CN106203360A - Intensive scene crowd based on multistage filtering model hives off detection algorithm - Google Patents

Abstract

A kind of intensive scene crowd based on multistage filtering model hives off detection algorithm, relates to image analysis technology field, and solved is the technical problem improving the Detection results that hives off.The foreground area that this algorithm obtains in the background subtraction of gauss hybrid models extracts KLT characteristic point, by analyzing the kinetic characteristic of characteristic point, time-space matrix is used to filter step by step adjacent to filter operator, velocity attitude filter operator and motion relevance filter operator, travel through the characteristic point in all prospects, it is achieved detection of hiving off.The algorithm that the present invention provides, it is adaptable to video image analysis.

Description

Intensive scene crowd based on multistage filtering model hives off detection algorithm

Technical field

The present invention relates to image analysis technology, particularly relate to a kind of intensive scene crowd based on multistage filtering model and divide The technology of group's detection algorithm.

Background technology

Crowd movement's analysis, social event and unusual checking in intensive scene are current intelligent monitoring research fields Hot subject, be also video monitoring system and one of the important directions of intelligent transportation research.Maintaining public order, improving city The aspects such as the science of city's traffic programme have important researching value.

Generally in dense population moving scene, crowd's common manifestation is a kind of irregular movement combination form, the most unordered Motion.The characteristic of this complexity substantially increases the difficulty of research, and therefore the rarest achievement in research relates to before.For close The traditional analysis thinking of collection scene is that scene is regarded as the motion set of target, to realize extracting target trajectory and Activity recognition Purpose.Its processing method has two classes, a class to be processing methods based on target's center's object, and another kind of is process based on scene Method.

Processing method based on target's center's object regards dense population as a large amount of group of individuals, from ontoanalysis angle Set out extraction its speed and direction, then use the connection figure of weighting or multi-level clustering method from bottom to up to realize colony's inspection Survey；The effectiveness guaranteed conditions of this algorithm is that target gather density is low, and the pixel resolution of target individual is high.But disordered motion Blocking very serious in intensive scene between crowd and hiding relation is the most unknown, cause target is correctly partitioned into a difficulty Point, effect of therefore hiving off is poor.

Processing method based on scene mainly has optical flow method, dynamic texture and grid particle etc., and these algorithms are for intensive The treatment effect that there is the scene of serious occlusion issue and compound movement pattern scene in crowd is poor, but also needs single Pedestrian carries out splitting and sample training, needs some prior informations are provided in advance；And the less stable of neighborhood characteristics point, The Detection results that hives off needs to be improved further.

Summary of the invention

For defect present in above-mentioned prior art, the technical problem to be solved is to provide a kind of to intensive Crowd exists scene and the high treating effect of compound movement pattern scene of serious occlusion issue, and need not single row People carries out splitting and sample training, it is not necessary to prior information is provided in advance；And the good stability of neighborhood characteristics point, inspection of hiving off Survey effective based on multistage filtering model intensive scene crowd to hive off detection algorithm.

In order to solve above-mentioned technical problem, a kind of intensive scene crowd based on multistage filtering model provided by the present invention Hive off detection algorithm, it is characterised in that specifically comprises the following steps that

S1) to target video image, first use mixed Gauss model that each picture frame is carried out background modeling, then use Prospect, the background of each picture frame are split by Background difference, thus obtain the foreground area of each picture frame；

S2) to target video image, KLT track algorithm is used to extract feature from the foreground area of each picture frame Point, thus obtain the characteristic point coordinate set of each picture frame；

S3) the characteristic point coordinate set to each picture frame, each characteristic point that this feature point coordinates is concentrated work one by one For impact point, and obtain initial same group's point set of each impact point according to the method for step S3.1 to S3.9, final obtain each Initial same group's point set of individual impact point constitutes initial same group's sequence of point sets of this picture frame；

The step of the initial same group's point set obtaining impact point is as follows:

S3.1) from ζ_tIn take a characteristic point as impact point i, and setFor the nearest-neighbor of impact point i, wherein, ζ_tCharacteristic point coordinate set for t picture frame in target video image；

S3.2) impact point i and ζ is calculated_tIn the Gauss weights of each characteristic point, specific formula for calculation is:

$w_{i,j}=\exp (-\frac{dist(i,j)}{r})=\exp (-\frac{\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}}{r})$

Wherein, w_i,jFor the Gauss weights of impact point i Yu characteristic point j, j ∈ ζ_tAnd j ≠ i, (x_i,y_i) it is the seat of impact point i Mark, (x_j,y_j) it being characterized the coordinate of a j, the value of r is 20；

S3.3) to ζ_tIn each characteristic point in addition to impact point i implement that time-space matrix is neighbouring to be filtered, filter method is:

To ζ_tIn any one characteristic point j, j ∈ ζ_tAnd j ≠ i, if there being w_i,j ≥0.5w_max, then characteristic point j is included intoW therein_maxFor impact point i and ζ_tIn further feature point Gauss weights in maximum；

S3.4) setOccur simultaneously for impact point i neighborhood between t → t+d, d=therein 3；

S3.5) calculateIn each characteristic point and impact point i from the velocity angle of t → t+d, specific formula for calculation For:

${\mathrm{\θ}}_{i,j}=\frac{1}{d+1}\underset{\mathrm{\τ}=t}{\overset{t+d}{\Σ}}abs({\mathrm{\θ}}_i^{\mathrm{\τ}}-{\mathrm{\θ}}_j^{\mathrm{\τ}})$

IfAndThen make

IfAndThen make

IfOrThen make

IfAndThen

IfAndThen

IfOrThen

Wherein, θ_i,jIt is characterized a j and impact point i from the velocity angle of t → t+d, d=3, It is characterized The velocity attitude of some j τ picture frame in target video image,It is characterized a j in target video image The coordinate of τ picture frame,It is characterized the seat of a j+1 picture frame of τ in target video image Mark；For the velocity attitude of impact point i τ picture frame in target video image,For impact point i in target The coordinate of τ picture frame in video image,For impact point i τ+1 figure in target video image Coordinate as frame；

S3.6) forIn any one characteristic point j, if there being θ_i,j>=45 or (360-θ_i,j) >=45, then will This feature point j fromIn shave and remove；

S3.7) calculateIn each characteristic point and impact point i from the motion relevance of t → t+d, specifically calculate public affairs Formula is:

$C_{i,j}=\frac{1}{d+1}\underset{\mathrm{\τ}=t}{\overset{t+d}{\Σ}}\frac{v_{\mathrm{\τ}}^i\·v_{\mathrm{\τ}}^j}{\left|\right|v_{\mathrm{\τ}}^i\left|\right|\·\left|\right|v_{\mathrm{\τ}}^j\left|\right|}$

$v_{\mathrm{\τ}}^i=(v_{x_i}^{\mathrm{\τ}},v_{y_i}^{\mathrm{\τ}})=(x_i^{\mathrm{\τ}+1}-x_i^{\mathrm{\τ}},y_i^{\mathrm{\τ}+1}-y_i^{\mathrm{\τ}})$

$\left|\right|v_{\mathrm{\τ}}^i\left|\right|=\sqrt{{\left(v_{x_i}^{\mathrm{\τ}}\right)}^2+{\left(v_{y_i}^{\mathrm{\τ}}\right)}^2}$

$v_{\mathrm{\τ}}^j=(v_{x_j}^{\mathrm{\τ}},v_{y_j}^{\mathrm{\τ}})=(x_j^{\mathrm{\τ}+1}-x_j^{\mathrm{\τ}},y_j^{\mathrm{\τ}+1}-y_j^{\mathrm{\τ}})$

$\left|\right|v_{\mathrm{\τ}}^j\left|\right|=\sqrt{{\left(v_{x_j}^{\mathrm{\τ}}\right)}^2+{\left(v_{y_j}^{\mathrm{\τ}}\right)}^2}$

Wherein, C_i,jIt is characterized a j and impact point i from the motion relevance of t → t+d, d=3, For mesh The speed of punctuate i τ picture frame in target video image,It is characterized a j the τ figure in target video image As the speed of frame,For the coordinate of impact point i τ picture frame in target video image,For The coordinate of the impact point i+1 picture frame of τ in target video image；It is characterized a j at target video image In the coordinate of τ picture frame,It is characterized the seat of a j+1 picture frame of τ in target video image Mark；

S3.8) C is set_th=0.6, forIn any one characteristic point j, if there being C_i,j≤C_th, then should Characteristic point j fromIn shave and remove；

S3.9) willIt is defined as initial same group's point set of impact point i

S4) the initial same group's sequence of point sets to each picture frame, this is initially initially same with each in group's sequence of point sets Group's point set is according to characteristic point number descending；

S5) the initial same group's sequence of point sets to each picture frame, first to this according to the method for step S5.1 to step S5.3 Begin initially to carry out key words sorting with group's point set with each in group's sequence of point sets；

S5.1) make K=1, L=1, all characteristic points that k-th initially puts concentration with group are labeled as L；

S5.2) K=K+1 is made；

If k-th initially puts the most unmarked mistake of all characteristic points of concentration with group, then make L=L+1, and by the beginning of k-th The all characteristic points beginning to put concentration with group are collectively labeled as L；

If k-th is initially put a concentration at least characteristic point with group and is had been labeled as L, then by k-th initially with group's point The all characteristic points concentrated are collectively labeled as L；

S5.3) step S5.2 is repeated, until initially each the most complete with the equal labelling of group's point set with in group's sequence of point sets；

S6) the characteristic point coordinate set to each picture frame, is classified as same class colony characteristic point identical for key words sorting, and This picture frame indicates characteristic point identical for key words sorting by identical color, and contingency table in this picture frame Remember that different characteristic point indicates by different color, thus the detection that realizes hiving off.

Based on multistage filtering model the intensive scene crowd that the present invention provides hives off detection algorithm, has following useful effect Really:

1) kinestate of the state estimation crowd of the characteristic point on employing moving target crowd, can avoid intensive Occlusion issue serious in crowd, the high treating effect to compound movement pattern scene；

2) need not single pedestrian is split and sample training, it is not required that any prior information；

3) time-space matrix is neighbouring filter in the feature of nearest-neighbor count can be according to the beeline of neighbor distance characteristic point Automatically adjust, and the neighborhood that the neighborhood choice continuous videos image of characteristic point obtains occurs simultaneously, it is ensured that neighborhood characteristics point Stability；

4) algorithm not only considers the dependency of group movement, and considers velocity attitude concordance, detection effect of hiving off The best.

Accompanying drawing explanation

Fig. 1 is that based on multistage filtering model the intensive scene crowd of the embodiment of the present invention hives off the principle of detection algorithm Figure.

Detailed description of the invention

Illustrate embodiments of the invention are described in further detail below in conjunction with accompanying drawing, but the present embodiment is not used to limit The present invention processed, the analog structure of every employing present invention and similar change thereof, protection scope of the present invention, the present invention all should be listed in In pause mark all represent the relation of sum.

As it is shown in figure 1, a kind of based on multistage filtering model the intensive scene crowd that the embodiment of the present invention is provided hives off Detection algorithm, it is characterised in that specifically comprise the following steps that

S1) to target video image, first use mixed Gauss model that each picture frame is carried out background modeling, then use Prospect, the background of each picture frame are split by Background difference, thus obtain the foreground area of each picture frame；

S2) to target video image, KLT track algorithm is used to extract feature from the foreground area of each picture frame Point, thus obtain the characteristic point coordinate set of each picture frame；

S3) the characteristic point coordinate set to each picture frame, each characteristic point that this feature point coordinates is concentrated work one by one For impact point, and obtain initial same group's point set of each impact point according to the method for step S3.1 to S3.9, final obtain each Initial same group's point set of individual impact point constitutes initial same group's sequence of point sets of this picture frame；

The step of the initial same group's point set obtaining impact point is as follows:

S3.1) from ζ_tIn take a characteristic point as impact point i, and setFor the nearest-neighbor of impact point i, wherein, ζ_tCharacteristic point coordinate set for t picture frame in target video image；

S3.2) impact point i and ζ is calculated_tIn the Gauss weights of each characteristic point, specific formula for calculation is:

$w_{i,j}=\exp (-\frac{dist(i,j)}{r})=\exp (-\frac{\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}}{r})$

Wherein, w_i,jFor the Gauss weights of impact point i Yu characteristic point j, j ∈ ζ_tAnd j ≠ i, (x_i,y_i) it is the seat of impact point i Mark, (x_j,y_j) it being characterized the coordinate of a j, r is constant, the representative value 20 of r；w_i,jWill along with distance dist (i, j) increase and subtract Little, on the contrary then increase, and this also just means that crowd is the most intensive, w_i,jThe biggest；

S3.3) to ζ_tIn each characteristic point in addition to impact point i implement that time-space matrix is neighbouring to be filtered, filter method is:

To ζ_tIn any one characteristic point j, j ∈ ζ_tAnd j ≠ i, if there being w_i,j ≥0.5w_max, then characteristic point j is included intoW therein_maxFor impact point i and ζ_tIn further feature point Gauss weights in maximum；

ObtainIn, the number of characteristic point can be according to w_maxIt is automatically adjusted, w_maxIt is impact point i and the spy nearest away from it Gauss weights between levying a little；

S3.4) setOccur simultaneously for impact point i neighborhood between t → t+d, d=therein 3；T → t+d refers to that the t picture frame of target video image is to the t+d picture frame；

S3.5) calculateIn each characteristic point and impact point i from the velocity angle of t → t+d, specific formula for calculation For:

${\mathrm{\θ}}_{i,j}=\frac{1}{d+1}\underset{\mathrm{\τ}=t}{\overset{t+d}{\Σ}}abs({\mathrm{\θ}}_i^{\mathrm{\τ}}-{\mathrm{\θ}}_j^{\mathrm{\τ}})$

IfAndThen make

IfAndThen make

IfOrThen make

IfAndThen

IfAndThen

IfOrThen

Wherein, θ_i,jIt is characterized a j and impact point i from the velocity angle of t → t+d, d=3, It is characterized The velocity attitude of some j τ picture frame in target video image,It is characterized a j in target video image The coordinate of τ picture frame,It is characterized the seat of a j+1 picture frame of τ in target video image Mark；For the velocity attitude of impact point i τ picture frame in target video image,For impact point i in target The coordinate of τ picture frame in video image,For impact point i τ+1 figure in target video image Coordinate as frame；

S3.6) forIn any one characteristic point j, if hadOr (360-θ_i,j) >=45, then will This feature point j fromIn shave and remove；

S3.7) calculateIn each characteristic point and impact point i from the motion relevance of t → t+d, specifically calculate Formula is:

$C_{i,j}=\frac{1}{d+1}\underset{\mathrm{\τ}=t}{\overset{t+d}{\Σ}}\frac{v_{\mathrm{\τ}}^i\·v_{\mathrm{\τ}}^j}{\left|\right|v_{\mathrm{\τ}}^i\left|\right|\·\left|\right|v_{\mathrm{\τ}}^j\left|\right|}$

$v_{\mathrm{\τ}}^i=(v_{x_i}^{\mathrm{\τ}},v_{y_i}^{\mathrm{\τ}})=(x_i^{\mathrm{\τ}+1}-x_i^{\mathrm{\τ}},y_i^{\mathrm{\τ}+1}-y_i^{\mathrm{\τ}})$

$\left|\right|v_{\mathrm{\τ}}^i\left|\right|=\sqrt{{\left(v_{x_i}^{\mathrm{\τ}}\right)}^2+{\left(v_{y_i}^{\mathrm{\τ}}\right)}^2}$

$v_{\mathrm{\τ}}^j=(v_{x_j}^{\mathrm{\τ}},v_{y_j}^{\mathrm{\τ}})=(x_j^{\mathrm{\τ}+1}-x_j^{\mathrm{\τ}},y_j^{\mathrm{\τ}+1}-y_j^{\mathrm{\τ}})$

$\left|\right|v_{\mathrm{\τ}}^j\left|\right|=\sqrt{{\left(v_{x_j}^{\mathrm{\τ}}\right)}^2+{\left(v_{y_j}^{\mathrm{\τ}}\right)}^2}$

Wherein, C_i,jIt is characterized a j and impact point i from the motion relevance of t → t+d, d=3, For mesh The speed of punctuate i τ picture frame in target video image,It is characterized a j the τ figure in target video image As the speed of frame,For the coordinate of impact point i τ picture frame in target video image,For The coordinate of the impact point i+1 picture frame of τ in target video image；It is characterized a j at target video image In the coordinate of τ picture frame,It is characterized the seat of a j+1 picture frame of τ in target video image Mark；

S3.8) C is set_th=0.6, forIn any one characteristic point j, if there being C_i,j≤C_th, then should Characteristic point j fromIn shave and remove；

S3.9) willIt is defined as initial same group's point set of impact point i In all characteristic points all and mesh The same group of punctuate i；

S4) the initial same group's sequence of point sets to each picture frame, this is initially initially same with each in group's sequence of point sets Group's point set is according to characteristic point number descending；

S5) the initial same group's sequence of point sets to each picture frame, first to this according to the method for step S5.1 to step S5.3 Begin initially to carry out key words sorting with group's point set with each in group's sequence of point sets；

S5.1) make K=1, L=1, all characteristic points that k-th initially puts concentration with group are labeled as L；

S5.2) K=K+1 is made；

If k-th initially puts the most unmarked mistake of all characteristic points of concentration with group, then make L=L+1, and by the beginning of k-th The all characteristic points beginning to put concentration with group are collectively labeled as L；

If k-th is initially put a concentration at least characteristic point with group and is had been labeled as L, then by k-th initially with group's point The all characteristic points concentrated are collectively labeled as L；

S5.3) step S5.2 is repeated, until initially each the most complete with the equal labelling of group's point set with in group's sequence of point sets；

S6) the characteristic point coordinate set to each picture frame, is classified as same class colony characteristic point identical for key words sorting, and This picture frame indicates characteristic point identical for key words sorting by identical color, and contingency table in this picture frame Remember that different characteristic point indicates by different color, thus the detection that realizes hiving off.

Claims (1)

1. an intensive scene crowd based on multistage filtering model hives off detection algorithm, it is characterised in that specifically comprise the following steps that

S1) to target video image, first use mixed Gauss model that each picture frame is carried out background modeling, then use background Prospect, the background of each picture frame are split by difference method, thus obtain the foreground area of each picture frame；

S2) to target video image, KLT track algorithm is used to extract characteristic point from the foreground area of each picture frame, from And obtain the characteristic point coordinate set of each picture frame；

S3) the characteristic point coordinate set to each picture frame, using this feature point coordinates concentrate each characteristic point one by one as mesh Punctuate, and initial same group's point set of each impact point is obtained according to the method for step S3.1 to S3.9, final each mesh obtained Initial same group's point set of punctuate constitutes initial same group's sequence of point sets of this picture frame；

The step of the initial same group's point set obtaining impact point is as follows:

S3.1) from ζ_tIn take a characteristic point as impact point i, and setFor the nearest-neighbor of impact point i, wherein, ζ_tFor The characteristic point coordinate set of t picture frame in target video image；

S3.2) impact point i and ζ is calculated_tIn the Gauss weights of each characteristic point, specific formula for calculation is:

$w_{i,j}=\exp \left(-\frac{dist\left(i,j\right)}{r}\right)=\exp \left(-\frac{\sqrt{{\left(x_i-x_j\right)}^2+{\left(y_i-y_j\right)}^2}}{r}\right)$

Wherein, w_i,jFor the Gauss weights of impact point i Yu characteristic point j, j ∈ ζ_tAnd j ≠ i, (x_i,y_i) it is the coordinate of impact point i, (x_j,y_j) it being characterized the coordinate of a j, the value of r is 20；

S3.3) to ζ_tIn each characteristic point in addition to impact point i implement that time-space matrix is neighbouring to be filtered, filter method is:

To ζ_tIn any one characteristic point j, j ∈ ζ_tAnd j ≠ i, if there being w_i,j>=0.5wmax, then be included into characteristic point j W therein_maxFor impact point i and ζ_tIn further feature point Gauss weights in maximum；

S3.4) setOccur simultaneously for impact point i neighborhood between t → t+d, d=3 therein；

S3.5) calculateIn each characteristic point and impact point i from the velocity angle of t → t+d, specific formula for calculation be:

${\mathrm{\θ}}_{i,j}=\frac{1}{d+1}\underset{\mathrm{\τ}=t}{\overset{t+d}{\Σ}}abs({\mathrm{\θ}}_i^{\mathrm{\τ}}-{\mathrm{\θ}}_j^{\mathrm{\τ}})$

IfAndThen make

IfAndThen make

IfOrThen make

IfAndThen

IfAndThen

IfOrThen

Wherein, θ_i,jIt is characterized a j and impact point i from the velocity angle of t → t+d, d=3, It is characterized a j The velocity attitude of τ picture frame in target video image,It is characterized a j in target video image The coordinate of τ picture frame,It is characterized the coordinate of a j+1 picture frame of τ in target video image； For the velocity attitude of impact point i τ picture frame in target video image,For impact point i at target video The coordinate of τ picture frame in image,For impact point i+1 picture frame of τ in target video image Coordinate；

S3.6) forIn any one characteristic point j, if there being θ_i,j>=45 or (360-θ_i,j) >=45, then by this spy Levy a j fromIn shave and remove；

S3.7) calculateIn each characteristic point and impact point i from the motion relevance of t → t+d, specific formula for calculation For:

$C_{i,j}=\frac{1}{d+1}\underset{\mathrm{\τ}=t}{\overset{t+d}{\Σ}}\frac{v_{\mathrm{\τ}}^i\·v_{\mathrm{\τ}}^j}{\left|\right|v_{\mathrm{\τ}}^i\left|\right|\·\left|\right|v_{\mathrm{\τ}}^j\left|\right|}$

$v_{\mathrm{\τ}}^i=(v_{x_i}^{\mathrm{\τ}},v_{y_i}^{\mathrm{\τ}})=(x_i^{\mathrm{\τ}+1}-x_i^{\mathrm{\τ}},y_i^{\mathrm{\τ}+1}-y_i^{\mathrm{\τ}})$

$\left|\right|v_{\mathrm{\τ}}^i\left|\right|=\sqrt{{\left(v_{x_i}^{\mathrm{\τ}}\right)}^2+{\left(v_{y_i}^{\mathrm{\τ}}\right)}^2}$

$v_{\mathrm{\τ}}^j=(v_{x_j}^{\mathrm{\τ}},v_{y_j}^{\mathrm{\τ}})=(x_j^{\mathrm{\τ}+1}-x_j^{\mathrm{\τ}},y_j^{\mathrm{\τ}+1}-y_j^{\mathrm{\τ}})$

$\left|\right|v_{\mathrm{\τ}}^j\left|\right|=\sqrt{{\left(v_{x_j}^{\mathrm{\τ}}\right)}^2+{\left(v_{y_j}^{\mathrm{\τ}}\right)}^2}$

Wherein, C_i,jIt is characterized a j and impact point i from the motion relevance of t → t+d, d=3, For impact point i The speed of τ picture frame in target video image,It is characterized a j the τ picture frame in target video image Speed,For the coordinate of impact point i τ picture frame in target video image,For target The coordinate of the some i+1 picture frame of τ in target video image；It is characterized a j in target video image The coordinate of τ picture frame,It is characterized the coordinate of a j+1 picture frame of τ in target video image；

S3.8) C is set_th=0.6, forIn any one characteristic point j, if there being C_i,j≤C_th, then by this feature Point j fromIn shave and remove；

S3.9) willIt is defined as initial same group's point set of impact point i

S4) the initial same group's sequence of point sets to each picture frame, by this initially with each in group's sequence of point sets initially with group's point Collection is according to characteristic point number descending；

S5) the initial same group's sequence of point sets to each picture frame, initially same to this according to the method for step S5.1 to step S5.3 Group's sequence of point sets each initially carries out key words sorting with group's point set；

S5.1) make K=1, L=1, all characteristic points that k-th initially puts concentration with group are labeled as L；

S5.2) K=K+1 is made；

If k-th initially puts the most unmarked mistake of all characteristic points of concentration with group, then make L=L+1, and k-th is initially same All characteristic points that group's point is concentrated are collectively labeled as L；

If k-th is initially put a concentration at least characteristic point with group and had been labeled as L, then k-th is initially put concentration with group All characteristic points be collectively labeled as L；

S5.3) step S5.2 is repeated, until initially each the most complete with the equal labelling of group's point set with in group's sequence of point sets；

S6) the characteristic point coordinate set to each picture frame, is classified as same class colony characteristic point identical for key words sorting, and at this Picture frame indicates characteristic point identical for key words sorting by identical color, and key words sorting phase in this picture frame Different characteristic point indicates by different color, thus the detection that realizes hiving off.