CN105574895A - Congestion detection method during the dynamic driving process of vehicle - Google Patents

Abstract

The invention relates to a congestion detection method during the dynamic driving process of a vehicle. The detection method comprises the steps: 1) foreground detection for the vehicle: by means of video shooting, performing frame selection of one area in every frame of a grey-scale map and utilizing a Vibe algorithm to perform background modeling, extracting the motion foregrounds of the vehicle, performing area filtering for every motion foreground, filtering the foregrounds with too large area or too small area, and maintaining the foreground which belongs to the vehicle most possibly; 2) vehicle tracking: after finishing extraction of FHOG characteristics, respectively using a translation filter and a scale filter to perform translation tracking and scale tracking for a target; and 3) according to the tracking result, calculating the real-time velocity for every vehicle target, and updating the accumulated real-time velocity Va, wherein the update rate is set as 0.1; determining whether Va is less than a threshold; if so, representing that congestion occurs and alarming; and if not, no alarming. The congestion detection method during the dynamic driving process of a vehicle has the advantages of being accurate in the detection effect and being able to quickly solve the congestion situation.

Description

A kind of detection method of blocking up in vehicle dynamic travels

Technical field

The invention belongs to vehicle congestion detection method technical field, particularly relate to a kind of detection method of blocking up in vehicle dynamic travels.

Background technology

Along with the continuous increase of private car quantity, urban transportation day by day can't bear the heavy load, and often in whole traffic circulation peak period, has to, by duty at each crossing of traffic police, solve the traffic problems day by day blocked up at present, consume a large amount of manpowers like this.But along with the progress of science and technology, have developed a kind of camera head, judge jam situation by the dense degree of the vehicular movement on backstage manual observation video and speed, cause accuracy not high, but also can not the look-ahead arrival of blocking up.

Summary of the invention

The object of the invention is to the deficiency overcoming prior art existence, and provide a kind of detection method of blocking up in vehicle dynamic travels, object is that Detection results is accurate, can quick solution jam situation.

The object of the invention is to have come by following technical solution, the method comprises the steps:

1), vehicle foreground detects: pass through video capture, each frame gray-scale map in input video surveyed area, frame selects a certain region of each frame gray-scale map to adopt Vibe algorithm to carry out background modeling, and extract vehicle movement prospect, area filtration is carried out to each sport foreground, the prospect that filtering area is too large or too little, reservation is most possibly the prospect of vehicle;

2), vehicle tracking: adopt Felzenszwalb algorithm to each vehicle foreground target, extract centered by target's center, be of a size of the FHOG feature of 33 different yardsticks of the same yardstick FHOG characteristic sum of the square region of target two times of sizes, extraction step is as follows:

(1) image is converted to gray-scale map, with the gradient operator [-1,0,1] on x direction and gradient operator [-1,0, the 1] T on y direction, filtering is carried out to gray-scale map;

(2) to each unique point on image, compute gradient direction and amplitude, formula is as follows:

$M(x,y)=\sqrt{I_x^2+I_y^2}$

Calculate during compute gradient direction simultaneously and have symbol (0-360 °) or the gradient direction without symbol (0-180 °);

(3) carry out weight votes by gradient direction and amplitude to there being 18 of symbol bin and signless 9 bin, obtain a length be 18 vector sum length be the vector of 9;

(4) every four unit combination become a block, are normalized, to eliminate illumination effect in block to vector;

(5) Felzenszwalb be extracted a large amount of unit without symbol gradient, each unit is totally 4 × 9=36 dimensional feature, and carry out principal component analysis (PCA) (PCA), find to use front 11 proper vectors substantially can comprise all information, but in order to quick calculating, obtain a kind of approximate PCA dimensionality reduction effect by the visual result of major component; Specifically, 36 dimensional vectors are regarded as the matrix of 4 × 9, to every a line, each row summation obtains 13 dimensional features, substantially can reach the Detection results that HOG feature 36 is tieed up; Be applicable to using the accuracy of detection having symbol gradient target to improve those, then have the summation of symbol gradient direction to obtain 18 dimensional vectors to 18, and add wherein, finally obtain the 13+18=31 dimensional feature vector in figure below;

(6), after FHOG feature extraction completes, with translate filter and scaling filter, translation tracking and yardstick tracking are carried out to target respectively;

3), according to the result of following the tracks of calculate the real-time speed of each vehicle target, and upgrade accumulation real-time speed Va, turnover rate can be arranged to 0.1; Judge whether Va is less than certain threshold value, if so, represents vehicle congestion and reports to the police, otherwise do not report to the police.

As preferably, described vehicle foreground detects by judging whether sampled pixel point pixel value in current frame pixel point and background model is apart from being greater than threshold value R to realize, and following formula is followed in its judgement:

#{S _R(v(x))∩{v ₁，v ₂，…，v _N}}≥#min

Optimum configurations is N=20, #min=2, R=20.

As preferably, the translation filter step that described translation is followed the tracks of is as follows:

1), the expection Gauss of initialized target exports G (two dimension);

2), centered by the old position of target, the sample Z that is of a size of target 2 times of sizes is gathered;

3), in sample each pixel calculates 28 dimension fusion features (1 dimension original gradation feature+27 ties up fhog feature), is multiplied by after two-dimentional hann window as test input Z;

4), adopt max (y) is asked to obtain target reposition.

As preferably, the scale filter step that described yardstick is followed the tracks of is as follows:

1), the expection Gauss of initialization yardstick exports G (three-dimensional);

2), centered by target reposition, the sample Z under 33 kinds of different scales is extracted;

3), each sample resize become fixed measure, extract 317 dimension fhog features respectively, all fhog features of each sample are connected into a proper vector again and form 33 layers of pyramid feature, are multiplied by after one dimension hann window as test input Z;

4), adopt max (y) is asked to obtain current yardstick.

Beneficial effect of the present invention is: can react vehicle jam situation fast, precisely, intuitively, and report to the police in advance to possible blocking up, thus improves the traffic efficiency of vehicle and ensure the unobstructed of road, and the fast passing for vehicle provides strong support.

Accompanying drawing explanation

Fig. 1 is the testing process schematic diagram that blocks up of the present invention.

Fig. 2 is that FHOG characteristic vector pickup of the present invention catches schematic diagram.

Fig. 3 is that gradient direction of the present invention and amplitude carry out weight votes schematic diagram.

Fig. 4 is the translation filter step schematic diagram that translation of the present invention is followed the tracks of.

Fig. 5 is the scale filter step schematic diagram that yardstick of the present invention is followed the tracks of.

Fig. 6 is vibe principle schematic of the present invention.

Embodiment

Do detailed introduction below in conjunction with accompanying drawing to the present invention: as shown in Figure 1, method of the present invention comprises the steps:

1), vehicle foreground detects: pass through video capture, each frame gray-scale map in input video surveyed area, frame selects a certain region of each frame gray-scale map to adopt Vibe algorithm to carry out background modeling, and extract vehicle movement prospect, area filtration is carried out to each sport foreground, the prospect that filtering area is too large or too little, reservation is most possibly the prospect of vehicle;

2), vehicle tracking: adopt Felzenszwalb algorithm to each vehicle foreground target, extract centered by target's center, be of a size of the FHOG feature of 33 different yardsticks of the same yardstick FHOG characteristic sum of the square region of target two times of sizes, extraction step is as follows:

(1) image is converted to gray-scale map, with the gradient operator [-1,0,1] on x direction and gradient operator [-1,0, the 1] T on y direction, filtering is carried out to gray-scale map;

(2) to each unique point on image, compute gradient direction and amplitude, formula is as follows:

$M(x,y)=\sqrt{I_x^2+I_y^2}$

Calculate during compute gradient direction simultaneously and have symbol (0-360 °) or the gradient direction without symbol (0-180 °);

(3) carry out weight votes by gradient direction and amplitude to there being 18 of symbol bin and signless 9 bin, obtain a length be 18 vector sum length be 9 vector (as shown in Figure 3);

(4) every four unit combination become a block, are normalized, to eliminate illumination effect in block to vector;

(5) Felzenszwalb be extracted a large amount of unit without symbol gradient, each unit is totally 4 × 9=36 dimensional feature, and carry out principal component analysis (PCA) (PCA), find to use front 11 proper vectors substantially can comprise all information, but in order to quick calculating, obtain a kind of approximate PCA dimensionality reduction effect by the visual result of major component; Specifically, 36 dimensional vectors are regarded as the matrix of 4 × 9, to every a line, each row summation obtains 13 dimensional features, substantially can reach the Detection results that HOG feature 36 is tieed up; Be applicable to using the accuracy of detection having symbol gradient target to improve those, then have the summation of symbol gradient direction to obtain 18 dimensional vectors to 18, and add wherein, finally obtain the 13+18=31 dimensional feature vector (as shown in Figure 2) in figure below;

(6), after FHOG feature extraction completes, with translate filter and scaling filter, translation tracking and yardstick tracking are carried out to target respectively; The wherein translation filter step following (as shown in Figure 4) of translation tracking:

1, the expection Gauss of initialized target exports G (two dimension);

2, centered by the old position of target, the sample Z that is of a size of target 2 times of sizes is gathered;

3, in sample, each pixel calculates 28 dimension fusion features (1 dimension original gradation feature+27 ties up fhog feature), is multiplied by after two-dimentional hann window as test input Z;

4, adopt max (y) is asked to obtain target reposition.

The scale filter step following (as shown in Figure 5) that yardstick is followed the tracks of:

1, the expection Gauss of initialization yardstick exports G (three-dimensional);

2, centered by target reposition, the sample Z under 33 kinds of different scales is extracted;

3, each sample resize is become fixed measure, extract 317 dimension fhog features respectively, all fhog features of each sample are connected into a proper vector again and form 33 layers of pyramid feature, are multiplied by after one dimension hann window as test input Z;

4, adopt max (y) is asked to obtain current yardstick.

3), according to the result of following the tracks of calculate the real-time speed of each vehicle target, and upgrade accumulation real-time speed Va, turnover rate can be arranged to 0.1; Judge whether Va is less than certain threshold value, if so, represents vehicle congestion and reports to the police, otherwise do not report to the police.

Described vehicle foreground detects by judging whether the sampled pixel point pixel value distance in current frame pixel point and background model is greater than threshold value R to realize, and following formula is followed in its judgement:

#{S _R(v(x))∩{v ₁，v ₂，…，v _N}}≥#min

Optimum configurations is N=20, #min=2, R=20.

When upgrading background, adopt following strategy:

1. when v (x) is judged as background, with the probability updating background sample set of 1/16;

2. when v (x) is selected be used for upgrading background time, replace at random a pixel value in background sample set with its;

3. when v (x) is selected be used for upgrading background time, get the background sample set of a pixel in its neighborhood, replace at random a pixel value in this background sample set with its.

As shown in Figure 6, vibe principle is: suppose the pixel value of v (x) for coordinate x point place; M (x)=and v1, v2 ..., vN} is the background sample set at x place, and sample size is N; The spheric region that SR (v (x)) is take x as centre of sphere R is radius; Vibe is by judging whether the distance of v (x) to vi is less than threshold value R and determines whether it being background dot, if be less than, is then judged as background dot, if be greater than, is then judged as foreground point.

Wave filter of the present invention differentiates:

Suppose f ₁..., f _ta series of target image block, using them as training sample.G is exported with target filtering ₁..., g _tas their label.The correlation filter h of optimum when time t _tobtain by minimizing mean square deviation:

Function f _j, g _jand h _tsize be all M × N, asterisk represent circumference be correlated with.Second equation follows Parseval's theorem.In formula, capitalization represents the DFT conversion of respective function. horizontal line above represents complex conjugate, it is the multiplication of point mode.After minimizing formula 2.1, obtain following wave filter:

$H_t=\frac{{\Σ}_{j=1}^T{\stackrel{\‾}{G}}_j{F}_j}{{\Σ}_{j=1}^T{\stackrel{\‾}{F}}_j{F}_j}.---2.2$

Target filtering exports g _jthat target f is dropped on summit _jthe Gaussian function at center.In practical application, H _tmolecule A _twith denominator B _talong with each new observation f _tinput upgraded respectively by the method for weighted average and obtain.

To the image block z that the size fixed in a new frame is M × N, the computing formula of the degree of confidence y that wave filter exports is here the inverse DFT operator of representative, the position of fresh target is positioned at the maximal value place of y.

Here adopt HOG feature as translate filter and got up in it and common gradation of image feature simultaneous.Certainly, other feature can also be comprised.In DSST algorithm, we only adopt 1 extra dimension wave filter to estimate yardstick, namely estimate translation with 2 dimension wave filters, search for the target in metric space with the wave filter of 3 dimensions.

We consider that the d dimensional feature figure of a signal represents, suppose that f is a square object block extracted inside this characteristic pattern, in f l dimension (l ∈ 1 ..., d}) on feature f ^lrepresent, target finds an optimization correlative filter device h, by a wave filter h on every one dimension ^lcomposition.This obtains by minimizing following cost function:

$\mathrm{\ϵ}=\left|\right|\underset{l=1}{\overset{d}{\Σ}}{h}^l*f^l-g|{|}^2+\mathrm{\λ}\underset{l=1}{\overset{d}{\Σ}}|\left|h^l\right|{|}^2---2.3$

Here, g is that the target filtering of training sample f exports, and parameter lambda >=0 controls the impact of regularization term.Minimize 2.3 to obtain:

$H^l=\frac{\stackrel{\‾}{G}{F}^l}{{\mathrm{\Σ}}_{k=1}^d\stackrel{\‾}{F^k}{F}^k+\mathrm{\λ}}---2.4$

As can be seen from above formula, prevent the adding of regularization term coefficient denominator be 0 situation.

Notice that formula 2.3 only considers a sample, optimal filter can be reached by the cost function minimized on all training samples, and this is excessive for online algorithm cost.In order to obtain the approximate of a robust, we adopt the molecule A in following formula renewal 2.4 respectively here _twith denominator B _t:

$A_t^l=(1-\mathrm{\η})A_{t-1}^l+\mathrm{\η}{\stackrel{\‾}{G}}_t{F}_t^l$

$B_t=(1-\mathrm{\η})B_{t-1}+\mathrm{\η}\underset{k=1}{\overset{d}{\Σ}}\stackrel{\‾}{F_t^k}{F}_t^k---2.5$

Here, η is a Study rate parameter.Output degree of confidence y on a square region z adopts formula 2.6 to represent:

First we use HOG features training translation estimation filter.In order to train this wave filter, we extract the characteristic pattern of object block, then wave filter h _transformula 2.5 is adopted to train.We come the position of estimating target in a new frame by the characteristic pattern being extracted in future position, and then degree of confidence output y adopts 2.6 to obtain.How following several trifles descriptions adopt these wave filters to carry out estimating target yardstick.

First a kind of associating translation-yardstick tracking based on study 3d metric space correlation filter is proposed.Filter size is fixed as the height and width that M × N × S, M and N is wave filter, and S is the number of yardstick.In order to upgrade wave filter, in our square region first around target, extract a feature pyramid.Training sample f is set to a cubical area above feature pyramid, and cube size is M × N × S and centered by target state estimator position out and yardstick.We export as target filtering with 3d Gaussian function.Finally, metric space tracking filter adopts formula 2.5 to upgrade.

In order to locate the position of target in a new frame, we extract one from feature pyramid and are of a size of M × N × S cube z, cubical position and the size being centrally located at prediction, then calculate degree of confidence according to formula 2.6 and export y.New target location is located in the maximal value place of y.

Be understandable that, for a person skilled in the art, technical scheme of the present invention and inventive concept be equal to and replace or change the protection domain that all should belong to the claim appended by the present invention.

Claims (4)

1. the detection method of blocking up in vehicle dynamic travels, is characterized in that: the method comprises the steps:

1), vehicle foreground detects: pass through video capture, each frame gray-scale map in input video surveyed area, frame selects a certain region of each frame gray-scale map to adopt Vibe algorithm to carry out background modeling, and extract vehicle movement prospect, area filtration is carried out to each sport foreground, the prospect that filtering area is too large or too little, reservation is most possibly the prospect of vehicle;

2), vehicle tracking: adopt Felzenszwalb algorithm to each vehicle foreground target, extract centered by target's center, be of a size of the FHOG feature of 33 different yardsticks of the same yardstick FHOG characteristic sum of the square region of target two times of sizes, extraction step is as follows:

(1) image is converted to gray-scale map, with the gradient operator [-1,0,1] on x direction and gradient operator [-1,0, the 1] T on y direction, filtering is carried out to gray-scale map;

(2) to each unique point on image, compute gradient direction and amplitude, formula is as follows:

Calculate during compute gradient direction simultaneously and have symbol (0-360 °) or the gradient direction without symbol (0-180 °);

(3) carry out weight votes by gradient direction and amplitude to there being 18 of symbol bin and signless 9 bin, obtain a length be 18 vector sum length be the vector of 9;

(4) every four unit combination become a block, are normalized, to eliminate illumination effect in block to vector;

(5) Felzenszwalb be extracted a large amount of unit without symbol gradient, each unit is totally 4 × 9=36 dimensional feature, and carry out principal component analysis (PCA) (PCA), find to use front 11 proper vectors substantially can comprise all information, but in order to quick calculating, obtain a kind of approximate PCA dimensionality reduction effect by the visual result of major component; Specifically, 36 dimensional vectors are regarded as the matrix of 4 × 9, to every a line, each row summation obtains 13 dimensional features, substantially can reach the Detection results that HOG feature 36 is tieed up; Be applicable to using the accuracy of detection having symbol gradient target to improve those, then have the summation of symbol gradient direction to obtain 18 dimensional vectors to 18, and add wherein, finally obtain the 13+18=31 dimensional feature vector in figure below;

(6), after FHOG feature extraction completes, with translate filter and scaling filter, translation tracking and yardstick tracking are carried out to target respectively;

3), according to the result of following the tracks of calculate the real-time speed of each vehicle target, and upgrade accumulation real-time speed Va, turnover rate can be arranged to 0.1; Judge whether Va is less than certain threshold value, if so, represents vehicle congestion and reports to the police, otherwise do not report to the police.

2. the detection method of blocking up in vehicle dynamic travels according to claim 1, it is characterized in that: described vehicle foreground detects by judging whether the sampled pixel point pixel value distance in current frame pixel point and background model is greater than threshold value R to realize, and following formula is followed in its judgement:

#{S _R(v(x))∩{v ₁，v ₂，…，v _N}}≥#min

Optimum configurations is N=20, #min=2, R=20.

3. the detection method of blocking up in vehicle dynamic travels according to claim 1, is characterized in that: the translation filter step that described translation is followed the tracks of is as follows:

1), the expection Gauss of initialized target exports G (two dimension);

2), centered by the old position of target, the sample Z that is of a size of target 2 times of sizes is gathered;

3), in sample each pixel calculates 28 dimension fusion features (1 dimension original gradation feature+27 ties up fhog feature), is multiplied by after two-dimentional hann window as test input Z;

4), adopt max (y) is asked to obtain target reposition.

4. the detection method of blocking up in vehicle dynamic travels according to claim 1, is characterized in that: the scale filter step that described yardstick is followed the tracks of is as follows:

1), the expection Gauss of initialization yardstick exports G (three-dimensional);

2), centered by target reposition, the sample Z under 33 kinds of different scales is extracted;

3), each sample resize become fixed measure, extract 317 dimension fhog features respectively, all fhog features of each sample are connected into a proper vector again and form 33 layers of pyramid feature, are multiplied by after one dimension hann window as test input Z;

4), adopt max (y) is asked to obtain current yardstick.