CN106096553A - A kind of pedestrian traffic statistical method based on multiple features - Google Patents

Abstract

The invention discloses a kind of pedestrian count method based on multiple features；It is characterized in that carrying out as follows: 1, gather several positive samples pictures comprising a number of people and do not comprise the negative sample picture foundation training storehouse of the number of people；2, the training sample picture in training storehouse is extracted Haar feature, use Adaboost algorithm one the Haar feature number of people grader of training improved；Training sample picture in training storehouse is extracted Hog feature, uses SVM algorithm one Hog feature number of people grader of training；3, in the preset belt-like zone of video, load Haar feature number of people grader, carry out pedestrian detection, it is thus achieved that number of people candidate region；4, number of people candidate region is loaded Hog feature classifiers and carry out cascade filtration, detect the number of people, set up headform.5, use temporal and spatial correlations parser that the number of people detected is tracked counting.The present invention is by merging pedestrian head Haar feature and Hog feature, thus ensures accuracy and the Statistical Speed of pedestrian count.

Description

A kind of pedestrian traffic statistical method based on multiple features

Technical field

The invention belongs to technical field of video processing based on computer vision technique, relate generally to a kind of based on multiple features Pedestrian traffic statistical method.

Background technology

Along with the fast development of urbanization, various economy, safety problem that population densification is brought also become and work as Modern society major issue.The statistics of pedestrian information is not only related to security protection industry, and also plays at the aspect such as traffic, business Huge effect.Management personnel can carry out the rational management of human and material resources by pedestrian stream statistics of variables information, and business is certainly Plan person can carry out next step investment decision, and the construction of traffic route can also make rationally rule according to the size of artificial abortion Draw.

It addition, along with traditional display, storage, monitoring system universal of playback mode, existing monitoring system carries Intelligent monitoring technology for value-added service becomes the focus of industry development.Utilization is arranged on heavy construction, amusement and leisure place, purchases The monitoring system of the public places such as thing facility, it is achieved the accurate metering of the volume of the flow of passengers and crowd density being estimated, monitor Group is movable, it is ensured that the safety of crowd and analyze the functions such as Trip distribution rule become in current monitor application in the urgent need to merit Energy.Thus, pedestrian count technology has the application value of reality and the biggest development prospect.

At present, it is thus achieved that the method for pedestrian's traffic statistics can be largely classified into three major types: 1 manually adds up 2. utilizes sensor to enter Every trade people flow rate statistical 3. pedestrian based on computer vision traffic statistics.Utilize and manually pedestrian's flow is added up, though should So method is simple, but when adding up scene pedestrian stream amount and being bigger, artificial statistics needs to put into substantial amounts of manpower and energy, not only Waste time and energy, and the accuracy rate added up also will not be the highest, the poor practicability of the method.Pedestrian's flow is carried out utilizing sensor In the field of statistics, research worker achieves some achievements in research, although utilize sensor can obtain pedestrian more accurately Traffic statistics, but the method is relatively costly.The method of main flow is pedestrian's flow based on computer vision system at present Counting, substantial amounts of achievement in research has been emerged in large numbers in pedestrian's traffic statistics field, but owing to there is illumination variation, target deformation, multiobject Block, high in cost of production factor, how reality scene realizes the pedestrian count system that accuracy rate is high, real-time is good and remains one Individual difficult point.

Summary of the invention

Accuracy and real time problems, and conventional pedestrian's skill cannot be taken into account present in above-mentioned prior art simultaneously The problem that art detects pedestrian's serious shielding in high-density scene and cannot be suitable for, the present invention proposes a kind of reality based on multiple features Time pedestrian traffic statistical method, to improve artificial abortion statistics accuracy rate and real-time, thus improve artificial abortion statistics practicality.

The present invention solves that technical problem adopts the following technical scheme that

The feature of a kind of pedestrian traffic statistical method based on multiple features of the present invention is to carry out as follows:

Step 1, utilize photographic head to shoot one group of pedestrian to monitor Sample video and one group of monitor and detection video, gather described row Number of people image in people's sample monitor video, thus constitute N width positive sample graph image set P, monitor in Sample video with described pedestrian Background image in addition to number of people image is as negative sample training image collection C；

Step 2, described N width positive sample graph image set P and M negative sample training image collection C is made normalized, and carry respectively Take image HOG feature, thus obtain HOG characteristic vector set F_g；

The Adaboost algorithm that step 3, utilization improve is to described N width positive sample graph image set P and M negative sample training image collection C is trained, it is thus achieved that the head grader P of Haar feature_r；

Step 4, utilization SVM algorithm are to described HOG characteristic vector set F_gIt is trained, it is thus achieved that the head part of HOG feature Class device P_g；

Step 5, the totalframes assuming described monitor and detection video are J；Defined variable j, and initialize j=1；

Step 6, in described monitor and detection video jth frame detection image one band-like detection area R is set_j, utilize described The head grader P of Haar feature_rTo described band-like detection area R_jCarry out one-level number of people detection, it is thus achieved that jth frame detection image Head candidate region ROI_j；

Step 7, utilize the number of people grader P of described Hog feature_gHead candidate region to described jth frame detection image ROI_jCarry out two grades of number of people detections, it is thus achieved that the head model set of jth frame detection image, be designated as1≤t_j≤T_j；Represent the t of jth frame detection image_j Individual head model；And have: Represent jth frame detection image The abscissa of the t head model center；Represent the t head model center vertical coordinate of jth frame detection image； The t head model of expression jth frame detection image motion cumulative amount in vertical direction；Represent jth frame detection figure The gray value of t head model of picture；Represent the head circular degree of t head model of jth frame detection image；Represent that jth frame detects the t the head model of image matching times in described band-like detection area；

Step 8, the standard head model set Ped ' of acquisition jth frame detection image_j；

Step 9, by jth frame detection image standard head model set Ped '_jIt is stored in pedestrian queuing model Q；

Step 10, judging whether j+1 ＞ J sets up, if setting up, then performing step 13；Otherwise, j+1 is assigned to j, and holds Row step 6 is to step 8, thus obtains the standard head model set Ped ' of jth+1 frame detection image_j+1；

Step 11, utilize described jth+1 frame detection image standard head model set Ped '_j+1To described pedestrian's queue Model Q is updated, thus obtains the pedestrian queuing model Q ' after renewal；

Step 12, judge whether j+2 ＞ J sets up, if setting up, then step 13；Otherwise, j+2 is assigned to j+1, and returns Step 11 performs；

Step 13, the head model traveled through in the pedestrian dummy queue Q ' after described renewal, if the coupling in head model Number of times more than set threshold value, then retains corresponding head model, otherwise, deletes corresponding head model, thus the most more New described pedestrian dummy queue Q '；

The head model in pedestrian dummy queue Q after step 13, traversal renewal, and statistically pedestrian's number and lower pedestrian Number, when motion cumulative amount is more than " 0 ", then add up up number；Otherwise, cumulative descending number.

The feature of real-time pedestrian count method based on multiple features of the present invention lies also in,

Described step 3 is to carry out as follows:

Step 3.1, to any i-th sample graph in described N width positive sample graph image set P and M negative sample training image collection C As e_iIt is marked, if described i-th sample image e_iFor positive sample image, then make i-th sample image e_iLabelling b_i=1； If described i-th sample image e_iFor negative sample image, then make i-th sample image e_iLabelling b_i=0；Thus obtain Haar Feature classifiers training set D={ (e₁,b₁),(e₂,b₂),…(e_i,b_i),…,(e_n,b_n)}；1≤i≤n；

Step 3.2, the number of definition strong classifier are U；Defined variable u；1≤u≤U；

Step 3.3, definition k_maxRepresent threshold value,The i-th sample of u strong classifier when representing kth time iteration Image e_iWeight,The normalization coefficient of all sample weights of u strong classifier when representing kth time iteration；

Step 3.4, initialization u=1

Step 3.5, initialization k=1；

Step 3.6, the weight of n sample image of u strong classifier is when initializing described kth time iteration

Step 3.7, utilize weight corresponding to each sample image that described training set D is sampled, it is thus achieved that u The Weak Classifier of strong classifier

Step 3.6, according to described training set D calculate the Weak Classifier of u strong classifierGrader error

Step 3.7, the weight renewal coefficient of u strong classifier when utilizing formula (1) to calculate kth time iteration

$a_k^{\left(u\right)}=\frac{1}{2}ln\[(1-E_k^{\left(u\right)})/E_k^{\left(u\right)}\]---\left(1\right)$

Step 3.8, the weight of the i-th sample image of u strong classifier when utilizing formula (2) to try to achieve+1 iteration of kthThus the weight of n sample image of u strong classifier when obtaining+1 iteration of kth:

$W_i^{(k+1)\left(u\right)}=\frac{W_i^{\left(k\right)\left(u\right)}}{Z_k^{\left(u\right)}}\left\{\begin{array}{cc}{e}^{-a_k^{\left(u\right)}},& h_k^{\left(u\right)}\left(e_i\right)=b_i\\ e^{a_k^{\left(u\right)}},& \begin{array}{cc}{h}_k^{\left(u\right)}\left(e_i\right)& b_i,h_k^{\left(u\right)}\left(e_i\right)<b_i\end{array}\\ 2e^{a_k^{\left(u\right)}},& h_k^{\left(u\right)}\left(e_i\right)\&NotEqual;b_i,h_k^{\left(u\right)}\left(e_i\right)>b_i\end{array}\right.---\left(2\right)$

In formula (2),Represent the Weak Classifier h of u strong classifier when utilizing kth time iteration_kTo i-th sample Image e_iCarry out the result detected, if testing result is positive sample, thenIf testing result is negative sample, then

Step 3.9, judge k+1 ＞ k_maxWhether set up, as set up, then perform step 3.10, otherwise, k+1 is assigned to k, And return step 3.6 and perform；

Step 3.10, formula (3) is utilized to obtain the u strong classifier f^(u):

$f^{\left(u\right)}=\&lsqb;\underset{k=1}{\overset{k_{max}}{\&Sigma;}}{a}_k^{\left(u\right)}{h}_k^{\left(u\right)}\&rsqb;---\left(3\right)$

Step 3.11, judge whether u+1 ＞ U sets up, if setting up, then by U strong classifier level cascaded series synthesis Haar spy Levy grader F separately_r；Otherwise, u+1 is assigned to u；And return step 3.5 and perform.

Described step 8 is to carry out as follows:

Step 8.1, initialization t=1；

Step 8.2, judge described jth frame detection image the t head modelHead circular degreeIt is No being in set scope, if being in, then performing step 8.3；Otherwise, from the head model of described jth frame detection image Set Ped_jThe t head model of middle deletion jth frame detection image

Step 8.3, judge described jth frame detection image the t head modelGray valueWhether it is in In set scope, if being in, then retain the t head model of described jth frame detection imageOtherwise from jth frame Ped is deleted in the head model set of detection image_jThe t head model

Step 8.4, judge whether t+1 ＞ T sets up, if setting up, then it represents that obtain the standard head mould of jth frame detection image Type set1≤p≤P；Otherwise t+1 is assigned to t, returns step Rapid 8.2 perform.

Described step 11 is to carry out as follows:

Step 11.1, assume in described pedestrian queuing model Q that arbitrarily a head model is designated as Q_l；And Q_l={ x_l,y_l,dy_l, clr_l,dgr_l,score_l}；x_lRepresent the abscissa of the l head model center；y_lRepresent the vertical seat of the l head model center Mark；dy_lRepresent the l head model motion cumulative amount in vertical direction；clr_lRepresent the gray value of l head model； dgr_lRepresent the head circular degree of l head model；score_lRepresent that the l head model is in described band-like detection area Matching times；

Step 11.2, initialization t_j+1=1；

Step 11.3, utilize formula (4) to jth+1 frame detection image standard head model set Ped '_j+1In t_j+1Individual Head modelMate, if formula (4) is set up, then it represents that pth standard head model Ped'_pIt is present in described row In people queuing model Q, and perform step 11.4；Otherwise, by t_j+1Individual head modelJoin described pedestrian's queue mould In type Q, thus obtain the pedestrian queuing model Q ' after renewal；

$(x_{t_{j+1}}^{(j+1)}-{\left(x_l\right)}^2+{(y_{t_{j+1}}^{(j+1)}-y_l)}^2<Dis---\left(4\right)$

In formula (4), Dis represents distance threshold；

Step 11.4, generalIt is assigned to x_l, willIt is assigned to dyl, by score_l+ 1 is assigned to score_l, thus update described the l head model Q_l:

Step 11.5, judge t_j+1+ 1 ＞ T_j+1Whether set up, if setting up, then it represents that complete described jth+1 frame detection figure The standard head model set Ped ' of picture_j+1In the process of all of head model, thus complete described pedestrian queuing model Q's Update, otherwise by t_j+1+ 1 is assigned to t_j+1After, return step 11.3 and perform.

Compared with prior art, the having the beneficial effects that of this law:

1., under height scene, pedestrian's blocks the overall profile Character losing that the most seriously can make pedestrian, conventional Pedestrian detection technology receives serious restriction under this scene, and pedestrian is entered by the present invention by extracting the head feature of pedestrian Row detecting and tracking, the problem effectively solving serious shielding.

2. the present invention devises the number of people detection of a kind of cascade filtration structure, and first movement velocity Haar feature faster is divided Class device carries out the one-level detection acquisition number of people and is selected region, then the HOG feature classifiers using accuracy rate high carries out two grades of screenings, has Effect solve accuracy rate and the unified problem of real-time.

3. the present invention proposes a kind of temporal and spatial correlations parser according to human body head feature in video, utilizes human body head Color, the feature such as shape, complete tracking and the counting of pedestrian in video sequence frame fast and accurately.

4. traditional Adaboost training algorithm is transformed by the present invention, carries out people in first order Haar tagsort During head detection, reduce loss, it is ensured that overall accuracy rate as far as possible.

Accompanying drawing explanation

Fig. 1 is the overall flow figure of the present invention；

Fig. 2 is that pedestrian's queue template of the present invention updates flow chart.

Detailed description of the invention

In the present embodiment, a kind of pedestrian traffic statistical method based on multiple features, regard by loading one section of scene monitoring Screen, arranges a banding number of people detection region in video, loads number of people grader and carry out cascade filtration, use in belt-like zone Temporal and spatial correlations parser is tracked technology to the number of people detected, completes pedestrian count, and overall flow is as it is shown in figure 1, also Carry out in accordance with the following steps

Step 1, utilize photographic head to shoot one group of pedestrian to monitor Sample video and one group of monitor and detection video, gather pedestrian's sample Number of people image in this monitor video, thus constitute N width positive sample graph image set P, monitor in Sample video except number of people figure with pedestrian As background image in addition is as negative sample training image collection C；

Step 2, sample graph image set P positive to N width and M negative sample training image collection C make normalized, all sample images It is normalized to the gray level image of 20 × 20 pixel sizes, and extracts image HOG feature respectively, thus obtain HOG characteristic vector Set F_g；

Step 2.1, the sample image after normalization is carried out gamma compression, thus reduce illumination to sample The impact of picture；

Step 2.2, utilization horizontal edge operator [-1,0,1]^TWith vertical edge operator [-1,0,1]^TCalculate sample image Gradient；

Step 2.3, by the cell factory being divided into several 4 × 4 pixel sizes of each width image uniform, for each carefully Born of the same parents' unit calculates a weighted gradient direction histogram, and wherein rectangular histogram comprises 9 bin, and demarcation interval is 0 °-360 °；

Step 2.4, by multiple neighbouring cell factory one block block of composition, seek its gradient orientation histogram vector, and Being normalized each block block, between block is shared, and i.e. one cell factory can be by multiple block Block is shared.It addition, each cell factory is block independence when being normalized, say, that each cell factory is belonging to it Block block in all can be normalized once, obtain a vector

Step 2.5, by piece image block block vector connect, formed piece image HOG characteristic vector describe Son, the HOG characteristic vector of final all samples describes son and collectively forms HOG characteristic vector set F_g；

Adaboost algorithm sample graph positive to N width image set P and M negative sample training image collection C that step 3, utilization improve enter Row training, it is thus achieved that the head grader P of Haar feature_r；

Any i-th sample image e in step 3.1, sample graph image set P positive to N width and M negative sample training image collection C_i It is marked, if i-th sample image e_iFor positive sample image, then make i-th sample image e_iLabelling b_i=1；If i-th Sample image e_iFor negative sample image, then make i-th sample image e_iLabelling b_i=0；Thus obtain Haar feature classifiers instruction Practice set D={ (e₁,b₁),(e₂,b₂),…(e_i,b_i),…,(e_n,b_n)}；1≤i≤n；

Step 3.2, the number of definition strong classifier are U；Defined variable u；1≤u≤U；

Step 3.3, definition k_maxRepresent threshold value,The i-th sample of u strong classifier when representing kth time iteration Image e_iWeight,The normalization coefficient of all sample weights of u strong classifier when representing kth time iteration；

Step 3.4, initialization u=1

Step 3.5, initialization k=1；

When step 3.6, initialization kth time iteration, the weight of n sample image of u strong classifier is

Step 3.7, utilize weight corresponding to each sample image that training set D is sampled, it is thus achieved that u strong point The Weak Classifier of class device

Step 3.6, according to training set D calculate the Weak Classifier of u strong classifierGrader error

Step 3.7, the weight renewal coefficient of u strong classifier when utilizing formula (1) to calculate kth time iteration

$a_k^{\left(u\right)}=\frac{1}{2}ln\&lsqb;(1-E_k^{\left(u\right)})/E_k^{\left(u\right)}\&rsqb;---\left(1\right)$

Step 3.8, the weight of the i-th sample image of u strong classifier when utilizing formula (2) to try to achieve+1 iteration of kthThus the weight of n sample image of u strong classifier when obtaining+1 iteration of kth:

$W_i^{(k+1)\left(u\right)}=\frac{W_i^{\left(k\right)\left(u\right)}}{Z_k^{\left(u\right)}}\left\{\begin{array}{cc}{e}^{-a_k^{\left(u\right)}},& h_k^{\left(u\right)}\left(e_i\right)=b_i\\ e^{a_k^{\left(u\right)}},& \begin{array}{cc}{h}_k^{\left(u\right)}\left(e_i\right)& b_i,h_k^{\left(u\right)}\left(e_i\right)<b_i\end{array}\\ 2e^{a_k^{\left(u\right)}},& h_k^{\left(u\right)}\left(e_i\right)\&NotEqual;b_i,h_k^{\left(u\right)}\left(e_i\right)>b_i\end{array}\right.---\left(2\right)$

In formula (2),Represent the Weak Classifier h of u strong classifier when utilizing kth time iteration_kTo i-th sample Image e_iCarry out the result detected, if testing result is positive sample, thenIf testing result is negative sample, then Represent that detection sample is flase drop sample,Represent that detection sample is missing inspection sample；

Step 3.9, judge k+1 ＞ k_maxWhether set up, as set up, represented the training of a strong classifier, then performed Step 3.10, otherwise, is assigned to k by k+1, and returns step 3.6 and perform；

Step 3.10, important for institute Weak Classifier is weighted combination, utilize formula (3) to obtain the u strong classifier f^(u):

$f^{\left(u\right)}=\&lsqb;\underset{k=1}{\overset{k_{\max }}{\&Sigma;}}{a}_k^{\left(u\right)}{h}_k^{\left(u\right)}\&rsqb;---\left(3\right)$

Step 3.11, judge whether u+1 ＞ U sets up, if setting up, then by U strong classifier level cascaded series synthesis Haar spy Levy grader F separately_r；Otherwise, u+1 is assigned to u；And return step 3.5 and perform.

Step 4, the present embodiment use Linear SVM algorithm to HOG characteristic vector set F_gIt is trained, it is thus achieved that HOG is special The head grader P levied_g；

Step 5, the totalframes assuming monitor and detection video are J；Defined variable j, and initialize j=1；

Step 6, in monitor and detection video jth frame detection image one band-like detection area R is set_j, wherein banding detection Region R_jWidth be detection image width, height be 40 pixel sizes, utilize the head grader P of Haar feature_rTo banding Detection region R_jCarry out one-level number of people detection, it is thus achieved that the head candidate region ROI of jth frame detection image_j；

Step 7, utilize the number of people grader P of Hog feature_gHead candidate region ROI to jth frame detection image_jCarry out two Level number of people detection, it is thus achieved that the head model set of jth frame detection image, is designated asRepresent the t of jth frame detection image_j Individual head model；And have: Represent jth frame detection image The abscissa of the t head model center；Represent the t head model center vertical coordinate of jth frame detection image； The t head model of expression jth frame detection image motion cumulative amount in vertical direction；Represent jth frame detection figure The gray value of t head model of picture；Represent the head circular degree of t head model of jth frame detection image；Represent that jth frame detects the t the head model of image matching times in band-like detection area；

Step 8, the standard head model set Ped ' of acquisition jth frame detection image_j；

Step 8.1, initialization t=1；

Step 8.2, judge jth frame detection image the t head modelHead circular degreeWhether locate In set scope, if being in, then perform step 8.3；Otherwise, from the head model set Ped of jth frame detection image_j The t head model of middle deletion jth frame detection image

Step 8.3, judge jth frame detection image the t head modelGray valueWhether it is in set In fixed scope, if being in, then retain the t head model of jth frame detection imageOtherwise detect image from jth frame Head model set in delete Ped_jThe t head model

Step 8.4, judge whether t+1 ＞ T sets up, if setting up, then it represents that obtain the standard head mould of jth frame detection image Type set1≤p≤P；Otherwise t+1 is assigned to t, returns step Rapid 8.2 perform.

Step 9, by jth frame detection image standard head model set Ped '_jIt is stored in pedestrian queuing model Q；

Step 10, judging whether j+1 ＞ J sets up, if setting up, then performing step 13；Otherwise, j+1 is assigned to j, and holds Row step 6 is to step 8, thus obtains the standard head model set Ped ' of jth+1 frame detection image_j+1；

Step 11, utilize jth+1 frame detection image standard head model set Ped '_j+1Pedestrian queuing model Q is carried out Update, thus obtain the pedestrian queuing model Q ' after renewal；

Step 11.1, assume in pedestrian queuing model Q that arbitrarily a head model is designated as Q_l；And Q_l={ x_l,y_l,dy_l,clr_l, dgr_l,score_l}；x_lRepresent the abscissa of the l head model center；y_lRepresent the vertical coordinate of the l head model center； dy_lRepresent the l head model motion cumulative amount in vertical direction；clr_lRepresent the gray value of l head model； dgr_lRepresent the head circular degree of l head model；score_lRepresent the l head model in band-like detection area Join number of times；

Step 11.2, initialization t_j+1=1；

Step 11.3, utilize formula (4) to jth+1 frame detection image standard head model set Ped '_j+1In t_j+1Individual Head modelMate, if formula (4) is set up, then it represents that pth standard head model Ped'_pIt is present in pedestrian team In row model Q, it is not necessary to create new head model, only need to update the headform matched in pedestrian's queue therewith, such as Fig. 2 Shown in, and perform step 11.4；Otherwise, by t_j+1Individual head modelJoin in pedestrian queuing model Q, thus obtain Pedestrian queuing model Q ' after must updating；

$(x_{t_{j+1}}^{(j+1)}-{\left(x_l\right)}^2+{(y_{t_{j+1}}^{(j+1)}-y_l)}^2<Dis---\left(4\right)$

In formula (4), Dis represents distance threshold；

Step 11.4, generalIt is assigned to x_l, willIt is assigned to dy_l, by score_l+ 1 is assigned to score_l, thus update the l head model Q_l:

Step 11.5, judge t_j+1+ 1 ＞ T_j+1Whether set up, if setting up, then it represents that complete jth+1 frame detection image Standard head model set Ped '_j+1In the process of all of head model, thus complete the renewal of pedestrian queuing model Q, otherwise By t_j+1+ 1 is assigned to t_j+1After, return step 11.3 and perform.

Step 12, judge whether j+2 ＞ J sets up, if setting up, then step 13；Otherwise, j+2 is assigned to j+1, and returns Step 11 performs；

The head model in pedestrian dummy queue Q ' after step 13, traversal renewal, if the matching times in head model More than set threshold value, represent that head model is real pedestrian head, then retain corresponding head model, otherwise, represent Head model is noise, deletes corresponding head model, thus again updates pedestrian dummy queue Q '；

The head model in pedestrian dummy queue Q after step 13, traversal renewal, and statistically pedestrian's number and lower pedestrian Number, when motion cumulative amount is more than " 0 ", then add up up number；Otherwise, cumulative descending number.

Claims (4)

1. a pedestrian traffic statistical method based on multiple features, is characterized in that carrying out as follows:

Step 1, utilize photographic head to shoot one group of pedestrian to monitor Sample video and one group of monitor and detection video, gather described pedestrian's sample Number of people image in this monitor video, thus constitute N width positive sample graph image set P, monitor in Sample video except people with described pedestrian Background image beyond head image is as negative sample training image collection C；

Step 2, described N width positive sample graph image set P and M negative sample training image collection C is made normalized, and extract figure respectively As HOG feature, thus obtain HOG characteristic vector set F_g；

Described N width positive sample graph image set P and M negative sample training image collection C is entered by the Adaboost algorithm that step 3, utilization improve Row training, it is thus achieved that the head grader P of Haar feature_r；

Step 4, utilization SVM algorithm are to described HOG characteristic vector set F_gIt is trained, it is thus achieved that the head grader of HOG feature P_g；

Step 5, the totalframes assuming described monitor and detection video are J；Defined variable j, and initialize j=1；

Step 6, in described monitor and detection video jth frame detection image one band-like detection area R is set_j, utilize described Haar special The head grader P levied_rTo described band-like detection area R_jCarry out one-level number of people detection, it is thus achieved that the head of jth frame detection image is waited Favored area ROI_j；

Step 7, utilize the number of people grader Pg of the described Hog feature head candidate region ROI to described jth frame detection image_jEnter Two grades of number of people detections of row, it is thus achieved that the head model set of jth frame detection image, are designated asRepresent the t of jth frame detection image_j Individual head model；And have: Represent jth frame detection image The abscissa of the t head model center；Represent the t head model center vertical coordinate of jth frame detection image； The t head model of expression jth frame detection image motion cumulative amount in vertical direction；Represent jth frame detection figure The gray value of t head model of picture；Represent the head circular degree of t head model of jth frame detection image；Represent that jth frame detects the t the head model of image matching times in described band-like detection area；

Step 8, the standard head model set Ped ' of acquisition jth frame detection image_j；

Step 9, by jth frame detection image standard head model set Ped '_jIt is stored in pedestrian queuing model Q；

Step 10, judging whether j+1 ＞ J sets up, if setting up, then performing step 13；Otherwise, j+1 is assigned to j, and performs step Rapid 6 to step 8, thus obtain the standard head model set Ped ' of jth+1 frame detection image_j+1；

Step 11, utilize described jth+1 frame detection image standard head model set Ped '_j+1To described pedestrian queuing model Q It is updated, thus obtains the pedestrian queuing model Q ' after renewal；

Step 12, judge whether j+2 ＞ J sets up, if setting up, then step 13；Otherwise, j+2 is assigned to j+1, and returns step 11 perform；

Step 13, the head model traveled through in the pedestrian dummy queue Q ' after described renewal, if the matching times in head model More than set threshold value, then retain corresponding head model, otherwise, delete corresponding head model, thus again update institute State pedestrian dummy queue Q '；

The head model in pedestrian dummy queue Q after step 13, traversal renewal, and statistically pedestrian's number and descending number, when When motion cumulative amount is more than " 0 ", then add up up number；Otherwise, cumulative descending number.

Real-time pedestrian count method based on multiple features the most according to claim 1, is characterized in that, described step 3 be by Following steps are carried out:

Step 3.1, to any i-th sample image e in described N width positive sample graph image set P and M negative sample training image collection C_i It is marked, if described i-th sample image e_iFor positive sample image, then make i-th sample image e_iLabelling b_i=1；If institute State i-th sample image e_iFor negative sample image, then make i-th sample image e_iLabelling b_i=0；Thus obtain Haar feature Classifier training set D={ (e₁,b₁),(e₂,b₂),…(e_i,b_i),…,(e_n,b_n)}；1≤i≤n；

Step 3.2, the number of definition strong classifier are U；Defined variable u；1≤u≤U；

Step 3.3, definition k_maxRepresent threshold value,The i-th sample image e of u strong classifier when representing kth time iteration_i Weight,The normalization coefficient of all sample weights of u strong classifier when representing kth time iteration；

Step 3.4, initialization u=1

Step 3.5, initialization k=1；

Step 3.6, the weight of n sample image of u strong classifier is when initializing described kth time iteration

Step 3.7, utilize weight corresponding to each sample image that described training set D is sampled, it is thus achieved that u strong point The Weak Classifier of class device

Step 3.6, according to described training set D calculate the Weak Classifier of u strong classifierGrader error Step 3.7, the weight renewal coefficient of u strong classifier when utilizing formula (1) to calculate kth time iteration

$a_k^{\left(u\right)}=\frac{1}{2}ln\&lsqb;(1-E_k^{\left(u\right)})/E_k^{\left(u\right)}\&rsqb;---\left(1\right)$

Step 3.8, the weight of the i-th sample image of u strong classifier when utilizing formula (2) to try to achieve+1 iteration of kthThus the weight of n sample image of u strong classifier when obtaining+1 iteration of kth:

$W_i^{(k+1)\left(u\right)}=\frac{W_i^{\left(k\right)\left(u\right)}}{Z_k^{\left(u\right)}}\left\{\begin{array}{cc}{e}^{-a_k^{\left(u\right)}},& h_k^{\left(u\right)}\left(e_i\right)=b_i\\ e^{a_k^{\left(u\right)}},& \begin{array}{cc}{h}_k^{\left(u\right)}\left(e_i\right)& b_i,h_k^{\left(u\right)}\left(e_i\right)<b_i\end{array}\\ 2e^{a_k^{\left(u\right)}},& h_k^{\left(u\right)}\left(e_i\right)\&NotEqual;b_i,h_k^{\left(u\right)}\left(e_i\right)>b_i\end{array}\right.---\left(2\right)$

In formula (2),Represent the Weak Classifier h of u strong classifier when utilizing kth time iteration_kTo i-th sample image e_iCarry out the result detected, if testing result is positive sample, thenIf testing result is negative sample, then

Step 3.9, judge k+1 ＞ k_maxWhether set up, as set up, then perform step 3.10, otherwise, k+1 is assigned to k, and returns Return step 3.6 to perform；

Step 3.10, formula (3) is utilized to obtain the u strong classifier f^(u):

$f^{\left(u\right)}=\&lsqb;\underset{k=1}{\overset{k_{\max }}{\&Sigma;}}{a}_k^{\left(u\right)}{h}_k^{\left(u\right)}\&rsqb;---\left(3\right)$

Step 3.11, judge whether u+1 ＞ U sets up, if setting up, then U strong classifier level cascaded series is synthesized Haar feature and divide Head grader F_r；Otherwise, u+1 is assigned to u；And return step 3.5 and perform.

Real-time pedestrian count method based on multiple features the most according to claim 1, is characterized in that, described step 8 be by Following steps are carried out:

Step 8.1, initialization t=1；

Step 8.2, judge described jth frame detection image the t head modelHead circular degreeWhether it is in In set scope, if being in, then perform step 8.3；Otherwise, from the head model set of described jth frame detection image Ped_jThe t head model of middle deletion jth frame detection image

Step 8.3, judge described jth frame detection image the t head modelGray valueWhether it is in set In fixed scope, if being in, then retain the t head model of described jth frame detection imageOtherwise detect from jth frame The head model set of image is deleted Ped_jThe t head model

Step 8.4, judge t+1 > whether T set up, if setting up, then it represents that obtain the standard head Models Sets of jth frame detection image Close1≤p≤P；Otherwise t+1 is assigned to t, returns step 8.2 Perform.

Real-time pedestrian count method based on multiple features the most according to claim 1, is characterized in that, described step 11 be by Following steps are carried out:

Step 11.1, assume in described pedestrian queuing model Q that arbitrarily a head model is designated as Q_l；And Q_l={ x_l,y_l,dy_l,clr_l, dgr_l,score_l}；x_lRepresent the abscissa of the l head model center；y_lRepresent the vertical coordinate of the l head model center； dy_lRepresent the l head model motion cumulative amount in vertical direction；clr_lRepresent the gray value of l head model； dgr_lRepresent the head circular degree of l head model；score_lRepresent that the l head model is in described band-like detection area Matching times；

Step 11.2, initialization t_j+1=1；

Step 11.3, utilize formula (4) to jth+1 frame detection image standard head model set Ped '_j+1In t_j+1Individual head ModelMate, if formula (4) is set up, then it represents that pth standard head model Ped'_pIt is present in described pedestrian team In row model Q, and perform step 11.4；Otherwise, by t_j+1Individual head modelJoin described pedestrian queuing model Q In, thus obtain the pedestrian queuing model Q ' after renewal；

$(x_{t_{j+1}}^{(j+1)}-{\left(x_l\right)}^2+{(y_{t_{j+1}}^{(j+1)}-y_l)}^2<Dis---\left(4\right)$

In formula (4), Dis represents distance threshold；

Step 11.4, generalIt is assigned to x_l, willIt is assigned to dy_l, by score_l+ 1 is assigned to score_l, from And update described the l head model Q_l:

Step 11.5, judge t_j+1+1>T_j+1Whether set up, if setting up, then it represents that complete the mark to described jth+1 frame detection image Accuracy portion model set Ped '_j+1In the process of all of head model, thus complete the renewal of described pedestrian queuing model Q, no Then by t_j+1+ 1 is assigned to t_j+1After, return step 11.3 and perform.