CN103971103A - People counting system - Google Patents

Abstract

The invention relates to a people counting system. The people counting system is characterized by comprising a human body target detection module and a human body target tracking counting module. The human body target detection module comprises a Kinect depth transducer, an interesting target detection module and a human face target detection module. The human body target tracking counting module comprises a target model set, a first matching module, a second matching module and a counting module and is used for counting human body target contours in the target model set. Compared with the prior art, the people counting system has the advantages that the system conducting people counting based on the Kinect depth sensor is adopted, counting results are accurate, and the system can adapt to any scene of a controlled environment and can conduct people counting in a detection space in real time.

Description

A kind of passenger number statistical system capable

Technical field

The present invention relates to a kind of passenger number statistical system capable.

Background technology

In recent years, along with the fast development of China's economy and the raising of living standards of the people, people's activity in public places constantly increased.On the one hand, the public environment such as market, public place of entertainment and service location are blocked up because the increase of number becomes, and this very easily causes the generation of accidents such as trampling.On the other hand, as the operation side of market, public place of entertainment and service location, in order to survive in fierce market competition, the situation that need to distribute according to number, adjusts the management tactics of oneself.As can be seen here, the number information in statistics public place has more and more heavier meaning.

In order to obtain the number information in public place, we often adopt the method for artificial counting, or according to registering, the information such as ticketing infers the number in public place.And these methods have the shortcomings and limitations of oneself.For the method for artificial counting, first, the method need to expend a large amount of manpowers; Secondly, be dynamic because number distributes, and artificial counting carries out in the at one end time often, we just cannot obtain the distribution situation of interior number of all time periods like this.And for the method for inferring numbers according to other information, only have minority public place to exist to register and ticketing etc. and the information such as number is relevant,, there is not this information in other open public place, and the method that Here it is has significant limitation.Therefore visible, we are badly in need of wanting a kind of passenger number statistical system capable automatically.

In recent years, along with the fast development of video monitoring counting, it is more and more extensive that the application places of video monitoring system is becoming---public transport, customs, airport, station, market and business hall.In the different application scenarios of these demands, video monitoring system is by obtaining the video information of monitoring objective, video is recorded, recalled and monitors, and according to video information automatic or manual make corresponding actions, with realize to monitoring objective control, supervision, intelligent management and safety precaution.

In video monitoring system numerous and complicated ground monitoring objective, human body target is important ingredient.Application scenarios in a lot of video monitoring systems enters in market and business hall, and statistics human body target number is one of most important function of system.Also just therefore, the passenger number statistical system capable based on Video Supervision Technique becomes with the fastest developing speed at present.Performance is passenger number statistical system capable the most reliably.

At present, the common demographic method based on video monitoring can be divided three classes:

1, the fitting process based on low-level image feature

The method is by extracting the bottom statistical nature of image and it being carried out to Function Fitting and carry out the human body target number in computed image.In the bottom statistical nature of image, the statistical nature of pixel can reflect human body target ground distribution situation in image preferably.Therefore, the statistical nature of pixel is often used as the fitting parameter of low-level image feature.Conventional pixels statistics feature comprises global characteristics and internal edge feature two classes.Global characteristics comprises the interior number of pixels of geological information, positional information and the agglomerate of the human body target agglomerate splitting in the middle of foreground image.And internal edge feature comprises the features such as edge pixel number, edge direction and Minkowski dimension.

Method based on low-level image feature matching is regarded the number of human body target in image as statistical information, and has neglected the meaning of single human body target, each independent human body target is not followed the tracks of.This makes this method be simple and easy to use, but also makes its counting precision decline, the occasion of the method common-use words guestimate number.

2, the tracing based on unique point

The method can be divided into the method based on light stream and the method based on descriptive model.First method based on light stream selects one group of unique point remaining unchanged in inter texture feature, has searched for tracking afterwards by local matching.The common method based on light stream comprises that Lucas and Kanade mate the method for correspondence image in stereoscopic vision and adopt the KLT algorithm of the poor quadratic sum of gradation of image as Feature Points Matching criterion based on in-plane displancement motion model.First method based on descriptive model is described modeling to all unique points, and the unique point set that has corresponding relation by Selection Model in each two field picture is afterwards to realize the tracking to target.

Tracking based on unique point has been described the motion of human body target effectively, makes passenger number statistical system capable counting more accurate, but also corresponding raising of computation complexity simultaneously.Meanwhile, based on light stream but method change under complicated situation often owing to occurring that mistake coupling causes following the tracks of unsuccessfully at textural characteristics.And the calculating of high time complexity in the loaded down with trivial details learning process that method based on model is obtained due to model and pattern search makes it often be difficult to meet the requirement of real-time of system.

3, based target detects the method for following the tracks of

First the method is carried out background modeling and is extracted the foreground target (normally moving target) in video, and get in foreground target human body target by the connected domain analysis template matches that lives, then adopt track algorithm to follow the tracks of to realize demographics to the human body target detecting.The method precision is high, dependable performance, but cannot overcome real time problems.

Summary of the invention

Technical matters to be solved by this invention for above-mentioned prior art provide a kind of in real time, add up accurate passenger number statistical system capable.

The present invention solves the problems of the technologies described above adopted technical scheme: a kind of passenger number statistical system capable, is characterized in that: comprise human body target detection module and human body target tracking counting module, wherein human body target detection module comprises:

Kinect depth transducer, Kinect depth transducer is aimed at detection space, for the image scene in real-time shot detection space, depth image and coloured image after Kinect depth transducer energy output needle aligns to same width image scene;

Interesting target detection module, the depth image of Kinect sensor output, as the input of interesting target detection module, by depth image being done to interesting target Check processing, obtains the interesting target image of a secondary rectangular shape;

Face module of target detection, the coloured image of the interesting target image of interesting target detection module output and the output of Kinect depth transducer is as two input parameters of face module of target detection, for detection of whether having face target in described interesting target image;

Human body target tracking counting module comprises:

Set of object models, for preserving human body objective contour, the initial value of set of object models is zero;

The first matching module, when described face module of target detection detects while having face target in described interesting target image, described interesting target image is mated with the human body target profile in set of object models by the first matching module, if the match is successful, described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If mate unsuccessfully, described interesting target image is fresh target, and this fresh target is added into set of object models;

The second matching module, when described face module of target detection does not detect while having face target in described interesting target image, described interesting target image is mated with the human body target profile in set of object models by the second matching module, if the match is successful, described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If mate unsuccessfully, described interesting target image is fresh target, and this fresh target is added into set of object models;

Counting module, counts for the human body target profile to set of object models.

As improvement, the process of depth image being done to interesting target Check processing is:

According to the coordinate system of depth image, in detection space, mark off in advance a space interested, this space projection interested becomes a rectangular extent in depth image, and this rectangular extent is called projection rectangle;

Suppose that a certain frame depth image is D (x, y), the projection rectangle upper left point coordinate of space interested in depth image is (x ₀, y ₀), the pixel wide of projection rectangle and be highly respectively w and h, minimum distance and the maximum distance of space length Kinect depth transducer interested are respectively d _minand d _max, according to following formula, depth image is carried out to binaryzation:

$B(x,y)=\left\{\begin{array}{cc}255& x_0\≤x{\≤x}_0+w,y_0\≤y\≤y_0+\mathrm{hand}{d}_{\min }\≤D(x,y)\≤d_{\max }\\ 0& \mathrm{else}\end{array}\right.$

Now, bianry image B (x, y) white portion represents the target in space interested, the target zone of now obtaining is irregular geometric configuration and has sometimes cavity, then using the boundary rectangle of bianry image B (x, the y) profile calculating as interesting target image transfer to face module of target detection.

Improve, described face module of target detection adopts the people's face detection algorithm based on Adaboost to detect in described interesting target image whether have face target again, and the people's face detection algorithm based on Adaboost is conventional method of the prior art.

Improve, described the first matching module is or/and the second matching module adopts following method to mate with the human body target profile in set of object models interesting target image again:

When described face module of target detection detects while having face target in described interesting target image, first described interesting target image being carried out to kinematic relation with all human body objective contours in set of object models mates, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, again all human body objective contours in described interesting target image and set of object models are carried out to Histogram Matching, if Histogram Matching success, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If Histogram Matching is unsuccessful, show that described interesting target image is fresh target, is added into this fresh target in set of object models;

When described face module of target detection detects while whether having face target in described interesting target image, described interesting target image is carried out to kinematic relation with all human body objective contours in set of object models to be mated, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, abandon this interesting target image.

Described the first matching module is or/and the second matching module can also adopt following method to mate with the human body target profile in set of object models interesting target image:

When described face module of target detection detects while having face target in described interesting target image, first described interesting target image being carried out to kinematic relation with all human body objective contours in set of object models mates, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, again all human body objective contours in described interesting target image and set of object models are carried out to Feature Points Matching, if Feature Points Matching success, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If Feature Points Matching is unsuccessful, show that described interesting target image is fresh target, is added into this fresh target in set of object models;

When described face module of target detection detects while whether having face target in described interesting target image, described interesting target image is carried out to kinematic relation with all human body objective contours in set of object models to be mated, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, abandon this interesting target image.

Described kinematic relation matching process is as follows:

Suppose that current interesting target image is t _c(x _c, y _c, w _c, h _c, z _c), wherein, (x _c, y _c) be the pixel coordinate of current interesting target image rectangular extent upper left point, w _cand h _cbe respectively the current interesting target image of target t _cpixel wide and height, z _crepresent whether current interesting target image contains face target, z _c∈ { 0,1}, z _cbe in the current interesting target image of 0 expression without face target, and z _cbe to have face target in the current interesting target image of 1 expression;

Suppose that the current goal set of having found is T={t ₁, t ₂, t ₃t _n, the element t in goal set _i(x _i, y _i, w _i, h _i, z _i), i ∈ 1,2,3 ... n}, wherein for target t in goal set _ipixel coordinate, w _iand h _ibe respectively target t _ipixel wide and height, z _irepresent target t _iwhether contain face target, z _i∈ { 0,1}, z _ibe 0 expression current goal t _imiddle without face target, and z _ibe 1 expression target t _iin have face target;

Current interesting target image t _cwith target t _iparameter while meeting following formula, judge current interesting target image t _cwith target t _ikinematic relation the match is successful:

$\left\{\begin{array}{c}(x_c+\frac{w_c}{2})-(x_i+\frac{w_i}{2})\≤\mathrm{\α}\\ (y_c+\frac{h_c}{2})-(y_i+\frac{h_i}{2})\≤\mathrm{\β}\end{array}\right.,$ In this formula, α and β are predefined parameter, and the numerical value that different scene α and β arrange can be different.

Compared with prior art, the invention has the advantages that: adopt based on Kinect depth transducer and carry out the system of demographics, statistics is more accurate, can be applicable to the scene of any controlled environment, and number in can real-time statistics detection space.

Brief description of the drawings

Fig. 1 is the block diagram of passenger number statistical system capable in the embodiment of the present invention.

Fig. 2 is the optional process flow diagram of the one of matching process in the embodiment of the present invention.

Fig. 3 is the optional process flow diagram of the another kind of matching process in the embodiment of the present invention.

Embodiment

Below in conjunction with accompanying drawing, embodiment is described in further detail the present invention.

Passenger number statistical system capable provided by the invention, as shown in Figure 1, comprises human body target detection module and human body target tracking counting module, and wherein human body target detection module comprises:

Kinect depth transducer 1, Kinect depth transducer is aimed at detection space, for the image scene in real-time shot detection space, depth image and coloured image after Kinect depth transducer energy output needle aligns to same width image scene;

Interesting target detection module 2, the depth image of Kinect sensor output, as the input of interesting target detection module 2, by depth image being done to interesting target Check processing, obtains the interesting target image of a secondary rectangular shape;

Face module of target detection 3, the coloured image that the interesting target image that interesting target detection module 2 is exported and Kinect depth transducer 1 are exported is as two input parameters of face module of target detection, for detection of whether having face target in described interesting target image;

Human body target tracking counting module comprises:

Set of object models 4, for preserving human body objective contour, the initial value of set of object models is zero;

The first matching module 5, when described face module of target detection detects while having face target in described interesting target image, described interesting target image is mated with the human body target profile in set of object models by the first matching module, if the match is successful, described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If mate unsuccessfully, described interesting target image is fresh target, and this fresh target is added into set of object models;

The second matching module 6, when described face module of target detection does not detect while having face target in described interesting target image, described interesting target image is mated with the human body target profile in set of object models by the second matching module, if the match is successful, described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If mate unsuccessfully, described interesting target image is fresh target, and this fresh target is added into set of object models;

Counting module 7, counts for the human body target profile to set of object models 3.

The process that wherein interesting target detection module 2 does interesting target Check processing to depth image is:

According to the coordinate system of depth image, in detection space, mark off in advance a space interested, this space projection interested becomes a rectangular extent in depth image, and this rectangular extent is called projection rectangle;

Suppose that a certain frame depth image is D (x, y), the projection rectangle upper left point coordinate of space interested in depth image is (x ₀, y ₀), the pixel wide of projection rectangle and be highly respectively w and h, minimum distance and the maximum distance of space length Kinect depth transducer interested are respectively d _minand d _max, according to following formula, depth image is carried out to binaryzation:

$B(x,y)=\left\{\begin{array}{cc}255& x_0\≤x{\≤x}_0+w,y_0\≤y\≤y_0+\mathrm{hand}{d}_{\min }\≤D(x,y)\≤d_{\max }\\ 0& \mathrm{else}\end{array}\right.$

Now, bianry image B (x, y) white portion represents the target in space interested, the target zone of now obtaining is irregular geometric configuration and has sometimes cavity, then using the boundary rectangle of bianry image B (x, the y) profile calculating as interesting target image transfer to face module of target detection.

And described face module of target detection 3 adopts the people's face detection algorithm based on Adaboost to detect in described interesting target image whether have face target.People's face detection algorithm based on Adaboost is conventional algorithm of the prior art.

The method that described the first matching module 5 and the second matching module 6 mate with the human body target profile in set of object models interesting target image has two kinds, and wherein one is, shown in Figure 2:

When described face module of target detection detects while having face target in described interesting target image, first described interesting target image being carried out to kinematic relation with all human body objective contours in set of object models mates, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, again all human body objective contours in described interesting target image and set of object models are carried out to Histogram Matching, if Histogram Matching success, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If Histogram Matching is unsuccessful, show that described interesting target image is fresh target, is added into this fresh target in set of object models;

When described face module of target detection detects while whether having face target in described interesting target image, described interesting target image is carried out to kinematic relation with all human body objective contours in set of object models to be mated, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, abandon this interesting target image.

Algorithm based on kinematic relation and color histogram coupling can be applicable to the scene of any controlled environment, the human body target overwhelming majority who occurs in this scene is moving target, and owing to being subject to the environmental restraint of scene, the human body target movement locus scope of living in of porch is relatively fixing, movement velocity is slower, therefore first system is used kinematic relation to mate target in this scene, but because the movement velocity of human body target also there will be the situation that exceedes matching range sometimes, system has been carried out again color histogram and has been mated to guarantee the accuracy to human body target tracking.

The method that described the first matching module and the second matching module mate with the human body target profile in set of object models interesting target image, another mode is, shown in Figure 3:

When described face module of target detection detects while having face target in described interesting target image, first described interesting target image being carried out to kinematic relation with all human body objective contours in set of object models mates, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, again all human body objective contours in described interesting target image and set of object models are carried out to Feature Points Matching, if Feature Points Matching success, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If Feature Points Matching is unsuccessful, show that described interesting target image is fresh target, is added into this fresh target in set of object models;

When described face module of target detection detects while whether having face target in described interesting target image, described interesting target image is carried out to kinematic relation with all human body objective contours in set of object models to be mated, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, abandon this interesting target image.

Human body target with entrance scene is compared, although the target movement velocity that target pause place occurs is slower, sometimes even in static state, the target time of occurrence under this scene is long, and often has rotation, partial occlusion or even block completely.Therefore, in order to ensure that the target of target pause place is carried out to correct counting, system provided by the invention has been used the method based on Feature Points Matching to replace based on the histogrammic matching process of color distribution and has been combined and realize the tracking to human body target with kinematic relation matching method in this scene.

Aforementioned movement relationship match method is as follows:

Suppose that current interesting target image is t _c(x _c, y _c, w _c, h _c, z _c), wherein, (x _c, y _c) be the pixel coordinate of current interesting target image rectangular extent upper left point, w _cand h _cbe respectively the current interesting target image of target t _cpixel wide and height, z _crepresent whether current interesting target image contains face target, z _c∈ { 0,1}, z _cbe in the current interesting target image of 0 expression without face target, and z _cbe to have face target in the current interesting target image of 1 expression;

Suppose that the current goal set of having found is T={t ₁, t ₂, t ₃t _n, the element t in goal set _i(x _i, y _i, w _i, h _i, z _i), i ∈ 1,2,3 ... n}, wherein for target t in goal set _ipixel coordinate, w _iand h _ibe respectively target t _ipixel wide and height, z _irepresent target t _iwhether contain face target, z _i∈ { 0,1}, z _ibe 0 expression current goal t _imiddle without face target, and z _ibe 1 expression target t _iin have face target;

Current interesting target image t _cwith target t _iparameter while meeting following formula, judge current interesting target image t _cwith target t _ikinematic relation the match is successful:

$\left\{\begin{array}{c}(x_c+\frac{w_c}{2})-(x_i+\frac{w_i}{2})\≤\mathrm{\α}\\ (y_c+\frac{h_c}{2})-(y_i+\frac{h_i}{2})\≤\mathrm{\β}\end{array}\right.,$ In this formula, α and β are predefined parameter, and the numerical value that different scene α and β arrange can be different.

In above-mentioned two kinds of methods, Histogram Matching and based on characteristic point matching method, all adopts routine techniques of the prior art.

In addition, set of object models can be deleted the overtime target of not upgrading, and the overtime time presets.

Claims (6)

1. a passenger number statistical system capable, is characterized in that: comprise human body target detection module and human body target tracking counting module, wherein human body target detection module comprises:

Kinect depth transducer, Kinect depth transducer is aimed at detection space, for the image scene in real-time shot detection space, depth image and coloured image after Kinect depth transducer energy output needle aligns to same width image scene;

Interesting target detection module, the depth image of Kinect sensor output, as the input of interesting target detection module, by depth image being done to interesting target Check processing, obtains the interesting target image of a secondary rectangular shape;

Face module of target detection, the coloured image of the interesting target image of interesting target detection module output and the output of Kinect depth transducer is as two input parameters of face module of target detection, for detection of whether having face target in described interesting target image;

Human body target tracking counting module comprises:

Set of object models, for preserving human body objective contour, the initial value of set of object models is zero;

The first matching module, when described face module of target detection detects while having face target in described interesting target image, described interesting target image is mated with the human body target profile in set of object models by the first matching module, if the match is successful, described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If mate unsuccessfully, described interesting target image is fresh target, and this fresh target is added into set of object models;

The second matching module, when described face module of target detection does not detect while having face target in described interesting target image, described interesting target image is mated with the human body target profile in set of object models by the second matching module, if the match is successful, described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If mate unsuccessfully, described interesting target image is fresh target, and this fresh target is added into set of object models;

Counting module, counts for the human body target profile to set of object models.

2. passenger number statistical system capable according to claim 1, is characterized in that: the process of depth image being done to interesting target Check processing is:

According to the coordinate system of depth image, in detection space, mark off in advance a space interested, this space projection interested becomes a rectangular extent in depth image, and this rectangular extent is called projection rectangle;

Suppose that a certain frame depth image is D (x, y), the projection rectangle upper left point coordinate of space interested in depth image is (x ₀, y ₀), the pixel wide of projection rectangle and be highly respectively w and h, minimum distance and the maximum distance of space length Kinect depth transducer interested are respectively d _minand d _max, according to following formula, depth image is carried out to binaryzation:

$B(x,y)=\left\{\begin{array}{cc}255& x_0\≤x{\≤x}_0+w,y_0\≤y\≤y_0+\mathrm{hand}{d}_{\min }\≤D(x,y)\≤d_{\max }\\ 0& \mathrm{else}\end{array}\right.$

Now, bianry image B (x, y) white portion represents the target in space interested, the target zone of now obtaining is irregular geometric configuration and has sometimes cavity, then using the boundary rectangle of bianry image B (x, the y) profile calculating as interesting target image transfer to face module of target detection.

3. passenger number statistical system capable according to claim 1, is characterized in that: described face module of target detection adopts the people's face detection algorithm based on Adaboost to detect in described interesting target image whether have face target.

4. passenger number statistical system capable according to claim 1, is characterized in that: described the first matching module is or/and the second matching module adopts following method to mate with the human body target profile in set of object models interesting target image:

When described face module of target detection detects while having face target in described interesting target image, first described interesting target image being carried out to kinematic relation with all human body objective contours in set of object models mates, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, again all human body objective contours in described interesting target image and set of object models are carried out to Histogram Matching, if Histogram Matching success, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If Histogram Matching is unsuccessful, show that described interesting target image is fresh target, is added into this fresh target in set of object models;

When described face module of target detection detects while whether having face target in described interesting target image, described interesting target image is carried out to kinematic relation with all human body objective contours in set of object models to be mated, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, abandon this interesting target image.

5. passenger number statistical system capable according to claim 1, is characterized in that: described the first matching module is or/and the second matching module adopts following method to mate with the human body target profile in set of object models interesting target image:

When described face module of target detection detects while having face target in described interesting target image, first described interesting target image being carried out to kinematic relation with all human body objective contours in set of object models mates, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, again all human body objective contours in described interesting target image and set of object models are carried out to Feature Points Matching, if Feature Points Matching success, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If Feature Points Matching is unsuccessful, show that described interesting target image is fresh target, is added into this fresh target in set of object models;

When described face module of target detection detects while whether having face target in described interesting target image, described interesting target image is carried out to kinematic relation with all human body objective contours in set of object models to be mated, if the match is successful for kinematic relation, show that described interesting target image is non-fresh target, with the human body target profile matching in this non-fresh target renewal set of object models; If kinematic relation coupling is unsuccessful, abandon this interesting target image.

6. according to the passenger number statistical system capable described in claim 4 or 5, it is characterized in that: described kinematic relation matching process is as follows:

Suppose that current interesting target image is t _c(x _c, y _c, w _c, h _c, z _c), wherein, (x _c, y _c) be the pixel coordinate of current interesting target image rectangular extent upper left point, w _cand h _cbe respectively the current interesting target image of target t _cpixel wide and height, z _crepresent whether current interesting target image contains face target, z _c∈ { 0,1}, z _cbe in the current interesting target image of 0 expression without face target, and z _cbe to have face target in the current interesting target image of 1 expression;

Suppose that the current goal set of having found is T={t ₁, t ₂, t ₃t _n, the element t in goal set _i(x _i, y _i, w _i, h _i, z _i), i ∈ 1,2,3 ... n}, wherein for target t in goal set _ipixel coordinate, w _iand h _ibe respectively target t _ipixel wide and height, z _irepresent target t _iwhether contain face target, z _i∈ { 0,1}, z _ibe 0 expression current goal t _imiddle without face target, and z _ibe 1 expression target t _iin have face target;

Current interesting target image t _cwith target t _iparameter while meeting following formula, judge current interesting target image t _cwith target t _ikinematic relation the match is successful:

$\left\{\begin{array}{c}(x_c+\frac{w_c}{2})-(x_i+\frac{w_i}{2})\≤\mathrm{\α}\\ (y_c+\frac{h_c}{2})-(y_i+\frac{h_i}{2})\≤\mathrm{\β}\end{array}\right.,$ In this formula,, α and β are predefined parameter.