CN104637058A - Image information-based client flow volume identification statistic method - Google Patents

Abstract

The invention relates to a method for identifying and counting passenger flow based on image information. The method utilizes an improved self-adaptive moving target detection method. The following steps are processed: first, predict the search area of the target image based on Kalman filtering; then, match the target according to the improved cost function, and establish a target chain to determine the corresponding relationship between the target images; finally, according to the target The association and matching relationship of the target can be determined to determine whether the target enters or leaves the recognition area, or is stationary and continuously moves inside and outside the target recognition area, so as to achieve the purpose of passenger flow recognition and statistics. The invention has the characteristics of strong real-time performance, high recognition rate, comprehensive statistical data, low hardware requirements and the like.

Description

A Method of Recognition and Statistics of Passenger Flow Based on Image Information

technical field

The invention relates to the technical field of image recognition, in particular to a method for identifying and counting passenger flow based on image information.

Background technique

Passenger flow has always been an indispensable and important data for management and decision-making in public places such as shopping malls, airports, bus stations, and subway stations. With the development of my country's economy and technology, the demand for passenger flow statistics in many industries is also growing. The traditional manual statistical method is time-consuming and laborious, and the timeliness is not strong. The follow-up data processing is cumbersome and cannot provide real-time statistical data. Therefore, it is of great significance to develop a real-time passenger flow intelligent statistical system.

The traditional infrared photoelectric detection passenger flow statistics system is simple in principle and widely used, but has obvious defects: the active type is greatly affected by natural conditions such as temperature and light, and has poor reliability; the passive type has strong reliability, but is expensive. And limited by the venue. Passenger flow recognition statistics based on image information has the advantages of low cost, simple installation and maintenance, etc., and can also be connected with the original monitoring system without purchasing additional equipment.

The system design mainly solves the problem of multi-target recognition and tracking counting in dynamic images. At present, a variety of passenger flow statistics and counting methods based on dynamic images have been proposed. The representative methods are as follows:

(1) Establish a mathematical statistical model of the real trajectory of pedestrians. For example, the foreground module is obtained by combining the background modeling method and the prior assumption algorithm, and then the trajectory model is filtered to exclude non-pedestrian trajectories. This algorithm saves time compared to direct detection of pedestrians, but it has higher requirements on the trajectory mathematical model and depends on the accuracy of foreground block recognition.

(2) Based on genetic algorithm. For example, "Research on Passenger Flow Analysis System Based on Genetic Algorithm" improves the accuracy of passenger flow by introducing genetic algorithm into the statistical system of passenger flow. Recognition difficulty increases.

(3) Mathematical morphology algorithm. For example, "Customer Flow Statistics Based on Improved Feature Tracking" proposes a mathematical morphology-based passenger flow statistics method. First, analyze the connected domain of the target after mathematical morphology processing, and use the forward search principle to determine the new location of the target. , count by delineating the count line. This method is simple to implement, but it is easy to be confused with other non-human moving objects, which affects the accuracy of statistics.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method for identifying and counting passenger flow based on image information in view of the defects in the prior art.

The technical solution adopted by the present invention to solve the technical problem is: a method for identifying and counting passenger flow based on image information, comprising the following steps:

1) Determine the target detection area; set four parameters of the ROI of the target image detection area: ROI.X represents the x-axis coordinate of the upper left corner point of the detection area ROI, ROI.Y represents the y-axis coordinate of the upper left corner point of the detection area ROI, ROI. Width represents the width of the detection area ROI, and ROI.Height represents the height of the detection area ROI;

2) Determine that the target detection area does not have the background image when the target to be detected exists; adopt the multi-image averaging method, calculate the average value for the superimposition of 10 to 20 frames of original images, and obtain the background image of the video image when the target does not exist, denoted as I ₀ ;

3) Process each frame of the video image collected, and store the image frames of the moving target into the array in order;

Process the i-th frame image of the collected video image into blocks, record it as the current frame image, and then select the target image detection region ROI;

After binarization and distance transformation of the current frame image, the enhanced image of the current frame image is obtained through mean filtering or sub-sampling; , record the difference of the ROI of the selected target image detection region as NZ;

Compare the motion detection difference NZ with the set alarm minimum element quantity limit value N0: if NZ<N0, then there is no motion, automatically collect the i+1 frame original image, and return to step 3); if NZ≥N0 , then save the current frame image, denoted as I1;

Differentiate the current frame image and the background image I ₀ , and record the difference of the selected target image detection region ROI as N _B ;

Compare the target detection difference N _B with the set alarm minimum element number limit N ₀ ': if N _B <N ₀ ', there is no target to be detected; if N _B ≥ N ₀ ', there is movement in the alarm target, and then store the current frame image with the moving target in the array MovingImages[i][j];

4) according to whether the image frames are continuous, the image frames in the array are divided into blocks;

(1) Save the minimum and maximum values of the abscissa of the target image detection region ROI into the corresponding variables xin and xout respectively, then read the original image sequence of the first frame in the array MovingImages[i][j], record The target image of the current frame counted for multi-target tracking statistics;

(2) Define the counting variables PersonIn and PersonOut, respectively record the number of objects entering and leaving, and initialize it to 0, and stipulate that along the X direction is in, and against the X direction is out;

(3) Carry out adaptive background segmentation based on the difference method on the target image of the current frame, binarize the obtained moving region image, and use a 2×2 structure operator (commensurate with the size of the head) to binarize the target image The image is morphologically corroded to remove false targets (ie, non-head images), and then the size of the actual target is restored through morphological expansion to obtain a binary image that only contains the connected domain of the moving target;

(4) Traverse the binary image that only contains the connected domain of the moving target, obtain the eigenvalues of each target connected domain Area[n][i], and store them in the corresponding linked list. The parameters include: image index Num, target index Index, The abscissa X and ordinate Y of the target centroid correspond to the gray average Gray of the target area of the original image, the Length and Width of the rectangular enclosing window of the target area, and the area size Space of the target area;

(5) According to the obtained eigenvalues of the target connected domain Area[n][i], use Kalman filter to predict the centroid motion position p(xi, The search area of yi);

(6) gather next frame image, carry out the operation of step (2) to it;

(7) with step (3);

(8) According to the improved cost function, match the target connected domains of the next frame image in the Kalman prediction search area of the current frame image, and calculate the target connected domain Area[n][i] of the current frame and the corresponding search area of the next frame image The cost function values of all target connected domains in , and find the minimum value (assuming the minimum cost function value with the next frame target Area[n+1][j]);

(9) Calculate the centroid distance d of the target connected domain Area[n][i] and Area[n+1][j], and compare it with the centroid distance limit value d0. If d≤d0, the target connected domain Area[n+1][j] is the follow-up of the target connected domain Area[n][i], and the eigenvalue of the target connected domain Area[n+1][j] is used to replace the target The value of the connected domain Area[n][i], and mark the target connected domain Area[n+1][j] to establish the target chain;

(10) If d>d0, the target connected domain Area[n][i] has no follow-up in the next frame, and the target connected domain Area[n][i] may leave the image observation window or be temporarily static, and the target connected domain needs to be judged The position of the centroid coordinate X of Area[n][i]:

5) counting the image frames after the block;

details as follows:

5.1) Save the minimum and maximum values of the abscissa of the target image detection region ROI into the corresponding variables xin and xout respectively, and then read the first frame of the original image sequence in the array MovingImages[i][j], which is recorded as The current frame target image of multi-target tracking statistics;

5.2) Define the counting variables PersonIn and PersonOut, respectively record the number of objects entering and leaving, and initialize it to 0, and stipulate that along the positive direction of the X axis, it is in, and against the positive direction of the X axis, it is out;

5.3) The position of the centroid coordinate X of the target connected domain Area[n][i]:

When X≤xin, if the abscissa of the first centroid X1≤xin of the movement trajectory of the target connected domain Area[n][i], it means that the target has been wandering in the reserved area, and any counter does not act; if X1 >xin, it means that the target leaves the tracking window, the direction is out, the out counter PersonOut is increased by 1, and the target chain of the target connected domain Area[n][i] is cleared;

When X>xout, if the abscissa coordinate X1≥xout of the first centroid coordinate of the movement track of the target connected domain Area[n][i], it means that the target has been wandering in the reserved area, and any counter does not act; if X1 <xout, it means that the target enters the tracking window, the direction is forward, the forward counter PersonIn is increased by 1, and the target chain of the target connected domain Area[n][i] is cleared;

When xin<X<xout, it means that the target connected domain Area[n][i] is moving in the detection area, and its final movement direction is unknown, so keep its eigenvalues and wait for the tracking of the target in the next frame; if the next If there is a matching target in the frame, establish the target chain according to the above steps, otherwise it will be discarded as interference;

5.4) After matching all target connected domains in the tracking window, verify whether all targets in the current frame are tracked; if there is an untracked target, judge whether the abscissa of its centroid satisfies any of X≤xin and X>xout One, if it is satisfied, it means that there is a new target, establish a new target chain for it, obtain and record the characteristic value, and turn to (3) in step 4; if it is not satisfied, it will be judged as interference and discarded;

5.5) After all the target recognition counts in the current frame image are finished, calculate the difference between the current advance counter PersonIn and the difference value PersonOut of the counter, which is recorded as the current passenger flow CustomCounting, and the target recognition count of the current frame image ends;

5.6) Identify and automatically count the target connected domains of each frame image in the image array MovingImages[i][j] in turn, until the image sequences in the image array MovingImages[i][j] are all processed, then the entire target recognition count Finish.

According to the above scheme, the described adopting target chain cost function to track and match multiple targets, the steps are:

1) Obtain three eigenvalues of multiple targets in the k-th frame image: the coordinates of the center of mass of the target, the average gray level and the area of the target area, where the eigenvalues of the target i are respectively recorded as: Point[k][i], Gray[ k][i] and Space[k][i];

2) Use the Kalman filter to predict the search area of the target i in the k+1 frame image in the k-th frame image, traverse the search area of the target i in the k+1 frame image, find n target objects, and obtain their characteristics value, where the eigenvalues of target j are respectively recorded as: Point[k+1][j], Gray[k+1][j] and Space[k+1][j];

3) Calculate the degree of change between the n target objects in the search area of the target i in the t+1 frame image and the three eigenvalues of the target i in the t frame image, and obtain the maximum values respectively as MaxPoint, MaxGray and MaxSpace , the specific calculation formula is as follows:

MaxPoint[i]=Max(sqrt((Point[k][i].x-Point[k+1][m].x)*(Point[k][i].x-Point[k+1] [m].x)+(Point[k][i].y-Point[k+1][m].y)*(Point[k][i].y-Point[k+1][m ].y)))

MaxGray[i]=Max(sqrt((Gray[k][i]-Gray[k+1][m])*(Gray[k][i]-Gray[k+1][m])))

MaxSpace[i]=Max(sqrt((Space[k][i]-Space[k+1][m])*(Space[k][i]-Space[k+1][m])))

Among them, 1≤m≤n, which means any one of n targets in the search area of target i; where sqrt is the square root operation;

4) Calculate the centroid distance D[i][j], the degree of grayscale change H[i][j] and the degree of area change S[ i][j], the corresponding calculation formula is as follows:

D[i][j]=sqrt((Point[k][i].x-Point[k+1][j].x)*(Point[k][i].x-Point[k+1 ][j].x)+(Point[k][i].y-Point[k+1][j].y)*(Point[k][i].y-Point[k+1][ j].y))/MaxPoint[i];

H[i][j]=sqrt((Gray[k][i]-Gray[k+1][j])*(Gray[k][i]-Gray[k+1][j])) /MaxGray[i];

S[i][j]=sqrt((Space[k][i]-Space[k+1][j])*(Space[k][i]-Space[k+1][j])) /MaxSpace[i]

5) The cost function V[i][j] can be obtained by the three eigenvalue changes obtained in step 4.5.4):

V[i][j]=αD[i][j]+βH[i][j]+γS[i][j];

Among them, the three parameter coefficients of α, β, and γ represent the influencing factors of the three eigenvalues, which can be adjusted according to the actual situation.

According to the above scheme, the specific steps of using the Kalman filter to predict the search area are:

1) Assuming that the center of mass of each target in the image moves at a uniform speed, and the area of the surrounding window is basically unchanged, initialize the moving speed V _x of the target in the x direction, the moving speed V _y in the y direction, and the changing speed Vt of the length and width of the surrounding window , the image sampling step size T (that is, the time interval between each frame of image sequences);

2) Based on the target i in the k-1th frame image, obtain the optimal estimation result of the feature value corresponding to the target i in the k-th frame image: centroid position X coordinate Point[k][i]. x and Y coordinate Point[ k][i].y, the length Win[k][i].length and width Win[k][i].width of the surrounding window, and the deviation estimate P[k][i] of the two feature quantities;

3) Based on the results obtained in step 2), calculate the predicted value of the feature value and the deviation estimate of the target i in the k+1th frame image. The specific calculation formula is as follows:

Point[k+1][i].x=Point[k][i].x+V _x *T;

Point[k+1][i].y=Point[k][i].y+V _y *T;

Win[k+1][i].length=Win[k][i].length+Vt*T;

Win[k+1][i].width=Win[k][i].width+Vt*T;

P[k+1][i]=P[k][i]+Q;

4) Calculate the corresponding Kalman gain according to the eigenvalue prediction value and bias estimation of the target i in the k+1th frame image:

Kg[k+1][i]=P[k][i]/(P[k][i]+R)

Calculate the optimal estimated value of each eigenvalue and bias estimate of the target i in the k+1th frame image, and update their predicted values:

Point[k+1][i].x＝Point[k+1][i].x+Kg[k+1][i]*(Z[k+1][i].x-Point[k +1][i].x); Point[k+1][i].y=Point[k+1][i].y+Kg[k+1][i]*(Z[k+1 ][i].y-Point[k+1][i].y);

Win[k+1][i].length＝Win[k+1][i].length+Kg[k+1][i]*(Z[k+1][i].length-Win[k +1][i].length);

Win[k+1][i].width＝Win[k+1][i].width+Kg[k+1][i]*(Z[k+1][i].width-Win[k +1][i].width);

P[k+1][i]=(1-Kg[K+1][i])*P[k][i];

5) From this, it can be obtained that the predicted search area (X, Y) of the target i in the k+1th frame in the image of the kth frame is:

Point[k+1][i].x-1.5*Win[k+1][i].length/2≤X≤Point[k+1][i].x+1.5*Win[k+1] [i].length/2;

Point[k+1][i].y-1.5*Win[k+1][i].width/2≤Y≤Point[k+1][i].y+1.5*Win[k+1] [i].width/2.

According to the above-mentioned scheme, the realization step of the algorithm of the association matching of target and data association in the target chain in (8) of described step 4 is:

In the association matching of the target chain, if the cost function values of the jth target and the mth target in the prediction search area of the target i in the kth frame image and the target i in the k+1th frame image are equal, and at the same time is the smallest, then the tracking and matching of the target will conflict at this time, and the following improved algorithm is used to complete the correct tracking and matching of the target;

Calculate the cost function of the target i in the kth frame image and the target j and m in the k+1th frame image. If the cost function value is equal and the smallest, then record the gray feature values of the above three targets as Gray[k ][i], Gray[k+1][j], Gray[k+1][m];

Respectively calculate the average gray scale change of target i in the kth frame image and target j, m in the k+1th frame image, the calculation formula is:

Value1=sqrt((Gray[k][i]-Gray[k+1][j])*(Gray[k][i]-Gray[k+1][j]));

Value2=sqrt((Gray[k][i]-Gray[k+1][m])*(Gray[k][i]-Gray[k+1][m]));

If Value1<Value2, the target j in the k+1 frame image is the follow-up of the target i in the k frame image, and the feature value of the target j in the k+1 frame image is used to replace the target i in the k frame image Eigenvalue, update the target chain, otherwise, the target m in the k+1th frame image is the follow-up of the target i in the kth frame image;

If there is no match with the target i in the image of the kth frame, it can be matched with other targets in the image of the kth frame, or it can be reserved and wait for the matching target to be found in the next frame.

The beneficial effects produced by the present invention are: the passenger flow identification and statistics method based on image information provided by the present invention provides a solution for this kind of multi-target identification and tracking and counting problems for reference. Good stickiness, high intelligence and strong stability.

Description of drawings

The present invention will be further described below in conjunction with accompanying drawing and embodiment, in the accompanying drawing:

Figure 1 is a schematic diagram of hardware installation at the image acquisition site;

Fig. 2 is the field background image and detection area ROI setting of a certain experiment of the present invention;

Fig. 3 is the image of the target entering the detection area under the background of Fig. 1;

Fig. 4 is the difference image of moving target existence detection stage;

Fig. 5 is the result image of human head target recognition and feature extraction;

Figure 6 shows the results of human head target recognition displayed in the original image;

Fig. 7 is the target feature scale of a group of dynamic images;

Fig. 8 is the center-of-mass trajectory of this type of multi-target in Fig. 6;

Fig. 9 is the tracking window established by this type of multi-target in Fig. 6;

Fig. 10 is a flow chart of the method of the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention more clear, the present invention will be further described in detail below in conjunction with the examples. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

As shown in Figure 10, a method for identifying and counting passenger flow based on image information.

A camera 2 is installed on the top of the entrance 1 of a shopping mall, and the PC host communicates with the camera 2 through the image acquisition card 3, and displays the collected image information on the monitor 5 after processing, and uses the image recognition on the top of the person and the flow of the statistics personnel, The schematic diagram of on-site hardware installation is shown in Figure 1.

The first step, detection and recognition of moving targets

The image resolution used in the experiment is 600×560. In order to reduce the amount of image processing, the image is divided into blocks, and the size of the image is 150×140. The camera captures images at a speed of 4 frames/s. Specific steps are as follows:

1) Set the four parameters of the target image detection ROI area: ROI.X represents the x-axis coordinate of the upper left corner of the ROI area, ROI.Y represents the y-axis coordinate of the upper left corner of the ROI area, ROI.Width represents the width of the ROI area, and ROI.Height Represents the height of the ROI area. In this example, set: ROI.X=40, ROI.Y=50, ROI.Width=50, ROI.Height=60.

2) Using the multi-image averaging method, the M frame images are superimposed to calculate the average value, and the background image of a certain video image when there is no target is obtained, as shown in Figure 2.

3) Process the captured i-th frame of the video image into blocks, record it as the current frame parameter determination image, and select the target image to detect the ROI area, as shown in Figure 3.

4) After the current frame image is binarized, and then through distance transformation and mean filtering or sub-sampling, the enhanced image of the current frame image is obtained; Difference, the difference of the selected ROI area is recorded as N _Z .

5) Compare N _Z with the set limit value N ₀ of the minimum number of elements for alarm: if N _Z <N ₀ , then there is no motion, automatically collect i+1 frame images, return to step 2) and start again; if N _Z ≥ N ₀ , then save the i-th frame image, denoted as I ₁ .

6) Perform adaptive background segmentation based on the difference method on the i-th frame image I ₁ and the background image I ₀ , then binarize the obtained image, and perform morphological erosion, and record the difference of the selected ROI area as N _B , the difference image is shown in Figure 4.

7) Compare N _B with the set limit value N′ ₀ of the minimum number of elements for alarm: if N _B <N′ ₀ , then there is no target to be detected, automatically collect i+1 frame images, and return to step 2) to start again Start; if N _B ≥ N′ ₀ , alarm that there is a moving target, and store the image frame in the array MovingImages[i][j] for subsequent image recognition and processing.

The second step, multi-target tracking and statistical counting

According to the read video image, set the ROI detection area, save the minimum and maximum values of its abscissa to the variables xin and xout, and select consecutive image frames with moving targets, and store them in the array MovingImages[i] in turn In [j], read the first frame target image of the image sequence in the array, record it as the current frame image of multi-target tracking statistics, and then perform online recognition of passenger flow according to the video image above the head of the person according to the following steps Counting, store the counting results in the counting variables PersonIn and PersonOut, respectively record the number of targets entering and leaving, and the initial value of the counting variable is set to 0. In this example, set xin=40 and xout=100.

1) Start two entry and exit counters, and stipulate that the positive direction of x is the entry, and the opposite direction of x is the exit.

2) After binarizing the target image in the current frame, select a suitable structure operator to perform morphological erosion on it to remove false targets (ie, non-head images), and obtain a binary image that only contains connected regions of moving targets .

3) Traverse the binary image, obtain the eigenvalues of each target connected domain Area[n][i], and store them in the corresponding linked list. The parameters include: image and target index Num and Index, and the abscissa X of the target centroid , the ordinate Y corresponds to the gray average value Gray of the target area of the original image, the length Length and width Width of the rectangular enclosing window of the target area, and the area size Space of the target area, as shown in Figure 4, the centroid coordinates of the two targets are respectively : (83.8530, 90.2874), (118.8503, 33.7245); the average gray levels are: 85.9029, 77.204; the target enclosing windows are: 30*29, 34*27; the target area sizes are: 762, 755.

4) According to the obtained feature value of the target Area[n][i], predict the centroid motion position p(xi, yi) and area of the target connected domain Area[n+1][i] in the next frame image based on Kalman filter.

5) Collect the next frame of image, and perform the operation of step 1) on it.

6) Same as step 2) of the second step.

7) According to the improved cost function, the Kalman prediction search area of the current frame image is matched with each target connected domain of the next frame image, and all targets in the search area corresponding to the current frame target Area[n][i] and the next frame image are calculated The cost function value of , and find the minimum value (assuming that the value of the cost function is the smallest with the target Area[n+1][j] of the next frame).

8) Calculate the centroid distance d of the target Area[n][i] and Area[n+1][j], and compare it with the distance limit value d0. If d≤d0, the target Area[n+1][j] is the follow-up of the target Area[n][i], and the target Area[n][i] is replaced by the eigenvalue of the target Area[n+1][j] ], and mark the target Area[n+1][j] to establish a target chain.

9) If d>d0, the target Area[n][i] has no follow-up target in the next frame, and the target may leave the image observation window or be temporarily stationary. It is necessary to judge the coordinates of the center of mass of the target Area[n][i]:

When X≤xin, if the abscissa of the first center of mass of the target Area[n][i] trajectory X1≤xin, it means that the target has been wandering in the reserved area, and any counter does not count; if X1>xin , it means that the target leaves the tracking window, the direction is out, the out counter PersonOut is incremented by 1, and the target chain Area[n][i] is cleared.

When X>xout, if the abscissa X1≥xout of the first center of mass coordinate of the moving target track, any counter will not count; if X1<xout, it means that the target enters the tracking window, the direction is forward, and the forward counter PersonIn is incremented by 1, And clear the target chain Area[n][i];

When xin<X<xout, it means that the target Area[n][i] is moving in the detection area, and the direction of its final movement is unknown, so keep its characteristic value and wait for the tracking of the target in the next frame. If there is a matching target in the next frame, the target chain will be established according to the above steps, otherwise it will be discarded as interference.

10) After all the tracked targets are matched, verify whether all targets in the current frame are tracked. If there is an untracked target, judge whether the abscissa of its centroid satisfies: X≤xin or X>xout, if it is satisfied, a new target appears, establish a new target chain for it, and set the characteristic value; if it is not satisfied, it may be Interference, discard, as shown in Figure 5. In this example, the two targets match the feature values of the target in the next frame: the centroid coordinates are: (80.7355, 91.3539), (116.1797, 34.5599); the average gray levels are: 86.9632, 77.7487; the target enclosing windows are: 31 *29, 35*28; the size of the target area is: 760, 768 respectively.

11) After all target recognition counts in each frame of image are completed, the difference between the current forward counter PersonIn and the outgoing counter PersonOut can be calculated and assigned to the variable CustomCounting, which is the statistical result of the current passenger flow. In this example, PersonIn=2, PersonOut=0, CustomCounting=2-0=1 of the current frame.

The recognition and matching results are displayed in the original image, as shown in Figure 6, and the target recognition count of the current frame image is over.

Follow steps 1) to 10) to identify and automatically count the targets in each frame in turn until all the images in the image array MovingImages[i][j] are processed, and the values of the two counters PersonIn and PersonOut at each moment and their The difference CustomCounting is stored in the database, and the passenger flow can be statistically analyzed. In this example, the final passenger flow statistic value is CustomCounting=572. In a piece of video used, the feature quantities of two targets in each frame image of the detection area during the recognition and counting process are shown in Figure 7, which is reproduced according to the data obtained from the recognition The motion trajectories of the two target centroids and target enclosing windows are shown in Fig. 8 and Fig. 9 respectively.

It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should belong to the protection scope of the appended claims of the present invention.

Claims (9)

1., based on a volume of the flow of passengers identification statistical method for image information, it is characterized in that, comprise the following steps:

1) object detection area is determined; Four parameter: ROI.X of Offered target image detection region ROI represent the x-axis coordinate of surveyed area ROI upper left angle point, ROI.Y represents the y-axis coordinate of surveyed area ROI upper left angle point, ROI.Width represents the width of surveyed area ROI, and ROI.Height represents the height of surveyed area ROI;

2) background image when object detection area does not have a target to be detected is determined;

3) every two field picture of the video image gathered is processed, will the picture frame of moving target be there is in order stored in array;

4) judge that whether the picture frame in array is continuous, according to judged result, piecemeal is carried out to the picture frame in array;

5) statistical counting is carried out to the picture frame after piecemeal, determine the volume of the flow of passengers passed in and out.

2. volume of the flow of passengers identification statistical method according to claim 1, it is characterized in that, described step 2) in determine that the method for background image is: adopt the multiple image method of average, to multiframe original image superposition averaged, obtain the background image of video image when target does not exist.

3. volume of the flow of passengers identification statistical method according to claim 1, it is characterized in that, described step 3) in enhancing image procossing is treated to the every two field picture of video image gathered, be specially: after binaryzation and range conversion are carried out to two field picture, again by mean filter or subsampling, obtain the enhancing image of current frame image.

4. volume of the flow of passengers identification statistical method according to claim 3, is characterized in that, described step 3) in every two field picture of video image of gathering whether to there is the decision method of moving target as follows:

3.1) two field picture of the video image collected is carried out piecemeal process, the i-th two field picture is designated as current frame image, then selected target image detection region ROI;

3.2) the i-th two field picture and the i-th-1 two field picture processed are carried out difference, the difference of selected target image surveyed area ROI is designated as N _z;

3.3) motion is detected difference N _zwith the warning least member quantity limit value N of setting ₀make comparisons: if N _z< N ₀, then do not move; If N _z>=N ₀, then have motion to exist, retain current frame image;

3.4) to current frame image and background image I ₀carry out difference, the difference of selected target image surveyed area ROI is designated as N _b;

3.5) by target detection difference N _bwith the warning least member quantity limit value N of setting ₀' make comparisons: if N _b< N ₀', then there is no target to be detected; If N _b>=N ₀', then report to the police and there is the target of motion, then will there is the current frame image of moving target stored in array MovingImages [i] [j];

3.6) travel through all two field pictures, preserve all two field pictures having moving target to exist in order, stored in array MovingImages [i] [j].

5. volume of the flow of passengers identification statistical method according to claim 3, is characterized in that, described step 4) in judge whether the picture frame in array adopts following methods continuously:

4.1) from array, two field picture is got successively, carry out splitting based on the adaptive background of method of difference to present frame target image, the moving region image of acquisition is carried out binaryzation, with 2 × 2 construction operators, morphological erosion is carried out to the bianry image of target image, to remove pseudo-target (i.e. non-head image), again by the size of morphological dilations reduction realistic objective, obtain the bianry image only containing moving target connected domain;

4.2) bianry image of traversal only containing moving target connected domain, obtains the eigenwert of each target connected domain Area [n] [i], stored in corresponding chained list; Wherein, eigenwert parameter comprises: image index Num, target index Index, the horizontal ordinate X of target centroid and ordinate Y, the average gray Gray of corresponding original image target area, the length Length of target area rectangles encompass window and width W idth, the size Space of target area;

4.3) according to the eigenwert of target connected domain Area [n] [i] that obtain, the region of search of center of mass motion position p (xi, yi) of target connected domain Area [n+1] [i] in next frame image is predicted by Kalman filter;

4.4) eigenwert of target connected domain Area [n+1] [i] of next frame image is calculated

4.5) each target connected domain of plain Region Matching next frame image is searched in the Kalman prediction of current frame image, calculate the cost function value of all target connected domains in present frame target connected domain Area [n] [i] region of search corresponding to next frame image, and find out wherein minimum value;

4.6) calculate the centroid distance d of the arbitrary target area Area [n+1] [j] of target connected domain Area [n] [i] and the corresponding region of search of next frame image, and make comparisons with centroid distance limit value d0:

If d≤d0, what then target connected domain Area [n+1] [j] was target connected domain Area [n] [i] is follow-up, the value of target connected domain Area [n] [i] is substituted by the eigenwert of target connected domain Area [n+1] [j], and target connected domain Area [n+1] [j] is made marks, set up object chain;

If d > is d0, then target connected domain Area [n] [i] is not follow-up in the next frame;

4.7) according to judged result, piecemeal is carried out to the picture frame in array.

6. volume of the flow of passengers identification statistical method according to claim 5, is characterized in that, described step 5) in, carry out statistical counting to the picture frame after piecemeal, concrete grammar is as follows:

5.1) minimum value of the horizontal ordinate of target image surveyed area ROI and maximal value are saved in respectively in corresponding variable xin and xout, then read the first frame original sequence in array MovingImages [i] [j], be designated as the present frame target image of multiple target tracking statistical counting;

5.2) define counting variable PersonIn and PersonOut, record the destination number entering and leave respectively, Initialize installation is 0, and regulation along X-axis positive dirction for entering, inverse X-axis positive dirction is for going out;

5.3) to the position of the center-of-mass coordinate X of target connected domain Area [n] [i]:

As X≤xin, if first barycenter horizontal ordinate X1≤xin of the movement locus of target connected domain Area [n] [i], then illustrate that this target is hovered always in trough, any counter is failure to actuate; If X1>xin, then illustrate that target leaves tracking window, direction, for going out, goes out counter PersonOut and adds 1, and removes the object chain of target connected domain Area [n] [i]; Described first barycenter horizontal ordinate is the barycenter horizontal ordinate that target starts to enter tracking window most, namely at this target first barycenter horizontal ordinate that tracking window captures;

As X>xout, if the movement locus of target connected domain Area [n] [i] first center-of-mass coordinate horizontal ordinate X1 >=xout, then illustrate that this target is hovered always in trough, any counter is failure to actuate; If X1<xout, then illustrate that target enters tracking window, direction, for entering, is entered counter PersonIn and is added 1, and removes the object chain of target connected domain Area [n] [i];

Work as xin<X<xout, then illustrate that target connected domain Area [n] [i] is moved in detection zone, its last direction of motion is failed to understand, retains its eigenwert, waits for the tracking of this target in next frame; If there is coupling target in next frame, then by above-mentioned steps 4.3) ~ 4.6) set up object chain, otherwise as interference, abandon;

5.4), after all target connected domains in tracking window are mated, in checking present frame, whether all targets are tracked; If there is not tracked target, then judge any one whether its barycenter horizontal ordinate meet in X≤xin and X>xout, satisfied then indicate that fresh target occurs, for it sets up new object chain, obtain and recording feature value, proceed to step 4.3); Do not meet and be then judged as interference, abandon;

5.5), after in current frame image, all target recognizing and countings terminate, calculate the difference of the difference PersonOut working as advance counter PersonIn and go out counter, be designated as current volume of the flow of passengers CustomCounting, the target recognizing and counting of current frame image terminates;

5.6) successively the target connected domain of two field picture every in image array MovingImages [i] [j] is identified and Auto-counting, until the image sequence in image array MovingImages [i] [j] is all disposed, then whole target recognizing and counting terminates.

7. volume of the flow of passengers identification statistical method according to claim 5, is characterized in that, described step 4.5) in adopt improve object chain cost function tracking and matching is carried out to multiple goal, step is:

4.5.1) obtain three eigenwerts of multiple target in kth two field picture: the area of target centroid coordinate, average gray and target area, wherein the eigenwert of target i is designated as respectively: Point [k] [i], Gray [k] [i] and Space [k] [i];

4.5.2) Kalman filter is utilized to predict the region of search of target i in kth+1 two field picture in kth two field picture, the traversal region of search of target i in kth+1 two field picture, find n destination object, obtain their eigenwert, wherein the eigenwert of target j is designated as respectively: Point [k+1] [j], Gray [k+1] [j] and Space [k+1] [j];

4.5.3) intensity of variation of 3 eigenwerts of target i in n destination object and t two field picture in the region of search of t+1 two field picture target i is asked for respectively, ask for maximal value and be designated as MaxPoint, MaxGray and MaxSpace respectively, specific formula for calculation is as follows:

MaxPoint[i]＝Max(sqrt((Point[k][i].x-Point[k+1][m].x)*(Point[k][i].x-Point[k+1][m].x)+(Point[k][i].y-Point[k+1][m].y)*(Point[k][i].y-Point[k+1][m].y)))

MaxGray[i]＝Max(sqrt((Gray[k][i]-Gray[k+1][m])*(Gray[k][i]-Gray[k+1][m])))

MaxSpace[i]＝Max(sqrt((Space[k][i]-Space[k+1][m])*(Space[k][i]-Space[k+1][m])))

Wherein, 1≤m≤n, to represent in the region of search of target i any one in n target; Wherein sqrt is sqrt computing;

4.5.4) ask for centroid distance size D [i] [j], grey scale change degree H [i] [j] and area change degree S [i] [j] of target j in target i and t+1 two field picture in t two field picture, corresponding computing formula is as follows:

D[i][j]＝sqrt((Point[k][i].x-Point[k+1][j].x)*(Point[k][i].x-Point[k+1][j].x)+(Point[k][i].y-Point[k+1][j].y)*(Point[k][i].y-Point[k+1][j].y))/MaxPoint[i]；

H[i][j]＝sqrt((Gray[k][i]-Gray[k+1][j])*(Gray[k][i]-Gray[k+1][j]))/MaxGray[i]；

S[i][j]＝sqrt((Space[k][i]-Space[k+1][j])*(Space[k][i]-Space[k+1][j]))/MaxSpace[i]

4.5.5) cost function V [i] [j] can by step 4.5.4) required three eigenwert variable quantities try to achieve:

V[i][j]＝αD[i][j]+βH[i][j]+γS[i][j]；

Wherein α, β, γ tri-parameter coefficients represent the factor of influence of three eigenwerts, can in the light of actual conditions adjust.

8. volume of the flow of passengers identification statistical method according to claim 5, is characterized in that, described step 4.3) adopt the concrete steps in Kalman filter forecasting search region to be:

4.3.1) suppose that the motion of each target centroid in image is uniform motion, surround window area substantially constant, initialized target is at the movement velocity V in x direction _x, movement velocity V in y-direction _y, surround the pace of change Vt of window length and width, image sampling step-length T (time interval namely between every frame image sequence);

4.3.2) based on the target i in kth-1 two field picture, the maximum likelihood estimation result of the eigenwert of corresponding target i in kth two field picture is obtained: centroid position X-coordinate Point [k] [i] .x and Y-coordinate Point [k] [i] .y, encirclement length Win [k] [i] .length and width W in [k] [i] .width of window, estimation of deviation P [k] [i] of two characteristic quantities;

4.3.3) based on 4.3.2) result that step obtains, calculate eigenwert predicted value and the estimation of deviation of target i in kth+1 two field picture, concrete computing formula is as follows:

Point[k+1][i].x＝Point[k][i].x+V _x*T；

Point[k+1][i].y＝Point[k][i].y+V _y*T；

Win[k+1][i].length＝Win[k][i].length+Vt*T；

Win[k+1][i].width＝Win[k][i].width+Vt*T；

P[k+1][i]＝P[k][i]+Q；

4.3.4) according to eigenwert predicted value and the estimation of deviation of target i in kth+1 two field picture, corresponding Kalman gain is calculated:

Kg[k+1][i]＝P[k][i]/(P[k][i]+R)

Calculate each eigenwert of target i and the optimal estimation value of estimation of deviation in kth+1 two field picture, and upgrade their predicted value:

Point[k+1][i].x＝Point[k+1][i].x+Kg[k+1][i]*(Z[k+1][i].x-Point[k+1][i].x)；

Point[k+1][i].y＝Point[k+1][i].y+Kg[k+1][i]*(Z[k+1][i].y-Point[k+1][i].y)；

Win[k+1][i].length＝Win[k+1][i].length+Kg[k+1][i]*(Z[k+1][i].length-Win[k+1]

[i].length)；

Win[k+1][i].width＝Win[k+1][i].width+Kg[k+1][i]*(Z[k+1][i].width-Win[k+1]

[i].width)；

P[k+1][i]＝(1-Kg[K+1][i])*P[k][i]；

4.3.5) can obtain the forecasting search region (X, Y) of target i in kth+1 frame in kth two field picture is thus:

Point[k+1][i].x-1.5*Win[k+1][i].length/2≤X≤Point[k+1][i].x+1.5*Win[k+1][i].length/2；

Point[k+1][i].y-1.5*Win[k+1][i].width/2≤Y≤Point[k+1][i].y+1.5*Win[k+1][i].width/2。

9. volume of the flow of passengers identification statistical method according to claim 5, is characterized in that, described step 4.5) in object chain the performing step of the association coupling of target and the algorithm of data correlation be:

In the association coupling of target, if occur, in kth two field picture, target i is equal with the cost function value of a jth target in the forecasting search region of target i in kth+1 two field picture and m target, and be minimum simultaneously, then now the tracking and matching of target can clash, then adopt following innovatory algorithm, complete the correct tracking and matching of target;

Calculate the cost function of target j and m in target i and kth+1 two field picture in kth two field picture, if cost function value is equal and be minimum, then the gray feature value of above-mentioned 3 targets is designated as Gray [k] [i], Gray [k+1] [j], Gray [k+1] [m] respectively;

Ask for the average intensity change size of target j, m in target i and kth+1 two field picture in kth two field picture respectively, computing formula is:

Value1＝sqrt((Gray[k][i]-Gray[k+1][j])*(Gray[k][i]-Gray[k+1][j]))；

Value2＝sqrt((Gray[k][i]-Gray[k+1][m])*(Gray[k][i]-Gray[k+1][m]))；

If Value1<Value2, target j then in kth+1 two field picture is the follow-up of target i in kth two field picture, the eigenwert of target i in kth two field picture is replaced by the eigenwert of the target j in kth+1 two field picture, upgrade object chain, otherwise the target m then in kth+1 two field picture is the follow-up of target i in kth two field picture;

With the target i in kth two field picture there is no that the match is successful can with other object matchings in kth two field picture, or retain and wait in next frame and find coupling target.