CN103400142B - A kind of pedestrian counting method - Google Patents

Abstract

The present invention discloses the pedestrian counting method of a kind of computer video processing technology field, and step is: monitor video carries out pedestrian's foreground pixel and the extraction of light stream feature; Then the foreground pixel feature extracted is normalized; Carry out pedestrian's occlusion detection simultaneously and block hierarchical estimation; Comprehensive pedestrian's characteristic sum blocks grade, Statistics Bar people's number. The feature that the present invention gets over line pedestrian by statistics carries out people stream counting, thus avoids pedestrian detection process; The ROI region of statistical computation simultaneously around virtual door carries out, and decreases calculated amount; Finally by carrying out pedestrian's occlusion detection and block hierarchical estimation, it is to increase there is being the counting accuracy rate in the situation of blocking.

Description

A kind of pedestrian counting method

Technical field

The invention belongs to computer video processing technology field, it is specially a kind of pedestrian counting method, especially design a kind of pedestrian's more line method of counting being suitable for monitoring application.

Background technology

At present, people counting technology plays more and more important effect in monitoring, pedestrian monitor especially in key monitoring object.

People counting based on pedestrian detection and the people counting based on pixels statistics are two kinds of common pedestrian counting methods. But the pedestrian counting method based on pedestrian detection needs all to detect the pedestrian in scene usually, thus carry out counting (see: BoWu, R.Nevatia. " Detectionofmultiple; partiallyoccludedhumansinasingleimagebyBayesiancombinati onofedgeletpartdetectors; " ICCV2005, TenthIEEEInternationalConferenceon, 1:90 97,2005.) thus its calculated amount is bigger, real-time is relatively poor, can not directly apply to real-time monitoring system. The traditional pedestrian counting method based on pixels statistics (see: Lee, G.G., Kim, B.S., Kim, W.Y. " AutomaticEstimationofPedestrianFlow. " ACM/IEEEInternationalConferenceonDistributedSmartCameras, 291 296,2007.) pedestrian number information is obtained by pedestrian's feature is carried out statistical computation, thus avoid pedestrian detection process, decrease the complexity of computing, but it does not consider the situation of blocking, thus causing when there being pedestrian to block generation, the accuracy of counting reduces.

Summary of the invention

It is an object of the invention to overcome above-mentioned the deficiencies in the prior art part, it is proposed to a kind of new pedestrian counting method, the method is under the jurisdiction of the pedestrian counting method based on pixels statistics, avoids pedestrian detection process, decreases calculated amount; Add pedestrian's occlusion detection simultaneously and block hierarchical estimation, block the counting loss caused because of pedestrian for dynamic compensation, it is to increase the accuracy of people counting.

For achieving the above object, the technical solution used in the present invention is: monitor video carries out pedestrian's foreground pixel and the extraction of light stream feature; Then the foreground pixel extracted is normalized; Carry out pedestrian's occlusion detection simultaneously and block hierarchical estimation; Comprehensive pedestrian's characteristic sum blocks grade, Statistics Bar people's number. Confirming through experiment, the present invention has higher accuracy for the people counting under monitoring scene, and real-time meets bayonet socket monitoring device requirement.

The inventive method specifically comprises the following steps:

The first step: monitor video is carried out mixed Gaussian background modeling, obtains pedestrian's prospect image.

2nd step: extract pedestrian's light stream feature.

3rd step: pedestrian's foreground pixel feature normalization.

Concrete steps are: adopt the oval modeling of pedestrian, be mapped to an ellipse according to its area size according to Elliptic Mappings rule by each pedestrian, this ellipse area is considered as pedestrian e_iArea in current camera perspective, is designated as area (e_i), to each foreground pixel belonging to this pedestrian, its corresponding normalization coefficient w_i=1/area(e_i)��

4th step, pedestrian's occlusion detection and block hierarchical estimation.

Concrete steps are:

1. pedestrian's occlusion detection and block hierarchical estimation, its method is:

$\mathrm{\η}\left(x\right)=\left\{\begin{array}{cc}0,& h_i\≤h_{0i}\\ s_i{h}_i/A_{\mathrm{ROI}},& h_{0i}\≤h_i\≤2h_{0i}\end{array}\right.,x\∈{\mathrm{blob}}_i$

X represents pedestrian's foreground pixel, blob_iRepresenting pedestrian foreground pixel UNICOM region, �� (x) represents the shaded coefficient of foreground pixel x, and �� (x) is more big, shows that current scene blocks grade more high. s_iRepresent that pedestrian's foreground pixel is connected region blob_iProjecting length on virtual door, h_iRepresent that pedestrian's foreground pixel is connected region blob_iIn the projecting length being perpendicular on virtual door, ROI represents area-of-interest selected on image, A_ROIRepresent the image area of area-of-interest. h₀Represent pedestrian's center line average on the image plane, h_0iRepresent the h after projection correction₀��

2. pedestrian level projection correction, its method is:

Make the projection transform function that f represents between image coordinate system and world's system of coordinates, H₀Represent the average pedestrian level under world's system of coordinates, s_iRepresent that pedestrian's foreground pixel is connected region blob_iProjecting length on virtual door, p₀Represent s_iMid point, its two dimensional image coordinate isThree-dimensional world coordinate is (X₀,Y₀,Z₀), the two dimensional image coordinate of order point p isIts world's coordinate is (X_p,Y_p,Z_p)=(X₀,Y₀,Z₀+H₀), then line segmentAt the projection h being perpendicular on virtual door direction_0iIt is the pedestrian's center line average after projection correction.

Preferably, pedestrian level projection correction concrete steps are:

(1) calculation level p₀Three-dimensional coordinate under world's system of coordinates.

$\left[\begin{array}{c}{X}_0\\ Y_0\\ Z_0\end{array}\right]=f\left(\left[\begin{array}{c}{x}_0^{\mathrm{image}}\\ y_0^{\mathrm{image}}\end{array}\right]\right)$

(2) three-dimensional coordinate of calculation level p under world's system of coordinates.

$\left[\begin{array}{c}{X}_p\\ Y_p\\ Z_p\end{array}\right]=\left[\begin{array}{c}{X}_0\\ Y_0\\ Z_0+H_0\end{array}\right]$

(3) two-dimensional coordinate of calculation level p under image coordinate system.

$\left[\begin{array}{c}{x}_p^{\mathrm{image}}\\ y_p^{\mathrm{image}}\end{array}\right]=f^{-1}\left(\left[\begin{array}{c}{X}_p\\ Y_p\\ Z_p\end{array}\right]\right)$

(4) line segment is calculatedAt the projection h being perpendicular on virtual door direction_0i��

5th step: in conjunction with second and third, the normalization characteristic that obtains of four steps and block grade, obtain people's flow amount.

Concrete grammar is:

$\mathrm{flow}\left(i\right)=\underset{x\∈\mathrm{Gate}}{\mathrm{\Σ}}(1+a_0\mathrm{\η}\left(x\right))\·a\·w\left(x\right)\·\left|\stackrel{\→}{v_x}\right|\·\sin \left({\mathrm{\θ}}_{\mathrm{vx}}\right)\·\mathrm{FG}\left(x\right)$

Wherein, flow (i) represents the more line number of pixels of the i-th frame, and x is the pixel on virtual door Gate, a is fixing constant, for correcting pedestrian's Elliptic Mappings so that ellipse area, closer to true pedestrian, is taken as 0.925 usually, w (x) is then for overcoming pick up camera perspective effect, namely the normalization coefficient obtained in step 3, the prospect that FG (x) is pixel differentiates result, if prospect, it is then 1, otherwise it is 0.For the motion vector modulus value of current pixel, ��_vxIt is then the angle of the current pixel virtual door of motion vector sum, by sin (��_vx) symbol can judge pedestrian direction. a₀For fixing constant, for limiting the peak value scope blocking compensation, usually get the shaded coefficient that 0.5, �� (x) represents the foreground pixel x obtained in step 4.

Compared with prior art, the present invention has following useful effect:

1) the present invention is a kind of pedestrian counting method based on pixels statistics, avoids pedestrian detection process, decreases calculated amount; 2) add pedestrian's occlusion detection and block hierarchical estimation, block the counting loss caused because of pedestrian for dynamic compensation, it is to increase the accuracy of people counting; 3) pedestrian's feature is extracted and is only carried out in specific area-of-interest with demographics, avoids full figure is carried out computing, it is possible to reduce computational complexity further. The people counting that the present invention is particularly useful under monitoring scene.

Accompanying drawing explanation

By reading with reference to the detailed description that non-limiting example is done by the following drawings, the other features, objects and advantages of the present invention will become more obvious:

Fig. 1 is pedestrian counting method overall block flow diagram of the present invention.

Fig. 2 is pedestrian's foreground pixel schematic diagram in the embodiment of the present invention.

Fig. 3 is pedestrian's light stream feature schematic diagram in the embodiment of the present invention.

Fig. 4 is pedestrian's foreground pixel normalization method schematic diagram in the embodiment of the present invention.

Fig. 5 is pedestrian's occlusion detection and block hierarchical estimation schematic diagram in the embodiment of the present invention.

Fig. 6 is pedestrian center line average projection correction schematic diagram.

Fig. 7 is embodiment of the present invention monitoring scene schematic diagram, and camera and horizontal plane angle are less than 90 ��.

Embodiment

Below in conjunction with specific embodiment, the present invention is described in detail. The technician contributing to this area is understood the present invention by following examples further, but does not limit the present invention in any form. It should be appreciated that to those skilled in the art, without departing from the inventive concept of the premise, it is also possible to make some distortion and improvement. These all belong to protection scope of the present invention.

Embodiment

The video sequence that the present embodiment adopts is public security monitoring scene sequence.

The pedestrian counting method that the present embodiment relates to, comprises following concrete steps:

The first step: monitor video is carried out mixed Gaussian background modeling, obtains pedestrian's prospect image, as shown in Figure 2.

2nd step: extract pedestrian's light stream feature, as shown in Figure 3.

3rd step: pedestrian's foreground pixel feature normalization. Concrete steps are: adopt the oval modeling of pedestrian, be mapped to an ellipse according to its area size according to Elliptic Mappings rule by each pedestrian, this ellipse area is considered as pedestrian e_iArea in current camera perspective, is designated as area (e_i), to each foreground pixel belonging to this pedestrian, its corresponding normalization coefficient w_i=1/area(e_i), as shown in Figure 4.

4th step, pedestrian's occlusion detection and block hierarchical estimation.

Concrete steps are:

1. pedestrian's occlusion detection and block hierarchical estimation, its method is:

$\mathrm{\η}\left(x\right)=\left\{\begin{array}{cc}0,& h_i\≤h_{0i}\\ s_i{h}_i/A_{\mathrm{ROI}},& h_{0i}\≤h_i\≤2h_{0i}\end{array}\right.,x\∈{\mathrm{blob}}_i$

X represents pedestrian's foreground pixel; Blob_iRepresent pedestrian foreground pixel UNICOM region, as shown in elliptic region in Fig. 5; �� (x) represents the shaded coefficient of foreground pixel x, and �� (x) is more big, shows that current scene blocks grade more high; As shown in Figure 5, s_iRepresent that pedestrian's foreground pixel is connected region blob_iProjecting length on virtual door, h_iRepresent that pedestrian's foreground pixel is connected region blob_iIn the projecting length being perpendicular on virtual door, ROI represents area-of-interest selected on image, A_ROIRepresent the image area of area-of-interest; h₀Represent pedestrian's center line average on the image plane, h_0iRepresent the h after projection correction₀��

2. pedestrian level projection correction, its method is:

Make the projection transform function that f represents between image coordinate system and world's system of coordinates; As shown in Figure 6, H₀Represent the average pedestrian level under world's system of coordinates, s_iRepresent that pedestrian's foreground pixel is connected region blob_iProjecting length on virtual door, p₀Represent s_iMid point, its two dimensional image coordinate isThree-dimensional world coordinate is (X₀,Y₀,Z₀), the two dimensional image coordinate of order point p isIts world's coordinate is (X_p,Y_p,Z_p)=(X₀,Y₀,Z₀+H₀), then line segmentAt the projection h being perpendicular on virtual door direction_0iIt is the pedestrian's center line average after projection correction. Concrete steps are:

(1) calculation level p₀Three-dimensional coordinate under world's system of coordinates.

$\left[\begin{array}{c}{X}_0\\ Y_0\\ Z_0\end{array}\right]=f\left(\left[\begin{array}{c}{x}_0^{\mathrm{image}}\\ y_0^{\mathrm{image}}\end{array}\right]\right)$

(2) three-dimensional coordinate of calculation level p under world's system of coordinates.

$\left[\begin{array}{c}{X}_p\\ Y_p\\ Z_p\end{array}\right]=\left[\begin{array}{c}{X}_0\\ Y_0\\ Z_0+H_0\end{array}\right]$

(3) two-dimensional coordinate of calculation level p under image coordinate system.

$\left[\begin{array}{c}{x}_p^{\mathrm{image}}\\ y_p^{\mathrm{image}}\end{array}\right]=f^{-1}\left(\left[\begin{array}{c}{X}_p\\ Y_p\\ Z_p\end{array}\right]\right)$

(4) line segment is calculatedAt the projection h being perpendicular on virtual door direction_0i��

5th step: in conjunction with second and third, the normalization characteristic that obtains of four steps and block grade, calculate people's flow amount. Concrete grammar is:

$\mathrm{flow}\left(i\right)=\underset{x\∈\mathrm{Gate}}{\mathrm{\Σ}}(1+a_0\mathrm{\η}\left(x\right))\·a\·w\left(x\right)\·\left|\stackrel{\→}{v_x}\right|\·\sin \left({\mathrm{\θ}}_{\mathrm{vx}}\right)\·\mathrm{FG}\left(x\right)$

Wherein, flow (i) represents the more line number of pixels of the i-th frame, and x is the pixel on virtual door Gate, a is fixing constant, for correcting pedestrian's Elliptic Mappings so that ellipse area is closer to true pedestrian, and the present embodiment is taken as 0.925, w (x) is then for overcoming pick up camera perspective effect, namely the normalization coefficient obtained in step 3, the prospect that FG (x) is pixel differentiates result, if prospect, it is then 1, otherwise it is 0.For the motion vector modulus value of current pixel, ��_vxIt is then the angle of the current pixel virtual door of motion vector sum, by sin (��_vx) symbol can judge pedestrian direction. a₀For fixing constant, for limiting the peak value scope blocking compensation, the present embodiment gets the shaded coefficient that 0.5, �� (x) represents the foreground pixel x obtained in step 4.

The present embodiment to comprising the video sequence that 70 positive movement pedestrians and 42 are negative movement pedestrian test, test result is as shown in table 1.

Table 1. people counting result

The feature that the present invention gets over line pedestrian by statistics carries out people stream counting, thus avoids pedestrian detection process; The ROI region of statistical computation simultaneously around virtual door carries out, and decreases calculated amount; Finally by carrying out pedestrian's occlusion detection and block hierarchical estimation, it is to increase there is being the counting accuracy rate in the situation of blocking.

Above specific embodiments of the invention are described. It is understood that the present invention is not limited to above-mentioned particular implementation, those skilled in the art can make various distortion or amendment within the scope of the claims, and this does not affect the flesh and blood of the present invention.

Claims (2)

1. a pedestrian counting method, it is characterised in that, comprise the following steps:

The first step: monitor video is carried out mixed Gaussian background modeling, obtains pedestrian's prospect image;

2nd step: extract pedestrian's light stream feature;

3rd step: pedestrian's foreground pixel feature normalization;

Adopt the oval modeling of pedestrian, it is mapped to an ellipse according to its area size according to Elliptic Mappings rule by each pedestrian, this ellipse area is considered as pedestrian e_iArea in current camera perspective, is designated as area (e_i), to each foreground pixel belonging to this pedestrian, its corresponding normalization coefficient w_i=1/area (e_i);

4th step, pedestrian's occlusion detection and block hierarchical estimation;

Described pedestrian's occlusion detection and block hierarchical estimation method and be:

$\mathrm{\η}\left(x\right)=\left\{\begin{array}{cc}0& h_i\≤h_{0i}\\ s_i{h}_i/A_{ROI},& h_{0i}\≤h_i\≤2h_{0i}\end{array}\right.,x\∈{\mathrm{blob}}_i$

X represents pedestrian's foreground pixel, blob_iRepresenting pedestrian foreground pixel UNICOM region, �� (x) represents the shaded coefficient of foreground pixel x, and �� (x) is more big, shows that current scene blocks grade more high; s_iRepresent that pedestrian's foreground pixel is connected region blob_iProjecting length on virtual door, h_iRepresent that pedestrian's foreground pixel is connected region blob_iIn the projecting length being perpendicular on virtual door, h₀Represent pedestrian's center line average on the image plane, h_0iRepresent the h after projection correction₀, ROI represents area-of-interest selected on image, A_ROIRepresent the image area of area-of-interest;

In 4th step: the pedestrian center line average h in image plane₀Projection correction be: make the projection transform function that f represents between image coordinate system and world's system of coordinates, H₀Represent the average pedestrian level under world's system of coordinates, s_iRepresent that pedestrian's foreground pixel is connected region blob_iProjecting length on virtual door, p₀Represent s_iMid point, its two dimensional image coordinate isThree-dimensional world coordinate is (X₀,Y₀,Z₀), the two dimensional image coordinate of order point p isIts world's coordinate is (X_p,Y_p,Z_p)=(X₀,Y₀,Z₀+H₀), then line segmentAt the projection h being perpendicular on virtual door direction_0iIt is the pedestrian's center line average after projection correction;

5th step: in conjunction with second and third, the normalization characteristic that obtains of four steps and block grade, obtain people's flow amount; It is specially:

Wherein, flow (i) represents the more line number of pixels of the i-th frame, and x is the pixel on virtual door Gate; A is fixing constant, for correcting pedestrian's Elliptic Mappings so that ellipse area, closer to true pedestrian, is taken as 0.925; W (x) is then for overcoming pick up camera perspective effect, the normalization coefficient that namely the 3rd step is obtained; The prospect that FG (x) is pixel differentiates result, if prospect, is then 1, otherwise is 0;For the motion vector modulus value of current pixel; ��_vxIt is then the angle of the current pixel virtual door of motion vector sum, by sin (��_vx) symbol decision pedestrian direction; a₀For fixing constant, for limiting the peak value scope blocking compensation, get 0.5; �� (x) represents the shaded coefficient of foreground pixel x.

2. pedestrian counting method according to claim 1, is characterized in that: the pedestrian center line average h in image plane₀Projection correction, concrete steps are:

1. calculation level p₀Three-dimensional coordinate under world's system of coordinates;

$\left[\begin{array}{c}{X}_0\\ Y_0\\ Z_0\end{array}\right]=f\left(\left[\begin{array}{c}{x}_0^{image}\\ y_0^{image}\end{array}\right]\right)$

2. the three-dimensional coordinate of calculation level p under world's system of coordinates;

$\left[\begin{array}{c}{X}_p\\ Y_p\\ Z_p\end{array}\right]=\left[\begin{array}{c}{X}_0\\ Y_0\\ Z_0+H_0\end{array}\right]$

3. the two-dimensional coordinate of calculation level p under image coordinate system;

$\left[\begin{array}{c}{x}_p^{image}\\ y_p^{image}\end{array}\right]=f^{-1}\left(\left[\begin{array}{c}{X}_p\\ Y_p\\ Z_p\end{array}\right]\right)$

4. line segment is calculatedAt the projection h being perpendicular on virtual door direction_0i��