CN103559498A - Rapid man and vehicle target classification method based on multi-feature fusion - Google Patents

Abstract

The invention belongs to the field of image processing and photoelectric technology, and particularly relates to a rapid man and vehicle target classification method based on multi-feature fusion. The method comprises the following steps of acquiring a monitoring video image and obtaining a difference image of two adjacent frames of images and carrying out image segmenting, noise removing, hole filling to form a target image, determining target areas on the target image and calculating the area of each target areas, the area of an external rectangle, a centroid coordinate and rectangular saturation, judging two target areas with small area difference and a closest centroid coordinate Euclidean distance in two adjacent frames of images as a same target, and if the same target stably appears on a plurality of frames, judging the target as a man or vehicle according to the area, speed and rectangular saturation of the target. According to the method, the rapid discrimination of moving target man and vehicle attributes is realized, the classification accuracy is high, and the real-time performance is good.

Description

Quick people's car objective classification method based on multi-feature fusion

Technical field

The invention belongs to image processing, field of photoelectric technology, be specifically related to a kind of quick people's car objective classification method based on multi-feature fusion.

Background technology

In video monitoring system, there will be a large amount of moving targets, and in whole moving targets, take personnel targets and vehicle target as main, this two classes target is also the highest priority of monitoring simultaneously.Because the management expectancy for personnel targets and vehicle target has remarkable difference, therefore in video monitoring system, there is the demand of people's car target classification.

People's car objective classification method of the prior art mainly, based on statistics training, need to be collected a large amount of vehicles and personnel's image pattern, and recognition speed is slow, higher to arithmetic facility demand, has greatly affected recognition effect.

Summary of the invention

The technical issues that need to address of the present invention are: people's car objective classification method of the prior art mainly, based on statistics training, need to be collected a large amount of vehicles and personnel's image pattern, and recognition efficiency is low.

Technical scheme of the present invention is as described below:

Quick people's car objective classification method, comprises the steps:

Obtain monitor video image, adjacent two two field pictures are carried out to inter-frame difference, obtain difference image; Adopt maximum variance between clusters definite threshold, difference image is carried out to image and cut apart; The noise in image is cut apart in removal, and filling cavity forms target image; Target image is carried out to pixels statistics, remove the low region of target image pixel value, form target area, maximal value and minimum value according to target area pixel transverse and longitudinal coordinate, construct the boundary rectangle in this region, in calculating target image, the area of each target area is, the area of boundary rectangle, center-of-mass coordinate and rectangle saturation degree; Two target areas that in adjacent two two field pictures, area discrepancy is little, center-of-mass coordinate Euclidean distance is nearest are judged to be to same target; If more than the some frames of the stable appearance of same target, according to area, speed and the rectangle saturation degree of this target, judge that this target is for personnel or vehicle.

As preferred version, method of the present invention comprises the following steps:

Step 1.

Obtain monitor video image, k two field picture is designated as F _k; Adjacent two two field pictures are carried out to inter-frame difference: get the absolute value of adjacent two two field picture respective pixel values differences, after pixel difference, obtain difference image FD _k;

Step 2.

Calculate difference image FD _kaverage μ and standard deviation sigma, for difference image FD _kin be greater than the pixel value of μ+σ, with maximum between-cluster variance criterion, calculate optimum segmentation threshold value Th, according to threshold value Th to difference image FD _kcarry out image and cut apart, form and cut apart image FS _k:

${\mathrm{FS}}_k(i,j)=\left\{\begin{array}{cc}1& {\mathrm{FD}}_k(i,j)\&GreaterEqual;\mathrm{Th}\\ 0& {\mathrm{FD}}_k(i,j)<\mathrm{Th}\end{array}\right.$

Wherein,

FD _k(i, j) represents difference image FD _kthe pixel value of the capable j row of i pixel;

FS _k(i, j) represents to cut apart image FS _kthe pixel value of the capable j row of i pixel;

Step 3

Adopt morphological operation to remove and cut apart the noise in image, filling cavity, forms target image FM _k;

Step 4

By target image FM _kmiddle pixel value is that 1 connected region is carried out pixel count statistics, removal pixel count is less than the region of N1, to the target area retaining, maximal value and minimum value according to pixel transverse and longitudinal coordinate, form the boundary rectangle in this region, in order to identify this target area, calculate the area S of each target area in image, the area RS of boundary rectangle, center-of-mass coordinate (x, y) and rectangle saturation degree R

Step 5

Each two field picture to monitor video carries out above-mentioned steps 1～step 4, add up the features such as the area of each two field picture target area, the area of target area boundary rectangle and target area center-of-mass coordinate, according to feature, carry out target association: establish and in k frame, have N target area, for its n target area, in k-1 frame, target area meeting area discrepancy and be less than n target area Euclidean distance minimum of centroid distance k frame under 10% condition, is judged to be same target;

Step 6

If same target is stable there is N2 frame more than, by many features identification, judge that these targets are personnel or vehicle:

Wherein,

D _s, d _v, d _rbe followed successively by area features weighting coefficient, velocity characteristic weighting coefficient, rectangle saturation degree characteristic weighing coefficient;

ThS ₁, ThS ₂be followed successively by the high and low threshold value of target area;

ThV ₁, ThV ₂be followed successively by the high and low threshold value of target velocity;

ThR ₁, ThR ₂be followed successively by the high and low threshold value of target rectangle saturation degree;

S _nit is the area of n target area of k frame;

V _nbe the speed of n target area of k frame;

R _nit is the rectangle saturation degree of n target area of k frame;

ThS ₁, ThS ₂, ThV ₁, ThV ₂, ThR ₁, ThR ₂numerical value is specifically set according to monitoring scene;

Setting threshold Th ₀if, F _n>=Th ₀, judge that this target is as vehicle; If F _n<Th ₀, judge that this target is as personnel.

As preferred version,

In step 3, adopt 5 * 5 square template to FS _kcorrode operation, remove noise; To removing image after noise, adopt 7 * 7 square template to FS _kcarry out expansive working, fill up the cavity in image, form target image FM _k.

As preferred version,

In step 6,

The method of calculating n target area speed of k frame is as described below, k >=2:

$\left\{\begin{array}{c}{V}_{\mathrm{nk}}=a\&times;D_{n(k-1)}+b\&times;D_{\mathrm{nk}}\\ D_{\mathrm{nk}}=\sqrt{{({\mathrm{xn}}_k-{\mathrm{xn}}_{k-1})}^2+{({\mathrm{yn}}_k-{\mathrm{yn}}_{k-1})}^2}\\ D_{n1}=0\end{array}\right.$

Wherein,

V _nkbe the speed of n target area of k frame;

A, b are weighting coefficient, and preferred value is a=0.7, b=0.3;

D _nkit is the barycenter displacement of n target area of k frame;

D _{n (k-1)}be in k-1 frame, to be judged as the barycenter displacement of the target area of same target with n target area of k frame;

(xn _k, yn _k) be the center-of-mass coordinate of n target area of k frame;

(xn _k-1, yn _k-1) be in k-1 frame, to be judged as the center-of-mass coordinate of the target area of same target with n target area of k frame.

As preferred version,

N1=60; N2=10; d _s=0.3, d _v=0.3, d _r=0.4; ThS ₁=2000 pixel counts, ThS ₂=6000 pixel counts; ThV ₁=4 pixel/frame, ThV ₂=8 pixel/frame; ThR ₁=0.5, ThR ₂=0.8; Th ₀=0.5.

Beneficial effect of the present invention is:

A kind of quick people's car objective classification method based on multi-feature fusion of the present invention, by extracting the features such as area, speed and rectangle saturation degree of moving target, realizes the Quick of moving target people car attribute, classification accurately high, real-time is good.

Accompanying drawing explanation

Fig. 1 is a kind of quick people's car objective classification method process flow diagram based on multi-feature fusion of the present invention.

Embodiment

Below in conjunction with drawings and Examples, a kind of quick people's car objective classification method based on multi-feature fusion of the present invention is elaborated.

A kind of quick people's car objective classification method based on multi-feature fusion of the present invention, comprises the following steps:

Step 1.

Obtain monitor video image, k two field picture is designated as F _k; Adjacent two two field pictures are carried out to inter-frame difference: get the absolute value of adjacent two two field picture respective pixel values differences, after pixel difference, obtain difference image FD _k, i.e. FD _k=| F _k-F _k-1|.

Step 2.

If difference image FD _kfor the capable c row of r, FD _k(i, j) represents the pixel value of the capable j row of i pixel, calculates difference image FD _kaverage μ and standard deviation sigma:

$\left\{\begin{array}{c}\mathrm{\&mu;}=\frac{1}{r\&times;c}\underset{i=1}{\overset{r}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{c}{\mathrm{\&Sigma;}}}{\mathrm{FD}}_k(i,j)\\ \mathrm{\&sigma;}=\sqrt{\frac{1}{r\&times;c-1}\underset{i=1}{\overset{r}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{c}{\mathrm{\&Sigma;}}}{({\mathrm{FD}}_k(i,j)-\mathrm{\&mu;})}^2}\end{array}\right..$

For difference image FD _kin be greater than the pixel value of μ+σ, with maximum between-cluster variance (OTSU) criterion, calculate optimum segmentation threshold value Th.

According to threshold value Th to difference image FD _kcarry out image and cut apart, form and cut apart image FS _k.Cut apart rule as follows:

${\mathrm{FS}}_k(i,j)=\left\{\begin{array}{cc}1& {\mathrm{FD}}_k(i,j)\&GreaterEqual;\mathrm{Th}\\ 0& {\mathrm{FD}}_k(i,j)<\mathrm{Th}\end{array}\right.$

Wherein, FS _k(i, j) represents to cut apart image FS _kthe pixel value of the capable j row of i pixel.

Step 3

Adopt morphological operation to remove and cut apart the noise in image, filling cavity: adopt 5 * 5 square template to FS _kcorrode operation, remove noise; To removing image after noise, adopt 7 * 7 square template to FS _kcarry out expansive working, fill up the cavity in image, form target image FM _k.

Step 4

By target image FM _kmiddle pixel value is that 1 connected region is carried out pixel count statistics, removes the region that pixel count is less than N1, N1=60 in the present embodiment.To the target area retaining, according to maximal value and the minimum value of pixel transverse and longitudinal coordinate, form the boundary rectangle in this region, in order to identify this target area, calculate the area S of each target area in image, the area RS of boundary rectangle, center-of-mass coordinate (x, y) and rectangle saturation degree R

Step 5

Each two field picture to monitor video carries out above-mentioned steps 1～step 4, add up the features such as the area of each two field picture target area, the area of target area boundary rectangle and target area center-of-mass coordinate, according to feature, carry out target association: establish and in k frame, have N target area, for its n target area, in k-1 frame, target area meeting area discrepancy and be less than n target area Euclidean distance minimum of centroid distance k frame under 10% condition, is judged to be same target.

Step 6

If same target is stable there is N2 frame more than, by many features identification, judge that these targets are personnel or vehicle:

Wherein,

D _s, d _v, d _rbe followed successively by area features weighting coefficient, velocity characteristic weighting coefficient, rectangle saturation degree characteristic weighing coefficient;

ThS ₁, ThS ₂be followed successively by the high and low threshold value of target area;

ThV ₁, ThV ₂be followed successively by the high and low threshold value of target velocity;

ThR ₁, ThR ₂be followed successively by the high and low threshold value of target rectangle saturation degree;

S _nit is the area of n target area of k frame;

V _nbe the speed of n target area of k frame;

R _nit is the rectangle saturation degree of n target area of k frame.

ThS ₁, ThS ₂, ThV ₁, ThV ₂, ThR ₁, ThR ₂numerical value is specifically set according to monitoring scene.

In the present embodiment, N2=10; d _s=0.3, d _v=0.3, d _r=0.4; ThS ₁=2000 pixel counts, ThS ₂=6000 pixel counts; ThV ₁=4 pixel/frame, ThV ₂=8 pixel/frame; ThR ₁=0.5, ThR ₂=0.8.

Wherein, calculate the method for n target area speed of k frame (k >=2) as described below:

$\left\{\begin{array}{c}{V}_{\mathrm{nk}}=a\&times;D_{n(k-1)}+b\&times;D_{\mathrm{nk}}\\ D_{\mathrm{nk}}=\sqrt{{({\mathrm{xn}}_k-{\mathrm{xn}}_{k-1})}^2+{({\mathrm{yn}}_k-{\mathrm{yn}}_{k-1})}^2}\\ D_{n1}=0\end{array}\right.$

Wherein,

V _nkbe the speed of n target area of k frame;

A, b are weighting coefficient, in the present embodiment, and a=0.7, b=0.3;

D _nkit is the barycenter displacement of n target area of k frame;

D _{n (k-1)}be in k-1 frame, to be judged as the barycenter displacement of the target area of same target with n target area of k frame;

(xn _k, yn _k) be the center-of-mass coordinate of n target area of k frame;

(xn _k-1, yn _k-1) be in k-1 frame, to be judged as the center-of-mass coordinate of the target area of same target with n target area of k frame.

Consider that personnel targets area is little, vehicle target area is large, personnel targets movement velocity is little, vehicle target movement velocity is large, and personnel targets rectangle saturation degree is low, vehicle target rectangle saturation degree is high, can carry out the classification of people's car by following criterion:

Setting threshold Th ₀if, F _n>=Th ₀, judge that this target is as vehicle; If F _n<Th ₀, judge that this target is as personnel.In the present embodiment, Th ₀=0.5.

Claims (6)

1. quick people's car objective classification method based on multi-feature fusion, is characterized in that: comprise the steps:

Obtain monitor video image, adjacent two two field pictures are carried out to inter-frame difference, obtain difference image; Adopt maximum variance between clusters definite threshold, difference image is carried out to image and cut apart; The noise in image is cut apart in removal, and filling cavity forms target image; Target image is carried out to pixels statistics, remove the low region of target image pixel value, form target area, maximal value and minimum value according to target area pixel transverse and longitudinal coordinate, construct the boundary rectangle in this region, in calculating target image, the area of each target area is, the area of boundary rectangle, center-of-mass coordinate and rectangle saturation degree; Two target areas that in adjacent two two field pictures, area discrepancy is little, center-of-mass coordinate Euclidean distance is nearest are judged to be to same target; If more than the some frames of the stable appearance of same target, according to area, speed and the rectangle saturation degree of this target, judge that this target is for personnel or vehicle.

2. quick people's car objective classification method based on multi-feature fusion according to claim 1, is characterized in that: comprise the following steps:

Step 1.

Obtain monitor video image, k two field picture is designated as F _k; Adjacent two two field pictures are carried out to inter-frame difference: get the absolute value of adjacent two two field picture respective pixel values differences, after pixel difference, obtain difference image FD _k;

Step 2.

Calculate difference image FD _kaverage μ and standard deviation sigma, for difference image FD _kin be greater than the pixel value of μ+σ, with maximum between-cluster variance criterion, calculate optimum segmentation threshold value Th, according to threshold value Th to difference image FD _kcarry out image and cut apart, form and cut apart image FS _k:

Wherein,

FD _k(i, j) represents difference image FD _kthe pixel value of the capable j row of i pixel;

FS _k(i, j) represents to cut apart image FS _kthe pixel value of the capable j row of i pixel;

Step 3

Adopt morphological operation to remove and cut apart the noise in image, filling cavity, forms target image FM _k;

Step 4

By target image FM _kmiddle pixel value is that 1 connected region is carried out pixel count statistics, removal pixel count is less than the region of N1, to the target area retaining, maximal value and minimum value according to pixel transverse and longitudinal coordinate, form the boundary rectangle in this region, in order to identify this target area, calculate the area S of each target area in image, the area RS of boundary rectangle, center-of-mass coordinate (x, y) and rectangle saturation degree R

Step 5

Each two field picture to monitor video carries out above-mentioned steps 1～step 4, add up the features such as the area of each two field picture target area, the area of target area boundary rectangle and target area center-of-mass coordinate, according to feature, carry out target association: establish and in k frame, have N target area, for its n target area, in k-1 frame, target area meeting area discrepancy and be less than n target area Euclidean distance minimum of centroid distance k frame under 10% condition, is judged to be same target;

Step 6

If same target is stable there is N2 frame more than, by many features identification, judge that these targets are personnel or vehicle:

Wherein,

D _s, d _v, d _rbe followed successively by area features weighting coefficient, velocity characteristic weighting coefficient, rectangle saturation degree characteristic weighing coefficient;

ThS ₁, ThS ₂be followed successively by the high and low threshold value of target area;

ThV ₁, ThV ₂be followed successively by the high and low threshold value of target velocity;

ThR ₁, ThR ₂be followed successively by the high and low threshold value of target rectangle saturation degree;

S _nit is the area of n target area of k frame;

V _nbe the speed of n target area of k frame;

R _nit is the rectangle saturation degree of n target area of k frame;

ThS ₁, ThS ₂, ThV ₁, ThV ₂, ThR ₁, ThR ₂numerical value is specifically set according to monitoring scene;

Setting threshold Th ₀if, F _n>=Th ₀, judge that this target is as vehicle; If F _n<Th ₀, judge that this target is as personnel.

3. quick people's car objective classification method based on multi-feature fusion according to claim 2, is characterized in that:

In step 6,

The method of calculating n target area speed of k frame is as described below, k >=2:

Wherein,

V _nkbe the speed of n target area of k frame;

A, b are weighting coefficient;

D _nkit is the barycenter displacement of n target area of k frame;

D _{n (k-1)}be in k-1 frame, to be judged as the barycenter displacement of the target area of same target with n target area of k frame;

(xn _k, yn _k) be the center-of-mass coordinate of n target area of k frame;

(xn _k-1, yn _k-1) be in k-1 frame, to be judged as the center-of-mass coordinate of the target area of same target with n target area of k frame.

4. quick people's car objective classification method based on multi-feature fusion according to claim 3, is characterized in that: a=0.7, b=0.3.

5. according to the based on multi-feature fusion quick people's car objective classification method described in claim 2 or 3, it is characterized in that: in step 3, adopt 5 * 5 square template to FS _kcorrode operation, remove noise; To removing image after noise, adopt 7 * 7 square template to FS _kcarry out expansive working, fill up the cavity in image, form target image FM _k.

6. according to the based on multi-feature fusion quick people's car objective classification method described in claim 2 or 3 or 4, it is characterized in that: N1=60; N2=10; d _s=0.3, d _v=0.3, d _r=0.4; ThS ₁=2000 pixel counts, ThS ₂=6000 pixel counts; ThV ₁=4 pixel/frame, ThV ₂=8 pixel/frame; ThR ₁=0.5, ThR ₂=0.8; Th ₀=0.5.

Images