CN104239905A - Moving target recognition method and intelligent elevator billing system having moving target recognition function - Google Patents

Abstract

The invention discloses a moving target recognition method and an intelligent elevator billing system having a moving target recognition function. The intelligent elevator billing system having the moving target recognition function comprises a billing device for obtaining and statistics on effective billing information, a detection device for statistics on the number of people actually borne by an elevator, a comparison and processing device connected with the billing device and the detection device and used for controlling the running state of the elevator according to the effective billing information and a comparison result of the number of the actually borne people. If the effective billing information is greater than or equal to the number of the actually borne people, running of the elevator is controlled; if the effective billing information is smaller than the number of the actually borne people, the elevator is controlled to stop running. The detection device comprises a camera shooting module and a moving target recognition module. Unmanned, intelligentized and accurate elevator management is achieved by means of the intelligent elevator billing system.

Description

Motion estimate method and there is the elevator intelligent charge system of this function

Technical field

The invention belongs to image identification technical field, the application of a kind of recognition methods of moving target, and the method specifically in elevator intelligent charge system and this elevator intelligent charge system.

Background technology

In order to improve the running benefit of the means of transports such as elevator, comercial operation elevator starts to adopt automatic payment mode to carry out to charge and operate.In prior art, a kind of self-service charging mode is entirely paid with the conscious active of occupant, do not check occupant's number and paying monitoring apparatus, so charge degree of accuracy is not high, can not meet the object of reasonable charging; Another kind of self-service charging mode adds person number detection device to improve charge precision, such as publication number CN202523143U patent document discloses a kind of elevator intelligent Fare Collection System, this system has pick-up unit for carrying out detection quantitative statistics to the effective strength of elevator cage inside, and its pick-up unit can be checked the number of people by infrared thermal imaging sensing device.But because car is usually all very narrow, this infrared thermal imaging sensing device is more in number, and when being mutually adjacent between passenger, sensitivity is not high, easily mistake numeration; When utilizing the shortcoming of camera head to be that number is more, identification is lower, can not meet the technical purpose that intelligence is checked the number of people well.Above because mistake numeration often causes passenger to pay, but central controller also thinks payment amount and passengers quantity inconsistent (namely think and have people not pay) and do not run elevator; Certainly also likely there is contrary situation, namely have people not pay and elevator brings into operation.

So those skilled in the art are necessary to be improved existing elevator Fare Collection System, need the pick-up unit improving patronage especially.

Summary of the invention

The object of the present invention is to provide a kind of recognition methods of moving target and the application in elevator intelligent charge system thereof.

A recognition methods for moving target, comprises the steps:

The first step, combines inter-frame difference with rim detection, carries out moving Object Segmentation to image sequence;

Second step, based on the moving target parameter rapid extraction of thick sampling;

3rd step, based on the motion estimate of neural network.

Intelligent fee counting method, comprises the steps:

A) charging step, obtains and adds up effective charge information;

B) step is checked, according to recognition methods statistics actual bearer number recited above;

C) compare treatment step, the comparative result of the actual bearer number that the effective charge information obtained according to step a) and step b) obtain determines the running status of carrying tool; If effectively charge information is greater than or equal to actual bearer number, then carrying tool runs; If effectively charge information is less than actual bearer number, then carrying tool is out of service.

Invention further provides and a kind ofly apply motion estimate function thus improve the elevator intelligent charge system of smart charge precision, comprising:

Message accounting, for obtaining and adding up effective charge information;

Pick-up unit, for adding up the number of elevator actual bearer;

Relatively treating apparatus, is connected with above-mentioned message accounting and pick-up unit, controls the running status of elevator according to the comparative result of described effective charge information and described actual bearer number; If effectively charge information is greater than or equal to actual bearer number, then controls elevator and run; If effectively charge information is less than actual bearer number, then control elevator out of service;

It is characterized in that: pick-up unit comprises the identification module of photographing module and moving target.

Preferably, the identification module of moving target adopts the recognition methods of above-mentioned moving target.

Beneficial effect of the present invention there are provided a kind of recognition methods of moving target, this recognition methods all obtains satisfactory result in operational efficiency, recognition accuracy. simultaneously, this algorithm is for the impact of the extraneous factors such as light, shade and stream of people's change, there is comparatively strong adaptability, drastically increase billing accuracy.Intelligent fee counting method and elevator intelligent charge system are by comparing charge information with actual bearer number, and the running status of automatic control electric ladder, achieves unmanned, intelligentized management.

Accompanying drawing explanation

Fig. 1 is the block diagram of the recognition methods of moving target of the present invention;

Fig. 2 is the process flow diagram of the recognition methods of moving target of the present invention;

Fig. 3 is the structured flowchart of elevator intelligent charge system of the present invention.

Embodiment

In recent years, along with the development of video processing technique, Video Applications is also further extensive. wherein, adopt image sequence processing technology to the automatic detection and indentification of particular video frequency Moving Objects, particularly based on real-time stream of people's detection and indentification technology of video, because of its having wide application prospects of institute's tool in intelligent monitoring, intelligent traffic administration system etc., it is the focus paying close attention to research always.

In smart charge system, the video processing technique stream of people adds up, and be a very crucial link, the camera adopted is all similar, and to the process aspect of picture signal, mainly the algorithm of software is had nothing in common with each other, and total people stream counting flow process as shown in Figure 2.

In this flow process, it is the most difficult for following the tracks of and identifying, at light, when angle is certain, how to the turnover identification of the personnel in video and article, does an introduction below.

First, under the condition ensureing accuracy of identification, on VC++ platform, by Direct Show interface in Microsoft DirectX8.1, construct gateway pedestrian and automatically detect number system.

This software uses computer vision (CV) technology, by to the process of continuous videos image and analysis, moving object is detected, separates from video background, carry out again screening, filtering, obtain real moving object, then trace analysis is carried out to it, judge in conjunction with counting line, by the target of coincidence counting rule by behavior, count.By object classification algorithm, all moving object can be divided into " people ", " car ", " unknown object " by us, thus obtains testing result more accurately.Be equipped with multiple filtrator (filtering objects size again; Size rate of change; Shadow of object, movement direction of object; Object width, highly, length breadth ratio, region, feature, the filtration such as speed) and meticulous logic rules judge, truly, accurately can obtain alarming result.Data into and out of number can be transmitted by RS485, RS232, TCP/IP, USB dish.Built-in independent digit I/O mouth can be connected with the equipment of miscellaneous equipment or switch gate easily, as connected DVR, then can be marked with people through out-of-date video-frequency band, be convenient to the retrieval of playback afterwards, the precision that can make up calculating instrument is not enough, as the switch of connection door, then can arrange when the door closes, calculating instrument stops counting.Also optionally join Reports module, by real time and history number undertaken showing by powerful report capability, inquire about, add up, print.

Algorithm in this paper is all achieved within the system.

Practical application shows, compares with previous methods, and algorithm all improves a lot in counting yield and counting accuracy rate, and all can better process influential system such as light conversion, pedestrian's shades.The problems such as the moving Object Segmentation accuracy rate detect for the stream of people under visible ray, existed in recognizer is low, recognition effect is poor, proposed here a kind of new recognition and tracking method.

(1) first utilize Moving Targets in Sequent Images space-time consistency, interframe second order difference (SODP) is combined with rim detection and carries out moving Object Segmentation;

(2) again according to pedestrian movement's model and moving target locality characteristic, by thick method of sampling rapid extraction tracking characteristics vector;

(3) utilize the characteristic components such as moving target outline projection ratio, form factor, and the moving object classification device constructed based on artificial neural network identifies.

Shown by the actual test carried out megastore: the method all obtains satisfactory result in operational efficiency, recognition accuracy.Meanwhile, this algorithm, for the impact of the extraneous factors such as light, shade and stream of people's change, has comparatively strong adaptability.

Herein using hoistway door main entrance gateway video as research object, for avoiding mutually blocking between pedestrian and the calculating of reduction system to each moving target expends, its top view is adopted frequently to detect, by Real-time Collection and process, reach the real-time detection to turnover surveyed area pedestrian, follow the tracks of and statistics.

(1) surveyed area model

1. camera and background relative position are fixed, and there is local background's situation of change, background is transfixion in image sequence.

2. ambient light changes greatly, and is subject to comprising the factor impacts such as light, sunlight and pedestrian's shade.

3. in region may there is multiple moving target in synchronization.

(2) pedestrian movement's model

1. in traveling process, pedestrian's spacing changes greatly. and exist and merge (multiple pedestrian's spacing is from large to small until contact with each other with same speed translation) and division (multiple Moving Objects spacing contacted with each other is changed from small to big until be separated from each other) phenomenon.

2. mode of motion is based on translation motion, also there is rotary movement.

3. pedestrian's direction of primary motion is turnover both direction, and direction change is less.

4. pedestrian is attended by time motion generation (as four limbs swing) of part in main motion process.

5. stream of people's variable density is comparatively large, generally contactless between pedestrian, under large density case, there is the phenomenon that contacts with each other.

In order to effectively carry out following the tracks of to pedestrian and detect, according to surveyed area and stream of people's movement characteristic and different processing functions. surveyed area is divided into motion target tracking region and moving target processing region.

(1) motion target tracking region

Carry out moving object detection, if judge whether, pedestrian occurs congested. occur, proceed to congestion prediction processing module. otherwise carry out target following to motion, to determine its position, direction, speed parameter, and marker motion target enters the initial position of tracing area.

(2) moving target processing region

With region asks that displacement (displacement between motion target tracking region inlet point and processing region inlet point) detects, judge whether moving target is shifting out surveyed area and having comparatively displacement between large regions, if satisfy condition, moving target is identified, and according to recognition result counting.

Whole algorithm forms primarily of three parts, as shown in Figure 1, respectively being (1) moving object detection, (2) follow the tracks of and (3) identify, carrying out respectively below, carrying out motion target tracking detection simultaneously, introducing its direction of motion.

Principle summary:

Pedestrian is large, the direction of motion no regularity of velocity of displacement difference and because of excessive the produced congestion phenomenon of pedestrian density between form changeable, individuality at the volley, all make to become a difficulties to pedestrian's real-time follow-up and identification. be directed to this, propose herein a kind of based on pedestrian tracking under the visible ray of tracing detection, identify and the new method counted.

First, at detection-phase, adopt interframe second order difference (SODP) and edge detection method to carry out moving Object Segmentation, improve the accuracy of moving Object Segmentation,

Secondly, at tracking phase, by pedestrian movement's model, adopt the parameter extraction algorithm based on thick sampling, effectively parameter extraction efficiency under the concurrent multiple goal of raising;

3rd, at cognitive phase, because non-rigid pedestrian's form is changeable, in addition other moving objects and ambient noise impact, utilize target essential characteristic amount (length, width etc.) and conventional linear sorting technique, be difficult to the discriminator to moving target. consider pedestrian's morphological character and algorithm realization efficiency, by introducing in literary composition

(1) moving target outline projection ratio,

(2) characteristic component such as form factor,

(3) Nonlinear Classification and the Serial Distribution Processing feature of artificial neural network is utilized

(4) structure is based on the moving object classification device of neural network, achieves satisfied classification accuracy and classification speed.

Experimental result shows, the method is real-time, and accuracy is high. simultaneously for extraneous factor impact, as light change, stream of people's change etc. have comparatively strong adaptability.

Moving target recognition

In order to carry out the recognition and tracking of moving target, first needing to carry out moving object detection and segmentation to video flowing. video flowing can be regarded as strict sequencing in time, by one of strong correlation group of image sequence that still image forms between consecutive frame, its segmentation result is not only required that in frame, rest image has the rationality of segmentation, but also the continuity of interframe movement object segmentation will be kept. therefore, adopt the Video object segmentation algorithm based on motion edge, in conjunction with movement detecting images and moving target spatial gradient Image Segmentation Using, obtain moving target.

But, due to the no regularity of stream of people's motion, for ensureing to carry out smoothly the tracking of moving target, system need adopt higher sampling frame per second to guarantee the short-term stationarity of motion target tracking proper vector, and namely interframe movement target feature vector difference is less than match error threshold.

But, when frame difference method detects under high frame per second, for the target of moving or deformation is less, often be less than given threshold value because it offsets between consecutive frame, cause detecting unsuccessfully. therefore, we, by utilizing second order difference image (SODP) i.e. laplacian image, change original difference image mode, carry out moving Object Segmentation.

If adjacent three two field pictures are respectively F in sequence image _i,j(t _n-1), F _i,j(t _n), F ( _i,j) (t _n+1), θ is threshold value, and its second order difference bianry image is

$L_{i,j}\left(t_n\right)=\left\{\begin{array}{c}255,|F_{(i.j)}\left(t_{n+1}\right)-2F_{i,j}\left(t_n\right)+F_{i,j}\left(t_{n-1}\right)|\&GreaterEqual;\mathrm{\&theta;}\\ 0,|F_{(i.j)}\left(t_{n+1}\right)-2F_{i,j}\left(t_n\right)+F_{i,j}\left(t_{n-1}\right)|<\mathrm{\&theta;}\end{array}\right\}$

Because segmentation make use of three frame image informations, improve quality and the precision of segmentation. frame per second of then sampling be 18 frames per second time, the method still can obtain and comparatively be satisfied with moving image time gradient image. in second order difference process, because relating to adjacent three two field pictures, a frame time can be produced delayed. because pickup system sampling interval is very little, a frame is delayed can not produce considerable influence to process in real time.

Real-time multi-target is followed the tracks of and process

Moving target after extraction, system need carry out real-time follow-up, determines its motion state at each moment t. and moving target is S at moment t tracking characteristics vector _t=[M (x, y), Attr (a, b, c)]

1. M (x, y) identifies moving target position of form center in x, y plane;

2. Attr (a, b, c) motion target tracking attribute, should meet interframe short-term stationarity.

People is the change of form moment when moving, also with the Merge and split phenomenon between pedestrian. and the change of its area of detection, direct of travel is then relatively little. and therefore, we select area of detection s and direction of motion v at the projection d=||v of direction of primary motion _d||. definition moving target is S at moment t tracking characteristics vector _t=[x, y, s, d]

But, from surveyed area and pedestrian movement's model, the concurrent motion conditions of multiple goal is often there is in scene, the foundation that in frame, total movement target following vector is shouted is completed in real time under high frame per second, this proposes higher rate request to parameter extraction algorithm. here, we propose a kind of moving target parameter Fast Extraction based on thick sampling. according to parameters precision requirement, not only can obtain moving target characteristic parameter to be fast similar to, and for the discontinuous situation that may exist in moving target profile, also can process. be key step below:

1. for moving target b _i(i=1 ..., M), (M is moving target number in image), extracts the boundary rectangle being parallel to direction of primary motion, is designated as; R (b _i) (i=1 ..., M)

2. R (b _i) inside slightly samples, sampling rectangular area is required to determine k value by parameters precision, select k=2 temporarily; Thick sample area is

Wherein,

R (b _i) _x, R (b _i) _ythe projection of moving target boundary rectangle on vertical direction of primary motion and direction of primary motion;

3. according to initial frame overall intensity, Otsu method determination segmentation threshold θ is adopted;

4. according to segmentation threshold θ to boundary rectangle R (b _i) in rectangular block carry out binary segmentation

$D\left(b_i\right)=\left\{\begin{array}{c}255,\mathrm{Count}\left(S_{m,n}\right)\&GreaterEqual;\mathrm{\&theta;}\\ 0,\mathrm{Count}\left(S_{m,n}\right)<\mathrm{\&theta;}\end{array}\right.,$ Wherein S _m,n∈ R (b _i);

5. binary segmentation image D (b in rectangle frame _i) carry out characteristic parameter extraction, as former moving target b _i(i=1 ... M) characteristic parameter is similar to.

Based on the motion estimate of neural network

Through the moving target of video Detection and Extraction, may be the turnover single or multiple pedestrian in doorway, also may be other moving objects (as trolley, bag etc.), need type decision be carried out, if be judged to be pedestrian, then need to judge number further, obtain the pedestrian information passing in and out surveyed area. because form during non-rigid pedestrian movement is changeable and moving target classification diversity (exists object of which movement, single, many people's motions that spacing is nearer), utilize the real-time grading that moving target essential characteristic attribute and conventional linear sorting technique have been difficult to it. map feature according to multidimensional nonlinear in non-rigid pedestrian morphological feature and identification process, by setting up the identification proper vector comprising outline projection ratio and form factor aliquot, and the moving object classification device set up based on BP neural network, completing by target feature vector target classification Nonlinear Mapping .BP neural network model is a kind of multi-level mapping neural network, typical BP network is 3 layers of feedforward stratum network, i.e. input layer, hidden layer (also claiming middle layer) and output layer, any n can be completed and tie up Euclidean space ties up Euclidean space non-linear mapping capability to m, simultaneously because of neural network concurrent distribution process, the needs of process in real time can be met. system is using moving target proper vector as input, after network operations, corresponding desired output is obtained at output layer, i.e. moving target type, reach target classification object.

Neural network designs

Neural network selects 3 layers of BP neural network, and transport function adopts sigmoid function: $f_{\left(x\right)}=\frac{1}{1+e^{-x}}$

(1) identification proper vector is set up

Win target identifiability for carrying, and reduce feature vector dimension.

Identification proper vector is defined as: d=(r, e, f, d, w, h), and each component concrete meaning is as follows:

1. e---the longest linear component and object video projection ratio in moving target profile.

Utilize Hough (Hough) to convert fast fetching and obtain the longest straight length h of moving target profile, with moving target at forward boundary rectangle hypotenuse ratio. d _x, D _ybe respectively the projection of moving region on x, y direction

2. f----form parameter.

According to pedestrian movement's model, select target parallel with projected length ratio on vertical target direction of primary motion.

F is form parameter, d _x, D _ybe respectively the projection of moving region on x, y direction

3. r---circle, wherein:

μ _tfor the mean distance from center of gravity to point, ${\mathrm{\&mu;}}_T=\frac{1}{K}{\mathrm{\&Sigma;}}_{K=D}^{K-1}\left|\right|(x_k,y_k)-(\stackrel{\&OverBar;}{x},\stackrel{\&OverBar;}{y})\left|\right|,$

σ _tfor the mean square deviation of the distance from regional barycenter to point,

${\mathrm{\&sigma;}}_r=\frac{1}{K}{\mathrm{\&Sigma;}}_{K=D}^{K-1}{\left[\right||(x_k,y_k)-(\stackrel{\&OverBar;}{x},\stackrel{\&OverBar;}{y})||-{\mathrm{\&mu;}}_r]}^2;$

4. d---dutycycle (moving target area and direction of primary motion boundary rectangle area ratio).

$d=\frac{a}{D_x{D}_y},$ a=∑ _(x,y)∈R1；

5. w, h---order, mark width, highly.

By form parameter f, area ratio d and circle r, effectively describe moving target shape facility. simultaneously, find in frame is split, moving object (as bag, go-cart) is usually containing more regular contour (linear edge), and non-rigid pedestrian is relatively less, therefore introduced by component e, improve algorithm to people and thing resolution characteristic. simultaneously, according to identification vector dimension, determine that input layer nodal point number is 6.

In BP neural network, node numbers of hidden layers object determines it is a difficult point always. number of network nodes is very few, can not set up complicated judgement border; Number of network nodes is too much, and learning time is long, and network generalization is reduced. by computing formula (m is node numbers of hidden layers order, and l is output node number, is the constant between 1 ~ 10) and real network are trained, and when Hidden unit number is 9, neural network global error has better convergence property.

Output layer is actual is the desired output of network training, and within a detection region, moving target may exist 5 kinds:

1. other classification objects (as the object such as go-cart, bag);

2. single walking;

3. many people's walking (carry on a shoulder pole according to surveyed area size, there are 2 people, 3 people, two kinds of situations) that spacing is nearer

4. local congestion's (occur in out population just to open the door the moment, at surveyed area, pedestrian locally occurs and excessively closely cause congested because of spacing).

Thus, neural network output layer selects 5 nodes.

Sample data process and training set

(1) change of scale

For avoiding, because characteristic component dimension difference causes " supersaturation " phenomenon in network training, carrying out change of scale to each component of proper vector, it being all mapped in [0,1] interval, making that notebook data is unified and standard is convenient to network training.

If target feature vector is s _i=(r _i, e _i, f _i, d _i, w _i, h _i), i=1,2 ..., n

Wherein i is proper vector number. proper vector S after conversion _i'=(r _i', e _i', f _i', d _i', w _i', h _i'), i=1,2 ..., n

Each component between proper vector has following relation:

a）r _i′=r _i,e _i′=e _i,f _i′=f _i,d _i′=d _i；

B) original width is w _i, height h _i, linear conversion

$w_i^{\&prime;}=\frac{w_i-w_{\min }}{w_{\max }-w_{\min }},h_i^{\&prime;}=\frac{h_i-h_{\min }}{h_{\max }-h_{\min }},$ Wherein

Wherein w _max=max (w _i), h _max=max (h _i), w _min=min (w _i); h _min=min (h _i),

In real transform, w _max, h _max, get the wide and high of moving region respectively, the area of detection lower limit s of pedestrian in being detected by motion _min

Get $w_{\min }=h_{\min }=2\sqrt{\frac{s_{\min }}{\mathrm{\&pi;}}}$

(2) training set structure

By carrying out video acquisition, according to the moving object detection of gained sequence frame and segmentation result, its proper vector is set up to moving target each in frame: d _i=(r _i, e _i, f _i, d _i, w _i, h _i) and manual sort is carried out to moving target. last, utilize the moving target proper vector after conversion and this moving target is carried out to the result of manual sort, composition learning sample p _i'=(r _i', e _i', f _i', d _i', w _i', h _i', c _i), wherein specify, as sample p _i'=(r _i', e _i', f _i', d _i', w _i', h _i', c _i) when belonging to c class, c ∈ [0, M-1], the desirable output vector of described neural network is $o_i=\left\{\begin{array}{c}0,i\&NotEqual;c\\ 1,i=c\end{array}\right.,$ I=0 ..., M-1, (M is moving target classification number, and C is classification belonging to sample).

Adopting said method, to all K sample, set up artificial neural network input, output vector pair, the i.e. mode of learning of neural network. this K mode of learning, just constitutes the training set of artificial neural network.

Systematic training and identification process

BP neural network is utilized to carry out classifying and mainly comprise network training and identify two benches. be the defect avoiding the slow and objective function of its speed of convergence to there is local minimum point, network adopts variable step momentum method to carry out BP neural metwork training in the training stage. and training parameter selection is as follows:

Learning rate 0.01;

Learning rate increases ratio 1.05;

Learning rate reduces ratio 0.7;

Momentum constant 0.9;

Error sum of squares index 0.1.

Neural network recognization flow process: in target treatment, to the moving target meeting displacement of targets, direction constrain condition, utilizes neural network classifier to carry out identification. carries out primarily of the following steps:

Stepl. the neural network having completed study is loaded into;

Step2. for the moving target in target treatment, set up target identification proper vector d=(c, e, f, d, w, h);

Step3. proper vector d is converted successively, if the proper vector after conversion is T;

Step4. make T be the input vector of neural network, obtain the output valve of each output unit of neural network

If step5. have then think moving target 6; Belong to classification C.

With reference to the accompanying drawings, the purposes of the intelligent charging method of serving for the maintenance management of carrying tool-elevator in this embodiment and intelligent toll system and charging method is explained.But should be understood that the explanation of embodiment just to technical solution of the present invention below, do not form the restriction to technical solution of the present invention.

In this specific implementation method, the elevator that carrying tool is house, office building runs.The maintenance management service of this elevator is by the owner of elevator or supvr, and namely infrastructure management company carries out, and infrastructure management company charges to resident family to the service provided.

The intelligent charging method that it uses, comprising:

Charging step, this step is by self-service counting equipment, as IC-card reader or fingerprint reader are added up the charging number of times reinstating IC-card reader or fingerprint reader;

Check step, this step is the checkout equipment by being arranged in elevator----intelligence is single, and binocular demographics instrument, carries out detection quantitative statistics to the effective strength in elevator;

Relatively treatment step; This step is by the charging number of times of quantitative statistics in charging step and checks quantitative statistics in step

Effective strength in carrying tool compares the running status determining elevator; If charging number of times is more than or equal to the effective strength in elevator, then elevator can run; If charging number of times is less than the effective strength in carrying tool, then elevator is out of service.

Charging number of times is less than the effective strength in carrying tool, then remind actual being in elevator when elevator is out of service, and the people not reinstating IC-card reader or fingerprint reader reinstates IC-card reader or fingerprint reader carries out paying confirmation.

Charging number of times is less than the effective strength in carrying tool, then, when elevator is out of service, if elevator does not still run after reminding, the keeper to elevator sends warning.

Above-mentioned intelligent fee counting method also comprises smart charge management process, after the charging number of times of every day and the amount of money have been added up by this step, automatically generates the statistical report on the same day.This step also can carry out the inquiry of several data and chart and printing and by Ethernet remote transmission.

Above-mentioned intelligent fee counting method also comprises interactive correspondence service step, and this step can realize the interactive correspondence taken in elevator between people and keeper, and keeper can by this step for the people that takes in elevator provides service.

Fig. 3 intelligent toll system shown in this embodiment is applied to the system principle structure of elevator management.As shown in Figure 3, native system is primarily of automatic accounting device, monocular or binocular camera, motion estimate module, system host (comprise contrast module), the composition such as warning system, intercommunication service system, speech prompting system, remote management and control terminal system, smart charge management software device.Comparison treatment facility passes through CAN or 485 buses and Ethernet and is connected with remote management and control terminal device.Relatively treatment facility is also connected with apparatus for controlling elevator, compares treatment facility and connects a panalarm.Between elevator and keeper, talk back equipment is linked up.Wherein, the identification module of described moving target comprises moving Object Segmentation module, and inter-frame difference is combined with rim detection by it, carries out moving Object Segmentation to image sequence; Based on the moving target parameter rapid extraction module of thick sampling; Based on the motion estimate submodule of neural network.

The area of common car is equipped with single camera and just can meets the demands, and be generally arranged on center or the upper end, entrance hall of car top, the inclination angle of camera can adjust, if car area is excessive or camera setting height(from bottom) is limited, can be equipped with dual camera.The present embodiment adopt single camera its there are 6 communication interfaces, to facilitate and system host, the communication contacts such as elevator control gear.Built-in independent digit I/O end can be connected with the equipment of miscellaneous equipment or switch gate, easily as quit work closing counting instrument behind the door with switch gate equipment connection.

After native system is opened, first the image information of camera is sent to image processing module, processing module uses computer vision (CV) technology, by the process of continuous videos image and analysis, moving object is detected, is separated from video background, carry out again screening, filtering, obtain real moving object, then trace analysis is carried out to it, judge in conjunction with counting line, by the target (passenger) of coincidence counting rule by behavior, count.All moving object can be divided into " people ", " car ", " unknown object " by image processing module, thus obtains testing result more accurately.Be equipped with picture information filtrator again, filtering objects size, size rate of change, shadow of object, movement direction of object, object width, highly, length breadth ratio, region, feature, the information such as speed are carried out meticulous logic rules and are judged, can obtain the incremental data needing paid passenger truly, accurately.

People's logarithmic data of paid passenger sends system host to through PORT COM, automatic accounting device simultaneously, the paid passenger quantity that the real-time payment information of such as IC-card reader is determined after data processing is also sent to system host, contrasted by contrast module, if paid passenger number is less than the determined number of real-time payment information, speech prompting system can be paid by reminding passengers, elevator lobby is opened simultaneously, elevator is out of service, if wait paid-for time what set, such as, after 15 seconds, paid passenger number is still less than the determined number of real-time payment information, warning system is reported to the police to keeper, processed by keeper.If paid passenger number is more than or equal to the determined number of real-time payment information, system host instruction elevator normally runs and completes real time billing work.

The present invention accurately can identify moving target, and be used in the charge such as elevator or bus and calculate in Fare Collection System, compared to existing technology, degree of accuracy improves greatly, makes native system more intelligent, advantageously in self-service toll administration, embodies outstanding technical progress.

Claims (23)

1. motion estimate method, is characterized in that, described recognition methods comprises the steps:

The first step, combines inter-frame difference with rim detection, carries out moving Object Segmentation to image sequence;

Second step, based on the moving target parameter rapid extraction of thick sampling;

3rd step, based on the motion estimate of neural network.

2. motion estimate method according to claim 1, is characterized in that, described inter-frame difference is interframe second order difference, and the bianry image of gained is $L_{i,j}\left(t_n\right)=\left\{\begin{array}{c}255,|F_{(i.j)}\left(t_{n+1}\right)-2F_{i,j}\left(t_n\right)+F_{i,j}\left(t_{n-1}\right)|\&GreaterEqual;\mathrm{\&theta;}\\ 0,|F_{(i.j)}\left(t_{n+1}\right)-2F_{i,j}\left(t_n\right)+F_{i,j}\left(t_{n-1}\right)|<\mathrm{\&theta;}\end{array}\right\},$ Wherein F _i,j(t _n-1), F _i,j(t _n), F ( _i,j) (t _n+1) being respectively three two field pictures adjacent in described image sequence, θ is threshold value.

3. motion estimate method according to claim 1, is characterized in that, described moving target parameter rapid extraction comprises the steps:

A) for moving target b _i(i=1 ..., M) and (M is moving target number in image), extract the boundary rectangle being parallel to direction of primary motion, be designated as R (b _i) (i=1 ..., M);

B) slightly sample in R (bi) inside, and sampling rectangular area is required to determine k value by parameters precision, thick sample area is wherein r (b _i) _x, R (b _i) _ythe projection of moving target boundary rectangle on vertical direction of primary motion and direction of primary motion;

C) according to initial frame overall intensity, Otsu algorithm determination segmentation threshold θ is adopted;

D) according to described segmentation threshold θ to boundary rectangle R (b _i) in rectangular block carry out binary segmentation, $D\left(b_i\right)=\left\{\begin{array}{c}255,\mathrm{Count}\left(S_{m,n}\right)\&GreaterEqual;\mathrm{\&theta;}\\ 0,\mathrm{Count}\left(S_{m,n}\right)<\mathrm{\&theta;}\end{array}\right.,$ Wherein S _m,n∈ R (b _i);

E) binary segmentation image D (b in rectangular block _i) carry out characteristic parameter extraction, as former moving target b _i(i=1 ..., M) characteristic parameter be similar to.

4. motion estimate method according to claim 1, is characterized in that, the described motion estimate based on neural network comprises the steps:

A) neural network design;

B) sample data process and training set;

C) systematic training and identification process.

5. recognition methods according to claim 4, is characterized in that, described neural network selects 3 layers of BP neural network, comprises input layer, hidden layer and output layer, and transport function adopts sigmoid function:

6. motion estimate method according to claim 5, is characterized in that, identification proper vector is defined as: d=(r, e, f, d, w, h), wherein:

A) e is the longest linear component and object video projection ratio in moving target profile,

Wherein h is the longest straight length of moving target profile utilizing Hough transformation fast fetching to obtain, D _x, D _ybe respectively the projection of moving region on x, y direction;

B) f is form parameter, $f=\frac{\min (D_X,D_Y)}{\max (D_X,D_Y)};$

C) r is circle, wherein:

μ _tfor the mean distance from center of gravity to point, ${\mathrm{\&mu;}}_T=\frac{1}{K}{\mathrm{\&Sigma;}}_{K=D}^{K-1}\left|\right|(x_k,y_k)-(\stackrel{\&OverBar;}{x},\stackrel{\&OverBar;}{y})\left|\right|,$

σ _tfor the mean square deviation of the distance from regional barycenter to point,

${\mathrm{\&sigma;}}_r=\frac{1}{K}{\mathrm{\&Sigma;}}_{K=D}^{K-1}{\left[\right||(x_k,y_k)-(\stackrel{\&OverBar;}{x},\stackrel{\&OverBar;}{y})||-{\mathrm{\&mu;}}_r]}^2;$

D) d is dutycycle, i.e. moving target area rectangular area ratio extraneous with direction of primary motion, wherein:

$d=\frac{a}{D_x{D}_y},$ a=∑ _(x,y)∈R1；

E) w, h are width and the height of target.

7. motion estimate method according to claim 6, is characterized in that, according to the dimension of described identification proper vector, determines that described neural network input layer nodes is 6.

8. motion estimate method according to claim 5, is characterized in that, described neural network node in hidden layer is m, wherein l is described neural network output layer nodes, and α is the constant between 1-10.

9. motion estimate method according to claim 8, is characterized in that, described neural network node in hidden layer is 9.

10. motion estimate method according to claim 5, is characterized in that, described neural network output layer nodes is 5.

11. motion estimate methods according to claim 4, is characterized in that, described sample data process and training set mainly comprise the steps:

A) change of scale, converts each component of proper vector, it is all mapped in [0,1] region;

B) training set structure.

12. motion estimate methods according to claim 11, is characterized in that, by proper vector s _i=(r _i, e _i, f _i, d _i, w _i, h _i) be transformed to proper vector S _i'=(r _i', e _i', f _i', d _i', w _i', h _i'), i is proper vector number, and each component wherein between proper vector has following relation:

a）r _i′=r _i,e _i′=e _i,f _i′=f _i,d _i′=d _i；

b） $w_i^{\&prime;}=\frac{w_i-w_{\min }}{w_{\max }-w_{\min }},h_i^{\&prime;}=\frac{h_i-h_{\min }}{h_{\max }-h_{\min }},$ Wherein

w _max=max(w _i),h _max=max(h _i),w _min=min(w _i);h _min=min(h _i)，

In real transform, w _max, h _maxget the wide and high of moving region respectively, get wherein S _minfor the area of detection lower limit of pedestrian in motion detection.

13. motion estimate methods according to claim 11, is characterized in that, set up proper vector d to moving target each in frame _i=(r _i, e _i, f _i, d _i, w _i, h _i), and manual sort is carried out to moving target, the proper vector after finally utilizing conversion and this moving target is carried out to the result c of manual sort _i, composition learning sample p _i'=(r _i', e _i', f _i', d _i', w _i', h _i', c _i), wherein specify, as sample p _i'=(r _i', e _i', f _i', d _i', w _i', h _i', c _i) when belonging to c class, c ∈ [0, M-1], the desirable output vector of described neural network is $o_i=\left\{\begin{array}{c}0,i\&NotEqual;c\\ 1,i=c\end{array}\right.,$ I=0 ..., M-1, M are moving target classification number, and C is classification belonging to sample.

14. motion estimate methods according to claim 4, it is characterized in that, described systematic training and identification process mainly comprise the steps:

A) neural metwork training, adopts variable step momentum method to carry out neural metwork training;

B) identification process, in target treatment, to the moving target meeting displacement of targets, direction constrain condition, utilizes neural network classifier to identify.

15. motion estimate methods according to claim 14, it is characterized in that, described identification process mainly comprises the steps:

A) neural network having completed study is loaded into;

B) for the moving target in target treatment, target identification proper vector d=(c, e, f, d, w, h) is set up;

C) proper vector d is converted successively, if the proper vector after conversion is T;

D) make T be the input vector of neural network, obtain the output valve of each output unit of neural network

If e) have then think that moving target belongs to classification C.

16., according to the arbitrary described motion estimate method of claim 1-15, is characterized in that, before carrying out described moving Object Segmentation to sequence image, first carry out pre-service to sequence image, specifically comprise filtering and noise reduction.

The 17. elevator intelligent charge systems with motion estimate function, comprising: message accounting, for obtaining and adding up effective charge information; Pick-up unit, for adding up the number of elevator actual bearer; Relatively treating apparatus, is connected with above-mentioned message accounting and pick-up unit, controls the running status of elevator according to the comparative result of described effective charge information and described actual bearer number; If described effective charge information is greater than or equal to actual bearer number, then controls described elevator and run; If described effective charge information is less than described actual bearer number, then control described elevator out of service; It is characterized in that: described pick-up unit comprises the identification module of photographing module and moving target.

The 18. elevator intelligent charge systems with motion estimate function according to claim 17, it is characterized in that, described photographing module can be equipped with single camera or multi-cam.

The 19. elevator intelligent charge systems with motion estimate function according to claim 17, it is characterized in that, the identification module of described moving target comprises moving Object Segmentation module, and inter-frame difference is combined with rim detection by it, carries out moving Object Segmentation to image sequence; Based on the moving target parameter rapid extraction module of thick sampling; Based on the motion estimate submodule of neural network.

The 20. elevator intelligent charge systems with motion estimate function according to claim 17, it is characterized in that, also comprise speech prompting device, be connected with the described treating apparatus that compares, when effective charge information is less than actual bearer number, the passenger not performing charging confirmation is reminded to carry out charging confirmation.

21. Intelligent billing devices with motion estimate function according to claim 17, it is characterized in that, also comprise panalarm, be connected with the described treating apparatus that compares, after the schedule time after alarm set sends prompting message, if elevator is off-duty still, then send warning message to keeper.

22. Intelligent billing devices with motion estimate function according to claim 17, is characterized in that, also comprise interactive correspondence equipment, in order to realize the interactive correspondence taken between people in keeper and elevator.

23. Intelligent billing devices with motion estimate function according to claim 17, is characterized in that, described message accounting can be arranged on entrance or the elevator interior of elevator.