, because each pixel has the mixed Gauss model of three passages, the result of determination that may appear at foreground pixel in three passages is different.Generally, if having any one to be judged to be foreground pixel in three passages, then think that this pixel is foreground pixel, if three passages are all judged to be background pixel, just think that this pixel is background pixel.

S103: foreground blocks authenticity judges:

Due to the change of illumination in monitoring image, in the prospect bianry image adopting background subtraction to obtain, pseudo-foreground target may be there is.In order to get rid of the impact of illumination instantaneous mutation on testing result, invention introduces histograms of oriented gradients (Histogram of Oriented Gradient, HOG) feature.The edge of the main Description Image of HOG feature and gradient information.This feature is a kind of Feature Descriptor being used for carrying out object detection in computer vision and pattern-recognition, and it forms gradient orientation histogram by the gradient information in calculating and statistical picture region.The normalization of this feature in computation process ensure that the optical conversion unchangeability of this descriptor, and during illumination variation the edge of image and gradient information all substantially constant, so adopt HOG feature can well get rid of illumination change suddenly impact on testing result, therefore the present invention adopts the distance of the HOG proper vector of background and current frame image to judge the authenticity of foreground target.

Fig. 4 is the process flow diagram that foreground blocks authenticity judges.As shown in Figure 4, the concrete grammar of foreground blocks authenticity in prospect bianry image is to utilize HOG feature to judge:

S401: analyze the connected domain obtaining prospect bianry image:

Connected domain analysis is carried out to the prospect bianry image that step S102 obtains, the region being linked to be block is classified as a class and extracts the positional information on four borders, obtain the boundary rectangle coordinate of every block connected domain.

Owing to there is noise in image, so general when carrying out connected domain analysis, too small connected domain can directly be given up, and a reservation size is greater than the connected domain of threshold value, and these connected domains just can be carried out piecemeal and be extracted HOG proper vector.

S402: extract background image:

Because the mixed Gauss model of each pixel is made up of several Gauss models, we need to extract one can represent the value of background to form background.According to gauss hybrid models, by the pixel value of the pixel average of Gauss model maximum for weight in the respective Gauss model of each pixel pixel as a setting, form a width background image.

Be the situation of gray level image for monitor video image, so from three Gauss models, extract weight when extraction represents background pixel value maximum, if monitor video image is coloured image, so just need from three Gauss models of each passage, extract weight separately maximum, then form colored background image.

S403: obtain gray level image:

The gray level image of the background image that acquisition detected image and step S402 obtain.

Visible, if monitor video image is originally as gray level image, the background image that so step S402 obtains also is gray level image naturally.If the monitor video image of colour, first greyscale image transitions is carried out to detected image and background image with regard to needs.

S404: two the HOG proper vectors extracting connected domain:

For each connected domain, from the gray level image of detected image and background image, extract corresponding gray level image block respectively according to its boundary rectangle coordinate, extract the HOG proper vector K of two gray level image blocks respectively ₁=(k ₁₁, k ₁₂..., k _1M) and K ₂=(k ₂₁, k ₂₂..., k _2M), the element number in M representation feature vector.

In the present embodiment, the concrete grammar of the extraction of HOG proper vector is: first divide carrying out block (Block) in image block.Because the size of connected domain boundary rectangle respectively has identical, therefore the present invention adopts fixed block number, and the method for pixel number adaptive change in block carrys out partitioned image block.The concrete grammar that block divides is: the pixels across quantity of note boundary rectangle is X, and longitudinal pixel quantity is Y, arranges horizontal piecemeal parameter P and longitudinal piecemeal parameter Q, P and Q are natural numbers, then pixels across quantity in each piece longitudinal pixel quantity represent and round downwards, from initial point transverse shifting block, transverse shifting step-length arrive transverse edge to vertically move, vertically moving step-length is again circulation like this is until complete piecemeal to image block.The visible number of blocks finally obtained is (2M _x-1) × (2M _y-1) individual.Gradient distribution in adding up each piece, final formation HOG proper vector.

S405: the Euclidean distance calculating two HOG proper vectors:

For two HOG proper vectors of each connected domain that step S404 obtains, calculate the Euclidean distance D of two proper vectors.Computing formula is:

$D=\sqrt{\underset{m=1}{\overset{M}{\mathrm{\&Sigma;}}}{(k_{1m}-k_{2m})}^2}$

S406: judge whether Euclidean distance D is greater than the threshold value T pre-set _d, if so, enter step S407, otherwise enter step S408.

S407: judge that this connected domain is as real foreground blocks.

S408: judge that this connected domain is pseudo-foreground blocks;

S104: abnormal conditions judge:

If analyze the result obtained in step S103 do not have connected domain, or each connected domain is pseudo-foreground blocks, then do not have abnormal conditions in this frame detected image, if having at least a connected domain to be real foreground blocks, then judges that this frame has abnormal conditions.In actual applications, it is current or safety door is of short duration opens in short-term to be there is pedestrian in escape way, therefore when not determining after single-frame images unusual circumstance that passageway for fire apparatus exists safety problem, and need to wait for a period of time, to get rid of of short duration abnormal conditions, the present invention is by judging that M frame detected image realizes continuously, if namely continuous N frame all exists abnormal conditions, M is default frame number, then passageway for fire apparatus exists safety problem, need warning or report to the police, otherwise passageway for fire apparatus safety.

In the safety detection process of long passageway for fire apparatus, due to the change of the environment such as light, scene can be made also to change, in order to adapt to these changes better, detected image can be adopted after every frame has detected to carry out study to the mixed Gauss model of pixel and upgrade.The residence time is longer in the picture obturator or the normally closed fire-proof door that is opened can be caused to be updated in background model for renewal due to foreground target pixel thus inspection does not measure abnormal next, so the mixed Gauss model of foreground target pixel can not upgrade, namely the present invention only upgrades background pixel point.Its update method is identical with the update method of training sample in step S101 to mixed Gauss model, namely first three Gauss models are sorted, then the pixel value of background pixel point is mated, namely this Gauss model is upgraded once the match is successful, if three Gauss models all can not the match is successful, then adopted by last Gauss model the pixel value of this pixel to carry out replacement to upgrade, upgraded the rear weight to three Gauss models and be normalized.

In online updating process, in order to prevent renewal process from making the too small foreground detection that causes of the standard deviation of Gauss model too responsive, a minimum value can be set to standard deviation, when the standard deviation calculated is less than minimum value, then standard deviation is set to minimum value, otherwise does not operate, keep initial value.Arranging minimum value in the present embodiment is 5.

Obviously, when channel monitoring image is gray level image, each pixel only has a mixed Gauss model to upgrade, if coloured image, then needs to upgrade the mixed Gauss model of RGB tri-passages.

Visible, when online updating, for each pixel, according to its attribute automatic decision, whether this is updated, the pixel belonging to prospect keeps original model constant, and the pixel only belonging to background just upgrades, and that is to say that foreground target region does not upgrade and only has context update like this, so both necessary renewal was carried out to scene, and turn avoid the overlong time that foreground target occurs and melt mistake for background gradually.

In actual applications, there are some flase drop situations seldom measured in the safety detection of passageway for fire apparatus.Such as: illumination variation causes pseudo-prospect, but by foreground target, this part pseudo-prospect is confirmed that step is rejected and is not mistakenly considered and detects obturator or normally closed safety door is opened.When carrying out online updating to mixed Gauss model, due to employing is prospect (although being at this moment pseudo-foreground target) not update strategy, so the background model of this part pseudo-foreground target part can not learn to upgrade, and the present image after illumination variation also remains unchanged substantially, at this moment this part becomes abnormality because pseudo-foreground target region that illumination causes will be falsely detected always, and cannot automatically revert to normal state.Although this part flase drop situation seldom occurs, the warning that this situation can cause detection system to continue always and cannot recover normal, causes very bad impact.So must be solved this situation.An artificial target can be set up in detection system carry out the execution of control program and then allow it revert to normal state.When there is this flase drop situation, staff finds that warning belongs to flase drop, then manually will restart control traffic sign placement and become rebooting status, original background mixed Gauss model is deleted, again from current monitoring video flow, extract N open image as training sample, again set up the mixed Gauss model of background, carry out safety detection based on new mixed Gauss model, at this moment testing result will recover normal condition.

Propose technique effect of the present invention to verify, real passageway for fire apparatus monitor video sequence has carried out a series of experiment.Conveniently statistic mixed-state rate and false drop rate, normally closes as positive sample using normal choke free passageway for fire apparatus, normally closed fire-proof door; The passageway for fire apparatus of obstruction, normally closed fire-proof door are opened as negative sample.

Next 3 of the passageway for fire apparatus monitoring video taken from same video camera do not intercept 3 scenes in the same time to carry out experimental verification, and its Scene 1 is normal clog-free situation, and scene 2 is illuminance abrupt variation situations, and scene 3 is passageway for fire apparatus blocking situation.Fig. 5 is the normal detected image of scene 1.Fig. 6 is the background image obtained according to background mixed Gauss model in the moment shown in Fig. 5.Fig. 7 is the prospect bianry image that scene 1 obtains.As shown in Figures 5 to 7, because scene 1 is normal clog-free situation, so there is no prospect in the prospect bianry image Fig. 7 obtained, therefore carry out there is no prospect connected domain in the analysis result of connected domain, so result of determination is situation without exception.

Fig. 8 is the detected image of scene 2 illuminance abrupt variation.Fig. 9 is the background image that the background mixture model of moment shown in Fig. 8 obtains.Figure 10 is the prospect bianry image that scene 2 obtains.As shown in Figure 10, owing to there is the situation of illuminance abrupt variation, in detected image, there is hot spot, therefore carrying out there is foreground area in the prospect bianry image that background subtraction obtains.

Figure 11 is the gray level image of detected image and the background image extracted after scene 2 connected domain analysis, and wherein Figure 11 (a) is the gray level image of detected image, and Figure 11 (b) is the gray level image of background image.Figure 12 is the histogram of the HOG proper vector of scene 2.As shown in figure 12, in scene 2 on the scene, closely, its Euclidean distance is less than the threshold value preset to the HOG proper vector of detected image and background image, and therefore judge that this connected domain is as pseudo-prospect, therefore the result of determination of scene 2 is situations without exception.

Figure 13 is the detected image of scene 3 channel block.As shown in figure 13, scene 3 times, has pedestrian to pass through, defines obstruction to passage in passageway for fire apparatus.Figure 14 is the background image that the background mixture model of moment shown in Figure 13 obtains.Figure 15 is the prospect bianry image that scene 2 obtains.As shown in figure 13, because pedestrian passes through, carrying out there is foreground area in the prospect bianry image that background subtraction obtains.

Figure 16 is the gray level image of detected image and the background image extracted after scene 3 connected domain analysis, and wherein Figure 16 (a) is the gray level image of detected image, and Figure 16 (b) is the gray level image of background image.Figure 17 is the histogram of the HOG feature trace of scene 3.As shown in figure 17, in scene 3, in scene 2, the HOG proper vector comparison in difference of detected image and background image is large, and its Euclidean distance is greater than the threshold value preset, therefore judge that this connected domain is as real prospect, therefore the result of determination of scene 3 there are abnormal conditions.But because this pedestrian can pass through passageway for fire apparatus at short notice, can not cause long-time obstruction, abnormal conditions continue frame number and are less than default frame number threshold value, therefore can not produce warning or report to the police.

Next the testing result of 6 sections of different video scenes is added up, comprising situations such as illuminance abrupt variation, normally closed fire-proof door abnormal start-up, passageway for fire apparatus obstructions.

Figure 18 is the scene schematic diagram of illuminance abrupt variation scene, and wherein Figure 18 (a) is detected image, and Figure 18 (b) is background image, and Figure 18 (c) is testing result figure.

Figure 19 is the scene schematic diagram that fire-proof door opens scene 1, and wherein Figure 19 (a) is detected image, and Figure 19 (b) is background image, and Figure 19 (c) is testing result figure.

Figure 20 is the scene schematic diagram that fire-proof door opens scene 2, and wherein Figure 20 (a) is detected image, and Figure 20 (b) is background image, and Figure 20 (c) is testing result figure.

Figure 21 is the scene schematic diagram that scene 1 is blocked in passageway for fire apparatus, and wherein Figure 21 (a) is detected image, and Figure 21 (b) is background image, and Figure 21 (c) is testing result figure.

Figure 22 is the scene schematic diagram that scene 2 is blocked in passageway for fire apparatus, and wherein Figure 22 (a) is detected image, and Figure 22 (b) is background image, and Figure 22 (c) is testing result figure.

Figure 23 is the scene schematic diagram of outdoor fire fighting passage scene 2, and wherein Figure 23 (a) is detected image, and Figure 23 (b) is background image, and Figure 23 (c) is testing result figure.

Table 1 is the quantitative testing result statistical form of 6 scene monitoring videos.

Scene	Verification and measurement ratio	False drop rate
			Illuminance abrupt variation scene	>95％	<2％
Fire-proof door opens scene 1	>95％	<2％
			Fire-proof door opens scene 2	>95％	<2％
Scene 1 is blocked in passageway for fire apparatus	>95％	<2％
			Scene 2 is blocked in passageway for fire apparatus	>95％	<2％
Outdoor fire fighting passage scene	>85％	<10％

Table 1

As can be seen from Table 1, because the video scene of 5 is above in indoor, various interference is less, and relatively simply, Detection results is fine.Visible the present invention can overcome the impact of illumination variation effectively, and the detection that abnormal start-up and passageway for fire apparatus for fire-proof door are blocked is all very accurate.For the 6th outdoor fire fighting passage scene, obviously, the interference such as personnel at all levels and vehicle is more, and relative complex, Detection results has decline a little, but maintains good Detection results in light change.

Although be described the illustrative embodiment of the present invention above; so that those skilled in the art understand the present invention; but should be clear; the invention is not restricted to the scope of embodiment; to those skilled in the art; as long as various change to limit and in the spirit and scope of the present invention determined, these changes are apparent, and all innovation and creation utilizing the present invention to conceive are all at the row of protection in appended claim.

Claims (5)

1., based on a passageway for fire apparatus safety detection method for adaptive background study, it is characterized in that, comprise the following steps:

S1: using N two field picture before the monitoring video flow of passageway for fire apparatus as training sample, training obtains the background mixed Gauss model mixed by three Gauss models, and the concrete steps of training comprise:

S1.1: mixed Gauss model initialization, wherein Gauss model 1 carries out initialization according to the 1st two field picture, and expression formula is:

${\mathrm{\&mu;}}_{j,1}^1=x_{j,1}$

${\mathrm{\&sigma;}}_{j,1}^1={\mathrm{\&sigma;}}_{\mathrm{init}}$

${\mathrm{\&omega;}}_{j,1}^1={\mathrm{\&omega;}}_{\mathrm{init}}$

Wherein, represent the average of pixel j after the 1st two field picture training in Gauss model 1, the span of j is j=1,2 ..., M, M are pixel quantity in image, x _{j, 1}represent the pixel value of pixel j in the 1st two field picture, represent the variance of pixel j after the 1st two field picture training in Gauss model 1, σ _initfor the initial variance preset, represent the weight of pixel j after the 1st two field picture training in Gauss model 1, ω _initfor the initial weight preset;

The initialized expression formula of Gauss model 2,3 is:

${\mathrm{\&mu;}}_{j,1}^2={\mathrm{\&mu;}}_{j,1}^3=-1$

${\mathrm{\&sigma;}}_{j,1}^2={\mathrm{\&sigma;}}_{j,1}^3={\mathrm{\&sigma;}}_{\mathrm{init}}$

${\mathrm{\&omega;}}_{j,1}^2={\mathrm{\&omega;}}_{j,1}^3=0$

represent the average of pixel j after the 1st two field picture training in Gauss model 2,3 respectively, represent the variance of pixel j after the 1st two field picture training in Gauss model 2,3, represent the weight of pixel j after the 1st two field picture training in Gauss model 2,3;

S1.2: adopt t=2 successively, 3 ... N two field picture is trained mixed Gauss model, obtains mixed Gauss model, and the method for t+1 two field picture training is:

By three Gauss models of pixel j obtained through front t two field picture, according to value descending sort, i=1,2,3, adopt the pixel value x of pixel j in t+1 two field picture _{j, t+1}mate each Gauss model successively, matching process is: if then think coupling, otherwise do not mate;

Once certain Gauss model the match is successful, then this Gauss model adopts pixel value x _{j, t+1}upgrade, other Gauss models remain unchanged, and more new formula is:

${\mathrm{\&omega;}}_{j,t+1}^i=(1-\mathrm{\&alpha;}){\mathrm{\&omega;}}_{j,t}^i+\mathrm{\&alpha;}$

${\mathrm{\&mu;}}_{j,t+1}^i=(1-\mathrm{\&alpha;}){\mathrm{\&mu;}}_{j,t}^i+\mathrm{\&alpha;}*x_{j,t+1},$

${\left({\mathrm{\&sigma;}}_{j,t+1}^i\right)}^2=(1-\mathrm{\&alpha;}){\left({\mathrm{\&sigma;}}_{j,t}^i\right)}^2+\mathrm{\&alpha;}*{(x_{j,t+1}-{\mathrm{\&mu;}}_{j,t}^i)}^T(x_{j,t+1}-{\mathrm{\&mu;}}_{j,t}^i)$

Wherein, α is default model learning rate, and span is 0 < α < 1, be illustrated respectively in the weight of i-th Gauss model in the mixed Gauss model of t frame and the rear pixel j of t+1 two field picture training, be illustrated respectively in the average of i-th Gauss model in the mixed Gauss model of t frame and the rear pixel j of t+1 two field picture training, be illustrated respectively in the variance of i-th Gauss model in the mixed Gauss model of t frame and the rear pixel j of t+1 two field picture training;

If pixel value x _{j, t+1}all do not mate with three Gauss models, then by last Gauss model according to pixel value x _{j, t+1}carry out replacement to upgrade, more new formula is:

${\mathrm{\&mu;}}_{j,t+1}^i=x_{j,t+1}$

${\mathrm{\&sigma;}}_{j,t+1}^i={\mathrm{\&sigma;}}_{\mathrm{init}}$

${\mathrm{\&omega;}}_{j,t+1}^i={\mathrm{\&omega;}}_{\mathrm{init}}$

After renewal completes, by the weight of three Gauss models be normalized, order remember that the average of the mixed Gauss model of having trained is variance is weight is

S2: to the detected image in fire-fighting channel monitoring video flowing, adopt background subtraction method to obtain the prospect bianry image of this two field picture according to mixed Gauss model, concrete grammar is:

S2.1: three of pixel j Gauss models are pressed carry out descending sort, get front B _jthe distribution as a setting of individual Gauss model, B _jcomputing formula be:

$B_j=\arg \underset{b}{\min }(\underset{i^{\&prime;}=1}{\overset{b}{\mathrm{\&Sigma;}}}{w}_j^{i^{\&prime;}}>T)$

T represents default threshold value;

S2.2: by the pixel value x ' of pixel j in detected image _jto B _jindividual Gauss model mates, if the match is successful any one Gauss model, then judge that pixel j is as background pixel, is set to 0 by the value of bianry image corresponding pixel points, otherwise judge that pixel j is as foreground pixel, is set to 1 by the value of bianry image corresponding pixel points;

S3: utilize HOG feature to judge the authenticity of foreground blocks in prospect bianry image, concrete grammar is:

S3.1: carry out connected domain analysis to the prospect bianry image that step S2 obtains, is classified as a class by the region being linked to be block and extracts the positional information on four borders, obtains the boundary rectangle coordinate of every block connected region;

S3.2: by the pixel value of the pixel average of Gauss model maximum for weight in three of each pixel Gauss models pixel as a setting, form a width background image;

S3.3: for each connected region, from the gray level image of detected image and background image and background image, corresponding gray level image block is extracted respectively according to its boundary rectangle coordinate, extract the HOG proper vector of two gray level image blocks respectively, calculate the Euclidean distance D of two proper vectors, if D is greater than the threshold value T pre-set _d, then judge that this connected region is as real foreground blocks, otherwise be pseudo-foreground blocks;

S4: if each connected region is pseudo-foreground blocks in step S3, then do not have abnormal conditions in this frame detected image, if having at least a connected region to be real foreground blocks, then judges that this frame has abnormal conditions; If continuous N frame all exists abnormal conditions, M is default frame number, then passageway for fire apparatus exists safety problem, otherwise passageway for fire apparatus safety.

2. passageway for fire apparatus according to claim 1 safety detection method, is characterized in that, in described step S3.4, the extracting method of HOG proper vector is:

The pixels across quantity of note boundary rectangle is X, and longitudinal pixel quantity is Y, arranges horizontal piecemeal parameter P and longitudinal piecemeal parameter Q, P and Q are natural numbers, then pixels across quantity in each piece longitudinal pixel quantity represent and round downwards, from initial point transverse shifting block, transverse shifting step-length arrive transverse edge to vertically move, vertically moving step-length is again circulation like this is until complete piecemeal to image block; Gradient distribution in adding up each piece, final formation HOG proper vector.

3. passageway for fire apparatus according to claim 1 safety detection method, is characterized in that, also comprise step S5:

After detection completes, mixed Gauss model according to detected image carry out study upgrade, concrete grammar is: first to three Gauss models according to value descending sort, then the pixel value of background pixel point is mated successively, namely this Gauss model is upgraded once the match is successful, if three Gauss models all can not the match is successful, then adopted by last Gauss model the pixel value of this pixel to carry out replacement to upgrade, upgraded the rear weight to three Gauss models and be normalized.

4. passageway for fire apparatus according to claim 3 safety detection method, is characterized in that, in the study renewal process of described step S5, for standard deviation sets a minimum value, when the standard deviation calculated is less than minimum value, then standard deviation is set to minimum value.

5. passageway for fire apparatus according to claim 1 safety detection method, it is characterized in that, also comprise the process to flase drop, concrete grammar is: when there is flase drop, original background mixed Gauss model is deleted, return to step S1 from current monitoring video flow, to extract N open image as training sample, again set up the mixed Gauss model of background, carry out safety detection based on new mixed Gauss model.