CN103839232B - A kind of pedestrian's cast shadow suppressing method based on agglomerate model - Google Patents

Abstract

The present invention proposes a kind of pedestrian's cast shadow suppressing method based on agglomerate model, and first described method builds pedestrian's agglomerate model, then by calculating moment characteristics and the agglomerate vertical histogram of agglomerate, the method by geometry obtains preliminary shadow region；Then, to whole pedestrian and shadow region, the gray scale of this shadow region, direction, center being carried out shade modeling as parameter, the shadow region obtaining primary segmentation is pruned.Difference is detected towards, the shade of varying number pedestrian and suppresses by the inventive method, and video processing speed is about 7.5 frames/s, it is possible to separated with its shade by moving target efficiently and accurately.

Description

A kind of pedestrian's cast shadow suppressing method based on agglomerate model

Technical field

The invention belongs to image procossing, video monitoring and traffic safety technology field, refer specifically to a kind of based on Pedestrian's cast shadow suppressing method of agglomerate model.

Background technology

Video Supervision Technique is contemplated to prevent the erroneous judgement of Security Personnel, utilizes automatic analysis technology to carry out video prison The technology of control.Current research method assumes that and there is not shade in video sequence, and in well-lighted scene, The shade of movement will be divided into foreground object by wrong.This will cause target location to be estimated, goal behavior analysis Error with subsequent treatment such as target identifications and difficulty.

At present, the method for moving shadow detection and suppression is roughly divided into following three types: method based on color model, Select a suitable color space, utilize the chromatic characteristic in color space of shadows pixels value to carry out shade Suppression, such as HSV space, color character invariant C1C2C3 space and normalization rgb space, but this A little methods are the most affected by noise and to light intensity sensitive；Method based on physical model, by the side of physics Method modeling or the specific appearance feature of study shadows pixels, the most double light source dichromatic reflection model BIDR, but it The situation that panel tone is identical with background cannot be processed；Method based on texture model, by shade spectral signature Obtaining candidate region, the correlation further according to these zone-textures distinguishes prospect and shade, such as normalization mutually Relevant, Gabor filters, orthogonal transformation etc., owing to multiple neighborhoods of a pixel are calculated, therefore fortune Calculation amount is relatively big, poor real.

Therefore, the most efficiently and accurately moving target is separated with its shade, have become as research at present Hot issue.

Summary of the invention

The technical problem to be solved is to overcome the deficiencies in the prior art, proposes a kind of based on agglomerate Pedestrian's cast shadow suppressing method of model.First the inventive method builds pedestrian's agglomerate model, then by calculating agglomerate Moment characteristics and agglomerate vertical histogram, by geometry method obtain preliminary shadow region.Then, by this moon The gray scale in territory, shadow zone, direction, center carry out shade modeling as parameter to whole pedestrian and shadow region, Former shadow region is pruned so that difference can be entered by algorithm towards, the shade of varying number pedestrian herein Row detection and suppression；The video processing speed of the inventive method is about 7.5 frames/s.

In order to solve above-mentioned technical problem, the technical solution adopted in the present invention is: a kind of based on agglomerate model Pedestrian's cast shadow suppressing method, comprises the steps:

Step A, builds pedestrian's agglomerate, makes each pedestrian represent with corresponding agglomerate, and it specifically comprises the following steps that

Step A-1, obtains current frame image, utilizes agglomerate extracting method, the scene color to current frame image Cluster, obtain image agglomerate；

Step A-2, according to described image agglomerate, utilizes mixed Gaussian background modeling method, obtains present frame figure The prospect masterplate of picture, obtains foreground moving agglomerate；

Step A-3, utilizes fuzzy clustering method to merge foreground moving agglomerate, it is thus achieved that pedestrian's agglomerate；

Step B, for pedestrian's agglomerate, utilizes shadow detection method that with its shade, pedestrian is carried out primary segmentation, The shadow region of acquisition pedestrian's agglomerate:

Step B-1, calculates the central moment of pedestrian's agglomerate, so obtain the shade of pedestrian's agglomerate towards；

Step B-2, calculates pedestrian's agglomerate vertical histogram, obtains pedestrian and the cut-point of shade in pedestrian's agglomerate；

Step B-3, according to the shade of pedestrian's agglomerate towards the cut-point with pedestrian Yu shade, determines pedestrian respectively Agglomerate pedestrian is minimum with the color average threshold value of shadow region and variance threshold values, selection color average and color variance Region, as the shadow region of pedestrian's agglomerate；

Step C, according to gray scale, the size and Orientation of shadow region, increases the shadow region of Preliminary detection Subtract, it is thus achieved that pedestrian shadow region；Its detailed process is as follows:

Step C-1, according to the color average of pedestrian's agglomerate shadow region, rejects face in pedestrian's agglomerate shadow region Look and this average differ by more than the part of shadow color threshold value, and by equal with this for pedestrian's agglomerate non-hatched area color Value difference is set to candidate's shadows pixels less than the pixel of shadow color threshold value；

Step C-2, with pedestrian and Shadow segmentation point as initial point, sets the N of shadow region a length of pedestrian height Times；Reject in pedestrian's agglomerate shadow region pixel and initial point distance more than the part of this shadow region length, and will In pedestrian's agglomerate non-hatched area, pixel and initial point distance are set to candidate's shade less than the part of shadow region length Pixel；

Step C-3, with pedestrian and Shadow segmentation point as initial point, according to pedestrian's shade towards, by pedestrian's agglomerate district In territory, pixel and initial point distance project elliptical coordinate system, reject pixel and initial point in pedestrian's agglomerate shadow region Distance is more than the part of threshold value, and the distance of pixel in pedestrian's agglomerate non-hatched area with initial point is less than threshold value Part is set to candidate's shadows pixels；

Step C-4, takes common factor, the pedestrian shadow region after composition increase and decrease by candidate's shadows pixels of above-mentioned steps；

Step D, returns step A, until video terminates.

In step C-2, N times of described shadow region a length of pedestrian height, the span of N is: 0 < N < 0.8.

Described agglomerate refers to that image pixel spatially connects and have the region of identical image feature, described group The parameter of block includes: agglomerate area coordinate, context marker, Shadow marks, center agglomerate mark, agglomerate is numbered, Agglomerate centre coordinate, agglomerate color average.

The invention has the beneficial effects as follows: the present invention proposes a kind of pedestrian's cast shadow suppressing method based on agglomerate model, First described method builds pedestrian's agglomerate model, then by calculating moment characteristics and the agglomerate vertical histogram of agglomerate, Method by geometry obtains preliminary shadow region；Then, by the gray scale of this shadow region, direction, centre bit Putting, as parameter, whole pedestrian and shadow region are carried out shade modeling, the shadow region obtaining primary segmentation is entered Row is pruned.Difference is detected and suppresses towards, the shade of varying number pedestrian, at video by the inventive method Reason speed is about 7.5 frames/s, it is possible to separated with its shade by moving target efficiently and accurately.

Accompanying drawing explanation

Fig. 1 is the flow chart of a kind of based on agglomerate model pedestrian's cast shadow suppressing method of the present invention.

Detailed description of the invention

Below in conjunction with the accompanying drawings, a kind of based on agglomerate model the pedestrian's cast shadow suppressing method proposed the present invention is carried out Describe in detail:

As it is shown in figure 1, a kind of based on agglomerate model pedestrian's cast shadow suppressing method of the present invention, its step is as follows: Step 1 pedestrian's agglomerate models, and comprises the steps:

1.1 pairs of each two field pictures, are divided into the agglomerate of N*N and to calculate the pixel color in each agglomerate equal Value and centre coordinate；

1.2 difference calculating the color average between each agglomerate and centre coordinate distances, by the difference of color less than color Threshold value is also while centre coordinate distance merges less than the agglomerate of distance threshold, and amendment center agglomerate mark, sign is entirely The distribution of color of scene；

1.3 utilize the foreground template that mixed Gaussian background modeling algorithm extracts, the agglomerate of marker motion；

1.4 with the centre coordinate of each motion agglomerate as node, utilize fuzzy clustering algorithm by its centre coordinate away from Again merge from the agglomerate less than distance threshold, amendment center agglomerate mark, obtain pedestrian's agglomerate.

Step 2 shadow Detection, comprises the steps:

2.1 according to pedestrian's agglomerate array R, calculates agglomerate vertical histogram H_R(x) and pedestrian and the angle of shade θ_R；

Described agglomerate vertical histogram H_RX (), by traveling through array R and can add up Vertical Square in the horizontal direction Obtain to the quantity of agglomerate.

Described angle theta_R, can be calculated by the central moment of pedestrian's agglomerate array R (x):

${\left({\mathrm{\&mu;}}_{p,q}\right)}_R=\underset{(x,y)\&Element;R}{\&Sigma;}{(x-\stackrel{\&OverBar;}{x})}^p{(y-\stackrel{\&OverBar;}{y})}^q,(\stackrel{\&OverBar;}{x},\stackrel{\&OverBar;}{y})=(\frac{1}{n}\underset{(x,y)\&Element;R}{\&Sigma;}x,\frac{1}{n}\underset{(x,y)\&Element;R}{\&Sigma;}y)$

${\mathrm{\&theta;}}_R=\frac{1}{2}{\tan }^{-1}\left(\frac{2{\mathrm{\&mu;}}_{1,1}}{{\mathrm{\&mu;}}_{2,0}-{\mathrm{\&mu;}}_{0,2}}\right)$

Wherein, (μ_p,q)_RFor the square of pedestrian's agglomerate array R,For the geometric center of pedestrian's agglomerate array R, (x, y) is the geometric center of pedestrian's agglomerate array element R [i], and n is the length of array R, and p, q are respectively x, y The exponent number of square.

2.2 according to agglomerate vertical histogram H_RX () calculates agglomerate histogram of difference dH_R(x)；

dH_R(x)=| H_R(x)-max(H_R)|

Wherein, H_RX () is agglomerate vertical histogram H_RXth row component

2.3 calculate pedestrian and Shadow segmentation point P_RAbscissa x_R,

Wherein, X_RbottomThe abscissa span of agglomerate histogram of difference, i.e. pedestrian's agglomerate array area of coverage The width in territory, works as x_RWhen having multiple value, retain the point that distance agglomerate histogram crest is nearest.

2.4 scanning agglomerate model arrays R, it is thus achieved that agglomerate centre coordinate set C_R(x)；

2.5 pass through P_RAbscissa and C_RX (), obtains P_ROrdinate y_R, y_R=C_R(x_R)

2.6 utilize angle theta_RAnd P_RObtain the cut-off rule of pedestrian and shade.Calculate the face of cut-off rule both sides picture Look average and variance, select the region of color average and variance all minimums as preliminary shadow region R₂；

Y=mx+c, m=tan θ_R, c=y_R-x_R tanθ_R

Wherein, m is straight slope, and c is Linear intercept, (x_R, y_R) it is cut-point P_RCoordinate.

Step 3 shade model, to pixel each in the shadow region of primary segmentation from the color of shade, size, Angle such as 3, direction etc. determines whether, comprises the steps of

3.1 shade gray scale criterions,μ_RIt is primary segmentation shadow region R₂'s Gray average, T_hIt is threshold value, I_k(x y) is pedestrian's agglomerate area pixel point (x, y) gray value at place.

3.2 shade size criterion,(x_R,y_R) it is P_RPoint coordinates, (x is y) that pedestrian rolls into a ball control area pixel point coordinates, T_dFor threshold value, less than 0.8 times of pedestrian's height.

3.3 shade direction criterions, θ_RFor shade direction during primary segmentation, (s is t) that pixel divides with shade Cutpoint composition distance vector (x-x_R,y-y_R) according to shade direction θ_R, original coordinate space project oval seat Expression behind mark space, T_oritFor T_dExpression at ellipsoidal coordinate space.

3.4 merge three quasi-sides, the discrimination formula of structure increase and decrease shadow region:

$d(x,y)=w_1\sqrt{s^2+t^2}+w_2|I_k(x,y)-{\mathrm{\&mu;}}_R|$

w₁=α w₁+(1-α)R_dist, w₂=α w₂+(1-α)R_grey

$w_1=\frac{w_1}{w_1+w_2},w_2=\frac{w_2}{w_1+w_2}$

Wherein, α is weight renewal rate, takes 0.1, w herein₁It is shadow length weight, w₂It it is shade Gray scale weight.Weights can adjust automatically according to the matching degree of each factor.

Step 4 shade mark and suppression, carry out shade modeling to each pixel of foreground area in present frame, as Really (x, y) less than threshold value T for d_s, then this pixel can be divided into shadows pixels.Here T_sTake preliminary point Shadow region R when cutting₂The average of interior all pixel d values, n is shadow region R₂The quantity of interior pixel, T_s Expression formula be:

$T_s=\frac{1}{n}\underset{(x,y)\&Element;R_2}{\&Sigma;}d(x,y).$

Claims (3)

1. pedestrian's cast shadow suppressing method based on agglomerate model, it is characterised in that comprise the steps:

Step A, builds pedestrian's agglomerate, makes each pedestrian represent with corresponding agglomerate, and it specifically comprises the following steps that

Step A-1, obtains current frame image, utilizes agglomerate extracting method, the scene color to current frame image Cluster, obtain image agglomerate；

Step A-2, according to described image agglomerate, utilizes mixed Gaussian background modeling method, obtains present frame figure The prospect masterplate of picture, obtains foreground moving agglomerate；

Step A-3, utilizes fuzzy clustering method to merge foreground moving agglomerate, it is thus achieved that pedestrian's agglomerate；

Step B, for pedestrian's agglomerate, utilizes shadow detection method that with its shade, pedestrian is carried out primary segmentation, The shadow region of acquisition pedestrian's agglomerate:

Step B-1, calculates the central moment of pedestrian's agglomerate, so obtain the shade of pedestrian's agglomerate towards；

Step B-2, calculates pedestrian's agglomerate vertical histogram, obtains pedestrian and the cut-point of shade in pedestrian's agglomerate；

Step B-3, according to the shade of pedestrian's agglomerate towards the cut-point with pedestrian Yu shade, determines pedestrian respectively Agglomerate pedestrian is minimum with the color average threshold value of shadow region and variance threshold values, selection color average and color variance Region, as the shadow region of pedestrian's agglomerate；

Step C, according to gray scale, the size and Orientation of shadow region, increases the shadow region of Preliminary detection Subtract, it is thus achieved that pedestrian shadow region；Its detailed process is as follows:

Step C-1, according to the color average of pedestrian's agglomerate shadow region, rejects face in pedestrian's agglomerate shadow region Look and this average differ by more than the part of shadow color threshold value, and by equal with this for pedestrian's agglomerate non-hatched area color Value difference is set to candidate's shadows pixels less than the pixel of shadow color threshold value；

Step C-2, with pedestrian and Shadow segmentation point as initial point, sets the N of shadow region a length of pedestrian height Times；Reject in pedestrian's agglomerate shadow region pixel and initial point distance more than the part of this shadow region length, and will In pedestrian's agglomerate non-hatched area, pixel and initial point distance are set to candidate's shade less than the part of shadow region length Pixel；

Step C-3, with pedestrian and Shadow segmentation point as initial point, according to pedestrian's shade towards, by pedestrian's agglomerate district In territory, pixel and initial point distance project elliptical coordinate system, reject pixel and initial point in pedestrian's agglomerate shadow region Distance is more than the part of threshold value, and the distance of pixel in pedestrian's agglomerate non-hatched area with initial point is less than threshold value Part is set to candidate's shadows pixels；

Step C-4, takes common factor, the pedestrian shadow region after composition increase and decrease by candidate's shadows pixels of above-mentioned steps；

Step D, returns step A, until video terminates.

A kind of pedestrian's cast shadow suppressing method based on agglomerate model the most according to claim 1, it is characterised in that In step C-2, N times of described shadow region a length of pedestrian height, the span of N is: 0 < N < 0.8.

A kind of pedestrian's cast shadow suppressing method based on agglomerate model the most according to claim 1, it is characterised in that Described agglomerate refers to that image pixel spatially connects and has the region of similar image features；Described agglomerate Parameter includes: agglomerate area coordinate, context marker, Shadow marks, center agglomerate mark, agglomerate numbering, group Block centre coordinate, agglomerate color average.