CN105447458A - Large scale crowd video analysis system and method thereof - Google Patents

Abstract

The invention relates to a large scale crowd video analysis system and a method thereof. The system comprises a crowd density calculation module, a crowd prospect segmentation module, a crowd tracking module, a crowd state analysis module and an event determination module. The crowd density calculation module, the crowd prospect segmentation module and the crowd tracking module process video image data and then the crowd number, a crowd area, a crowd motion direction and a speed are acquired respectively. The crowd state analysis module carries out processing analysis based on the acquired crowd number, the crowd area, the crowd motion direction and the speed and sends an analysis result to the event determination module. The event determination module is used for determining whether a crowd event is abnormal. The crowd event means that the crowd reaching a certain scale carries out one group characteristic behavior in a monitoring area. By using the system and the method, monitoring personnel can complete intelligent trend prediction, characteristic event positioning and other tasks; and effective help is provided for prevention of emergency, suspicious clue tracking and the like.

Description

A kind of large-scale crowd video analytic system and method

Technical field

The disclosure relates to computer vision field, particularly a kind of large-scale crowd video analytic system and method.

Background technology

Crowd than being easier to intensive place, such as railway station, bus station, cinema, large supermarket etc., often occur to trample due to crowd the injures and deaths event caused.Although the existing video monitoring system of China has possessed scale to a certain degree, these video monitoring systems play an important role in response occurred events of public safety and public security guarantee in time, and various video monitoring system is widely used in various scene, but monitoring relies on manual monitoring mostly now, need badly and explore and study intelligent video monitoring for security protection, monitor staff is helped to complete the task such as intelligent trend prediction, characteristic event location, for accident prevention, suspicious clue to be traced etc. and provided effective help.

Summary of the invention

For above-mentioned subproblem, present disclose provides a kind of large-scale crowd video analytic system and method.

A kind of large-scale crowd video analytic system, described system comprises crowd density computing module, crowd's foreground segmentation module, crowd's tracking module, crowd state analysis module, event determination module; Wherein:

Described crowd density computing module covers crowd's quantity within the vision for obtaining certain two field picture in monitor video view data;

Described crowd's foreground segmentation module for obtain certain two field picture in monitor video view data cover prospect within the vision and background segment after, crowd region in prospect;

Described crowd's tracking module covers crowd movement direction within the vision and speed for obtaining certain two field picture in monitor video view data;

Described crowd state analysis module carries out Treatment Analysis based on the crowd's quantity obtained, crowd region, crowd movement direction and speed, and analysis result is sent into event determination module;

Described event determination module is for judging that whether crowd's event is abnormal;

Described crowd's event refers in guarded region, and certain population characteristic sexual behaviour appears in the crowd reaching certain scale.

For realizing a method for large-scale crowd video analytic system, described method comprises the steps:

The disclosure can help monitor staff to complete the task such as intelligent trend prediction, characteristic event location, for accident prevention, suspicious clue to be traced etc. and provided effective help.

Accompanying drawing explanation

About large-scale crowd video analytic system data processing flow diagram in Fig. 1 disclosure embodiment;

The schematic network structure of complete convolutional neural networks in an embodiment of Fig. 2 disclosure system;

The schematic network structure of complete convolutional neural networks in an embodiment of Fig. 3 disclosure method.

Embodiment

Basic embodiments provides a kind of large-scale crowd video analytic system at one, described system comprises crowd density computing module, crowd's foreground segmentation module, crowd's tracking module, crowd state analysis module, event determination module; Wherein:

Described crowd density computing module covers crowd's quantity within the vision for obtaining certain two field picture in monitor video view data;

Described crowd's foreground segmentation module for obtain certain two field picture in monitor video view data cover prospect within the vision and background segment after, crowd region in prospect;

Described crowd's tracking module covers crowd movement direction within the vision and speed for obtaining certain two field picture in monitor video view data;

Described crowd state analysis module carries out Treatment Analysis based on the crowd's quantity obtained, crowd region, crowd movement direction and speed, and analysis result is sent into event determination module;

Described event determination module is for judging that whether crowd's event is abnormal;

Described crowd's event refers in guarded region, and certain population characteristic sexual behaviour appears in the crowd reaching certain scale.

In this embodiment, described system can provide Data support for monitor staff completes the task such as intelligent trend prediction, characteristic event location, for accident prevention, suspicious clue to be traced etc. and provided effective help.Described view data can be directly the video interception that a frame is complete, also can be the view data that processed one-tenth facilitates corresponding module to be further processed, can also be carry out the view data after compressing process for convenience of transmission, then in the module receiving this view data, corresponding decompression is carried out, etc.In crowd state analysis module, its Treatment Analysis of carrying out is mainly crowd's event of what type based on the crowd's quantity obtained, crowd region, crowd movement direction and the speed crowd's event analyzed residing for the crowd determining guarded region, and then the conclusion of this crowd's event is sent to event determination module, judge whether this crowd's event exists exception by event determination module according to the judgment rule formulated.The type of described crowd's event may be different to the classification difference of crowd's event because of different system.

In one embodiment, provide the preferred computation model that crowd density computing module calculates crowd's quantity, that is: described crowd density computing module adopts degree of depth convolutional neural networks (DCNN, DeepConvolutionalNeuralNetwork) model carry out crowd density estimation and then obtain crowd's quantity.

The learning objective of described degree of depth convolutional neural networks model is for mapping F:X → D, and wherein X represents the pixel characteristic of image in training set, and D represents crowd density image.This model has following features: on study crowd characteristic, have good validity and robustness, do not need extra mark work, and to split with display foreground be independently, therefore, it is possible to obtain more reasonably result of calculation.Use this model at crowd density computing module, reliable Data support can be provided for follow-up analysis, and then when there is abnormal crowd's event, can offer help for the decision-making of supervisor.

In one embodiment, provide the preferred parted pattern of crowd's foreground segmentation module, that is: described crowd's foreground segmentation module adopts complete convolutional neural networks (FCNN, FullyConvolutionalNeuralNetwork) model to calculate crowd region.

The convolutional neural networks that complete convolutional neural networks is relatively traditional, can be applicable to the situation of picture dimensional variation, uses more flexible.

In one embodiment, provide the preferred calculating crowd movement direction of crowd's tracking module and the method for speed, that is: described crowd's tracking module adopts KLT algorithm (Kanade-Lucas-Tomasi) to obtain crowd movement direction and speed.

Although it is consuming time that the problem adopting optical flow method to carry out moving object segmentation is mainly that most of optical flow method calculates, real-time and practicality are all poor, but the advantage of optical flow method is that light stream not only carries the movable information of moving object, but also the abundant information carried about scenery three-dimensional structure, it when not knowing any information of scene, can detect Moving Objects.

In one embodiment, crowd density image mapped expression formula is embodied as a core cell by described crowd density computing module, that is: described crowd density computing module comprises crowd density image mapped unit, and the expression formula of described crowd density image mapped unit is:

$D_i\left(p\right)=\underset{P\∈P_i}{\Σ}\frac{1}{\left|\right|Z\left|\right|}\left(N_h\right(p;P_h,{\mathrm{\σ}}_h)+N_b(p;P_b,\Σ\left)\right)$

Wherein:

D _ip () is the some p around i-th people in density image D;

Z is normalized parameter;

P _ibe the people of i-th mark;

N _hit is the normalized 2D gaussian kernel function as head model;

P _hfor the position of head;

σ _hfor N _hvariance;

N _bit is the two-variable normal distribution as body model;

P _bfor the position of the person;

∑ is N _bcovariance matrix.

Above-mentioned mapping can ensure that all density values are equal crowd's quantity in former figure on the whole in a density mappings.

In one embodiment, the population characteristic sexual behaviour of the crowd of generation event is divided, the crowd behaviour that further clear and definite disclosure system will be monitored, that is: described population characteristic sexual behaviour comprises that the crowd is dense, crowd massing, crowd is detained, crowd runs and crowd drives in the wrong direction;

It is described that the crowd is dense refers to the parameter threshold number mean value in T1 guarded region second in past being exceeded to setting;

The described crowd massing number referring to stop in guarded region in the crowd region that flocks together exceedes and arranges threshold value;

Described crowd is detained and refers to follow the tracks of the crowd characteristic point in monitoring image, the move distance of described crowd characteristic point in time T2 lower than the threshold value arranged;

Described crowd runs and refers to that the quantity that movement velocity in guarded region is greater than the people of threshold speed V exceedes amount threshold N1;

The quantity of the people moved along direction of driving in the wrong direction in the entry region that described crowd drives in the wrong direction in finger guarded region exceedes threshold value N2;

Wherein, T1, T2, V, N1 and N2 are the self-defining value that preset value maybe can be revised.

According to the division of above-mentioned population characteristic sexual behaviour, crowd's event that the disclosure will be monitored comprises that the crowd is dense, crowd massing, crowd is detained, crowd runs and crowd drives in the wrong direction, and then in systems in which in order to identify crowd's event, can be everyone group's event dispense event ID.

Further, by this embodiment, also imply and how to have judged crowd's event anomalies at the event judge module of system.

For the crowd is dense event, described event determination module is by carrying out statistics to number mean value in past T inner region second, if exceed the parameter threshold of setting, then in judging area, the crowd is dense; Wherein T is the self-defining value that preset value maybe can be revised.In one embodiment, described T is 10 seconds.

For crowd massing event, described event determination module, by adding up the number stopped in image in the crowd region that flocks together, arranges threshold value if exceeded, then generation abnormal aggregation in judging area.

Be detained event for crowd, described event determination module is by following the tracks of the crowd characteristic point in image; The explanation residence time that in a period of time, move distance is shorter is more of a specified duration, when the residence time exceedes the threshold value of setting, then has crowd to be detained in declare area.

To run event for crowd, described event determination module is by calculating the movement velocity of tracking characteristics point under scene, and statistics movement velocity is greater than the number of the unique point of threshold value T1, when feature point number is greater than threshold value T2, there occurs abnormal running under showing scene; Wherein T1, T2 are the self-defining value that preset value maybe can be revised.

To drive in the wrong direction event for crowd, described event determination module is by adding up along the number of the direction motion characteristics point that drives in the wrong direction in entry region, if it is N number of to exceed threshold value, then judging area occurs to drive in the wrong direction; Wherein N is the self-defining value that preset value maybe can be revised, and such as N is 10.

More preferably, described system determines the method whether crowd's event of monitoring continuously in one section of event is same crowd's event, that is: certain behavior in described population characteristic sexual behaviour is when continuous several times occurs, if adjacent twice interval greater than setting-up time, then this kind of behavior of adjacent twice appearance is considered to twice event, otherwise this kind of behavior of adjacent twice appearance is considered to an event.Such as, in the crowd is dense event, if event interval is greater than 60 seconds if adjacent twice the crowd is dense, then twice event is judged to be.In crowd massing event, if in phase overlay area twice the crowd is dense that if event interval is greater than 60 seconds, be then judged to be twice event.In crowd's delay event, if if the delay event occurred in identical retention areas is separated by 10 seconds, be then judged to be same delay event.Run in event crowd, if event interval is greater than 60 seconds if twice adjacent crowd runs, be then judged to be twice event.Drive in the wrong direction in event crowd, if event interval is greater than 60 seconds if adjacent twice crowd in identical retrograde region drives in the wrong direction, be then judged to be twice event.

In one embodiment, described system also comprises real-time monitoring module, and described real-time monitoring module is for reflecting current point in time, and in scene corresponding to the video of a certain road, the situation of crowd, shows the dense degree of crowd by illustrated mode.Wherein, the figure of described illustrated mode institute foundation can be based on the image in monitor video.

Preferably, the figure used in described real-time monitoring module comprises crowd and stops colored graph and Crowds Distribute thermodynamic chart; Described crowd stops colored graph according to residence time length, carries out Jet dyeing to the crowd in scene; Described Crowds Distribute thermodynamic chart, according to crowd's density, represents the real-time distribution situation of crowd in scene with the form of thermodynamic chart.

Further, described Jet dyeing is that the time length stopped according to personage in scene dyes to target, and maximal value shows as redness, and minimum value shows as blueness.Wherein: blue is the personage in mobile, yellow for there being the personage of short stay behavior, along with personage's residence time is more and more longer, its color also can more and more convergence be red.

This that is, blue is the personage in mobile, yellow for there being the personage of short stay behavior, and along with the residence time is more and more longer, its color also can more and more convergence be red.By this mode, monitor staff can find out crowd's residence time length of regional in scene at a glance, the abnormal crowd stopped of Timeliness coverage.

Further, the region that the color of described Crowds Distribute thermodynamic chart is redder shows that crowd density is herein higher; Color along with the reduction of crowd density by redness to blue gradual change.

In one embodiment, described system also comprises trend and finds module, and described trend finds that module locates by illustrated mode the region that scene servant clustering holds together.Wherein, the figure of described illustrated mode institute foundation can be based on the image in monitor video.A situation arises to find module monitor staff to be helped to understand the detailed event of a certain passage by trend, finds potential rule and aid decision making.After selected a certain passage, namely according to time, two, space dimension, to the event on this passage, a situation arises carries out analysiss displaying for system.Monitor staff is by page turning mode Switch Video channel selecting window.

Preferably, described trend finds that the figure used in module comprises crowd's quantity broken line graph and crowd is evenly distributed thermodynamic chart; Described crowd's quantity broken line graph represents the number of crowd within a period of time, comprises peak value number and average number; Supervisor also can select in 1 hour as required, in 1 day, in 7 days, equal time granularity is checked.Described crowd be evenly distributed thermodynamic chart with seclected time section for dimension, the situation that is evenly distributed of crowd is illustrated, according to different color regions, the region that people's clustering is disturbed easily occurs under can being easy to navigate to scene.

Further, the red area that described crowd is evenly distributed thermodynamic chart shows herein for region is held together in people's clustering; Color along with the reduction of number in a period of time by redness to blue gradual change.

In one embodiment, described system also comprises described system and also comprises event statistics module, described event statistics module in periodically temporal projection, occurs in periodically temporal rule with the event of checking based on a certain crowd's anomalous event nearest W week Frequency; Wherein W is the self-defining value that preset value maybe can be revised.Such as W is 4.

Preferably, described event statistics module comprises event query unit, described event query unit can in each monitor video event review and locate, as by select video channel, event type, generation time period, event can be inquired.

Preferably, the Query Result of described event query unit shows with the form of the sectional drawing of historical events in conjunction with the Word message of corresponding event.In the displaying result of an embodiment, except showing the sectional drawing of historical events, also key message can be shown, as time, gathering number etc. that event occurs.

Optionally, described Query Result at least comprises key frame sectional drawing, and described key frame obtains by carrying out tracking to the whole forming process of described event.In one embodiment, when the event selecting certain to inquire, the details of corresponding event can be opened in new window.Details are made a summary by key message, key frame sectional drawing forms, and wherein key frame carries out tracking by event judge module to whole event forming process and obtains.Monitor staff, by checking key frame, understands the whole generating process of initial, climax that event occurs, end, has basic understanding to the overall condition of event.

In one embodiment, described event determination module, when determining crowd's event anomalies, provides early warning.

Optionally, the form of described early warning comprises the array configuration of following a kind of or any various ways: static text, pattern or dynamically word, dynamic pattern, sound.Event that such as the crowd is dense, by calculating crowd saturated come reaction density situation.Crowd is wherein saturated can according to the ratio r=N/T of real-time scene number N and scene number capacity T: be divided into 5 grades: sparse (crowd's saturation degree is 0 ~ 30%), normal (crowd's saturation degree is 30 ~ 60%), higher (crowd's saturation degree is 60 ~ 80%), exceed standard (crowd's saturation degree is 80 ~ 120%), supersaturation (crowd's saturation degree is higher than 120%).When crowd's saturation degree lower than 60% time, be green; When crowd's saturation degree is yellow 60 ~ 80%, when crowd's saturation degree is shown in red higher than 80%, show saturation degree in real time simultaneously.Three Estate can be divided into according to the number of current aggregator: normal (below 20 people), merit attention (below 50 people), abnormal aggregation (more than 50 people), the crowd extremely event of running can be divided into 5 grades according to the severe degree of running for crowd massing event, except color tips, show grade of running simultaneously.Delay event can be divided into three different degree of meriting attention according to crowd from the time length of delay, shows hold-up time length in real time simultaneously.Retrograde event only has two state-events: occur and do not occur.Contribute to monitor staff like this and can take the corresponding measure intervention according to the order of severity of event.

In one embodiment, described system also comprises video acquisition and decoder module, and the video image for taking camera gathers and decodes.

In one embodiment, for helping monitor staff to find 5 class anomalous events fast, in time, the video image of described camera shooting represents with the form of real-time panel; Described panel shows content in the matrix form, and wherein each is classified as a video channel, and the picture of the first behavior camera shooting of matrix, remaining each behavior one class crowd event.Shown by the picture that camera is taken, contribute to the actual conditions intuitively seeing scene in real time.

In one embodiment, panel shows according to crowd's event class.Described crowd's event class divides and comprises panel and directly show saturation degree, panel and directly show crowd's incident duration, according to the number divided rank of group of participants's event with without associated ratings.Such as in the crowd is dense event, event class is that panel directly shows saturation degree, is shown as orange early warning more than 80%, is red alarm event more than 100%.In crowd massing event, event class is show in panel to assemble the duration.In crowd's delay event, event class is that Display panel is detained incident duration.Run in event crowd, event class is the relation according to feature point number on image and threshold value T2, will run and be divided into 5 grades, and described T2 is number.Drive in the wrong direction in event crowd, event class is without associated ratings, and when there is this crowd's event, panel directly identifies by the color can distinguished when occurring without crowd's event.

Optionally, described panel carries out split screen displaying by the video channel number arranging display in a screen automatically.This is convenient to monitor while multi-channel video, and simultaneously by this mode, it is extending transversely that system can effectively realize video channel number.In addition, optionally, also support to use by the full frame mode fully supporting whole display, facilitate monitor staff to find key message fast.

Optionally, the monitoring parameter that described video channel can carry out scene is arranged, and the monitoring parameter of described scene arranges the Initialize installation comprising scene and optimum configurations related with early warning; The Initialize installation of described scene comprises the overall situation response hot-zone specified in a certain video channel scene, and mode is in schematic picture, choose the region of a polygon as population analysis by mouse; Described be included in schematic picture with the related optimum configurations of early warning choose one or more polygonal region to carry out monitor and/or arrange a direction for responder to or for not responder to.

Optionally, the Initialize installation of described scene also comprises to be demarcated scene, and its mode is: choose at least 2 about height 170cm adults; If without the pedestrian satisfied condition in scene, then need to reselect the image for demarcating.In one embodiment, the adult choosing 2 height 170cm demarcates.In one embodiment, the adult choosing 2 height 175cm demarcates.

Optionally, the Initialize installation of described scene also comprises to be demarcated scene, and its mode is: according to the different scale size of people under the far and near scene of difference, estimate under same level coordinate, in actual scene, 100cm is mapped to the pixel count in image.

In one embodiment, by mouse mark rectangle will demarcate in the people that chooses just frame choosing, and according to the different scale size of people in scene distance, estimate under same level coordinate, in actual scene, one meter is mapped to the pixel count in image.

By demarcating, the parameters such as the real area of scene, degree of tilt, suggestion galleryful can be estimated.

In one embodiment, after demarcation, generate perspective matrix, described perspective matrix is mainly used to the size of the different far and near crowd of normalization scene.From the single-frame images of scene, select at least 2 adults randomly, if be less than 2, then change a uncalibrated image, and suppose that everyone average height is 175cm, use a linear regression model (LRM) to estimate a perspective matrix M.Pixel value M (p) in perspective matrix M to represent in the actual scene of pixel p position 1 meter of pixel quantity shared in the picture.

In one embodiment, video channel can be increased, and by inserting monitor video address at the interface of adding video channel, can complete the connection to new video passage.

Optionally, all existing video channels are shown with tabular form.

In one embodiment, described event determination module is also for the data write into Databasce by crowd's event.

Below in conjunction with accompanying drawing, the disclosure is set forth.

In one embodiment, as shown in Figure 1, provide a kind of large-scale crowd video analytic system, comprise camera and front end display screen within the system, by video acquisition and deciphering module, the image of camera collection is carried out gathering and decoding, and decoded image is passed to crowd density computing module, crowd's foreground segmentation module and crowd's tracking module respectively.Wherein, described crowd density computing module uses degree of depth convolutional neural networks model to carry out crowd density estimation, finally obtains crowd's quantity; Described crowd's foreground segmentation module uses complete convolutional network model to estimate the guarded region in image, finally obtains crowd region; Described crowd's tracking module uses KLT optical flow method to follow the tracks of, and finally obtains crowd movement direction and speed.The crowd's quantity obtained, crowd region and crowd movement direction and speed are delivered to crowd state analysis module, crowd's event analysis is carried out by based on the data received in this module, determine what crowd's event guarded region there occurs, and by this crowd's event transfer to event judge module, for determining whether this crowd's event there occurs exception further, if produce abnormal, then on the display screen of front end, provide front end early warning, early warning information write into Databasce is stored simultaneously; No matter but whether this crowd's event exception occurs, the data of crowd's event are all prepared against the use of the aspects such as inquiry by described event determination module stored in database.

In one embodiment, as shown in Figure 2, this figure once trains computation process or target identifying to the schematic network structure of the complete convolutional neural networks used in the system.In the drawings, ground floor P0 is former figure, and the second layer is the sub regions P0 ' in former figure P0, and be then 3 layers of convolutional layer, parameter 7*7 wherein, 5*5 represent the size of convolution kernel.The dimension of the binary file blob of the subgraph P0 ' obtained from P0 is 72 × 72 × 3, the dimension of the binary file blob obtained after first convolutional layer conv1 process is a result 72 × 72 × 32 for 3 dimensions, a result 36 × 36 × 32 for 3 dimensions is obtained through second convolutional layer conv2 process, a result 18 × 18 × 64 for 3 dimensions is obtained through the 3rd convolutional layer conv3 process, here pond layer has been lain in convolutional layer, can see that image size that convolutional layer obtains is all the half of front one deck.Two-layer full linking layer is entered after this layer, fc4 and fc5 is two-layer full linking layer, fc4 is the result of 1000 × 1, fc5 is the result of 400 × 1, fc6 is that regression target exports, fc7 is that classifying type target exports, last DenistyMap and Gobalnumber is the calibration result of original image, be manually provide mark and generated, fc6 (324 × 1) result that network generates can arrange by image is rows of the density image becoming (18 × 18), and the density_map counting loss concentrated with demarcation, for training network; Fc7 be 1 × 1 result and gobalnumber contrast and lost, also may be used for training network.Can switch between two of a training pattern convergence target as required, a convergence target loss loss tolerance, the computing method of described loss are as follows:

Result-the globanumber of loss=fc7

Another convergence target loses Eucldieanloss tolerance with Euler, and the computing method that described Euler loses Eucldieanloss are as follows:

Result-the densitymap of Eucldieanloss=fc6

When loss does not meet the demands, this process of iteration.

In one embodiment, provide the method for realizing said system, that is: a kind of large-scale crowd methods of video analyses, described method comprises the steps:

S100, calculating parameter crowd quantity, crowd region, crowd movement direction and speed:

Calculate certain two field picture in monitor video view data and cover crowd's quantity within the vision;

Calculate certain two field picture in monitor video view data cover prospect within the vision and background segment after, crowd region in prospect;

Calculate certain two field picture in monitor video view data and cover crowd movement direction within the vision and speed;

S200, analysis judgement crowd event category: carry out Treatment Analysis based on the crowd's quantity obtained, crowd region, crowd movement direction and speed, and judge crowd's event generic;

Described crowd's event refers in guarded region, and certain population characteristic sexual behaviour appears in the crowd reaching certain scale.

In this embodiment, the tasks such as intelligent trend prediction, characteristic event can be completed for monitor staff by described method and Data support is provided, for accident prevention, suspicious clue to be traced etc. and provided effective help.Described view data can be directly the video interception that a frame is complete, also can be the view data that processed one-tenth facilitates corresponding module to be further processed, can also be carry out the view data after compressing process for convenience of transmission, then after this view data of reception, corresponding decompression is carried out, etc.The analysis of described step S200 is mainly crowd's event of what type based on the crowd's quantity obtained, crowd region, crowd movement direction and the speed crowd's event analyzed residing for the crowd determining guarded region, and then carries out event type judgement based on the analysis conclusion of this crowd's event.When carrying out event type and judging, judge whether this crowd's event exists exception preferably by the judgment rule formulated.The type of described crowd's event may be different to the classification difference of crowd's event because of different system.

In one embodiment, provide the preferred computation model of calculating crowd quantity, that is: described crowd's quantity calculates further on the basis obtaining crowd density, described crowd density is estimated by adopting degree of depth convolutional neural networks (DCNN, DeepConvolutionalNeuralNetwork) model.

The learning objective of described degree of depth convolutional neural networks model is for mapping F:X → D, and wherein X represents the pixel characteristic of image in training set, and D represents crowd density image.This model has following features: on study crowd characteristic, have good validity and robustness, do not need extra mark work, and to split with display foreground be independently, therefore, it is possible to obtain more reasonably result of calculation.Use this model when crowd density calculates, reliable Data support can be provided for follow-up analysis, and then when there is abnormal crowd's event, can offer help for the decision-making of supervisor.

In one embodiment, provide one and carry out the preferred parted pattern of crowd's foreground segmentation, crowd region can be obtained by this model, that is: described crowd region calculates by adopting complete convolutional neural networks (FCNN, FullyConvolutionalNeuralNetwork) model.

The convolutional neural networks that complete convolutional neural networks is relatively traditional, can be applicable to the situation of picture dimensional variation, uses more flexible.

In one embodiment, the method preferably calculating crowd movement direction and speed is provided, that is:

Described crowd movement direction and speed are by adopting KLT algorithm (Kanade-Lucas-Tomasi) to calculate.

Although it is consuming time that the problem adopting optical flow method to carry out moving object segmentation is mainly that most of optical flow method calculates, real-time and practicality are all poor, but the advantage of optical flow method is that light stream not only carries the movable information of moving object, but also the abundant information carried about scenery three-dimensional structure, it when not knowing any information of scene, can detect Moving Objects.

Further, provide concrete crowd density computing method in one embodiment, that is: described crowd density is reflected by crowd density image, and the expression formula of described crowd density image is:

$D_i\left(p\right)=\underset{P\∈P_i}{\Σ}\frac{1}{\left|\right|Z\left|\right|}\left(N_h\right(p;P_h,{\mathrm{\σ}}_h)+N_b(p;P_b,\Σ\left)\right)$

Wherein:

D _ip () is the some p around i-th people in density image D;

Z is normalized parameter;

P _ibe the people of i-th mark;

N _hit is the normalized 2D gaussian kernel function as head model;

P _hfor the position of head;

σ _hfor N _hvariance;

N _bit is the two-variable normal distribution as body model;

P _bfor the position of the person;

∑ is N _bcovariance matrix.

Above-mentioned mapping can ensure that all density values are equal crowd's quantity in former figure on the whole in a density mappings.

In one embodiment, the population characteristic sexual behaviour of the crowd of generation event is divided, clearly use disclosure method to carry out the crowd behaviour of crowd regulation further, that is:

Described population characteristic sexual behaviour comprises that the crowd is dense, crowd massing, crowd is detained, crowd runs and crowd drives in the wrong direction;

It is described that the crowd is dense refers to the parameter threshold number mean value in T1 guarded region second in past being exceeded to setting;

The described crowd massing number referring to stop in guarded region in the crowd region that flocks together exceedes and arranges threshold value;

Described crowd is detained and refers to follow the tracks of the crowd characteristic point in monitoring image, the move distance of described crowd characteristic point in time T2 lower than the threshold value arranged;

Described crowd runs and refers to that the quantity that movement velocity in guarded region is greater than the people of threshold speed V exceedes amount threshold N1;

The quantity of the people moved along direction of driving in the wrong direction in the entry region that described crowd drives in the wrong direction in finger guarded region exceedes threshold value N2;

Wherein, T1, T2, V, N1 and N2 are the self-defining value that preset value maybe can be revised.

According to the division of above-mentioned population characteristic sexual behaviour, crowd's event that the disclosure will be monitored comprises that the crowd is dense, crowd massing, crowd is detained, crowd runs and crowd drives in the wrong direction, and then in order to identify crowd's event, can be that everyone group's event increases event id.

Further, by this embodiment, also imply and how judged crowd's event anomalies in disclosure method.

For the crowd is dense event, by carrying out statistics to number mean value in past T inner region second, if exceed the parameter threshold of setting, then in judging area, the crowd is dense; Wherein T is the self-defining value that preset value maybe can be revised.In one embodiment, described T is 10 seconds.

For crowd massing event, by adding up the number stopped in image in the crowd region that flocks together, threshold value is set if exceeded, then in judging area, abnormal aggregation occurs.

Event is detained for crowd, by following the tracks of the crowd characteristic point in image; The explanation residence time that in a period of time, move distance is shorter is more of a specified duration, when the residence time exceedes the threshold value of setting, then has crowd to be detained in declare area.

To run event for crowd, by calculating the movement velocity of tracking characteristics point under scene, statistics movement velocity is greater than the number of the unique point of threshold value T1, when feature point number is greater than threshold value T2, there occurs abnormal running under showing scene; Wherein T1, T2 are the self-defining value that preset value maybe can be revised.

To drive in the wrong direction event for crowd, by adding up along the number of the direction motion characteristics point that drives in the wrong direction in entry region, if it is N number of to exceed threshold value, then judging area occurs to drive in the wrong direction; Wherein N is the self-defining value that preset value maybe can be revised, and such as N is 10.

More preferably, described method further provides the judgment principle whether crowd's event of monitoring continuously in one section of event is same crowd's event, that is: certain behavior in described population characteristic sexual behaviour is when continuous several times occurs, if adjacent twice interval greater than setting-up time, then this kind of behavior of adjacent twice appearance is considered to twice event, otherwise this kind of behavior of adjacent twice appearance is considered to an event.Such as, in the crowd is dense event, if event interval is greater than 60 seconds if adjacent twice the crowd is dense, then twice event is judged to be.In crowd massing event, if in phase overlay area twice the crowd is dense that if event interval is greater than 60 seconds, be then judged to be twice event.In crowd's delay event, if if the delay event occurred in identical retention areas is separated by 10 seconds, be then judged to be same delay event.Run in event crowd, if event interval is greater than 60 seconds if twice adjacent crowd runs, be then judged to be twice event.Drive in the wrong direction in event crowd, if event interval is greater than 60 seconds if adjacent twice crowd in identical retrograde region drives in the wrong direction, be then judged to be twice event.

In one embodiment, in order to reflect in current point in time in real time, the situation of crowd in scene corresponding to the video of a certain road, after step S200, described method also comprises:

The dense degree of S300, in real time display crowd: the situation of crowd in scene corresponding to a certain road video being reflected current point in time by illustrated mode.

Wherein, the figure of described illustrated mode institute foundation can be based on the image in monitor video.

Preferably, the figure used in the dense degree of described real-time display crowd comprises crowd and stops colored graph and Crowds Distribute thermodynamic chart; Described crowd stops colored graph according to residence time length, carries out Jet dyeing to the crowd in scene; Described Crowds Distribute thermodynamic chart, according to crowd's density, represents the real-time distribution situation of crowd in scene with the form of thermodynamic chart.

Further, described Jet dyeing is that the time length stopped according to personage in scene dyes to target, and maximal value shows as redness, and minimum value shows as blueness.Wherein: blue is the personage in mobile, yellow for there being the personage of short stay behavior, along with personage's residence time is more and more longer, its color also can more and more convergence be red.

This that is, blue is the personage in mobile, yellow for there being the personage of short stay behavior, and along with the residence time is more and more longer, its color also can more and more convergence be red.By this mode, monitor staff can find out crowd's residence time length of regional in scene at a glance, the abnormal crowd stopped of Timeliness coverage.

Further, the region that the color of described Crowds Distribute thermodynamic chart is redder shows that crowd density is herein higher; Color along with the reduction of crowd density by redness to blue gradual change.

In one embodiment, conveniently locate the region that scene servant clustering is held together, after step S200, described method also comprises:

The region that S400, location scene servant clustering are held together: locate the region that scene servant clustering holds together by illustrated mode.

Wherein, the figure of described illustrated mode institute foundation can be based on the image in monitor video.A situation arises to find module monitor staff to be helped to understand the detailed event of a certain passage by trend, finds potential rule and aid decision making.After selected a certain passage, namely according to time, two, space dimension, to the event on this passage, a situation arises carries out analysiss displaying for system.Monitor staff is by page turning mode Switch Video channel selecting window.

Preferably, the figure used in the region that the scene servant clustering of described location is held together comprises crowd's quantity broken line graph and crowd is evenly distributed thermodynamic chart; Described crowd's quantity broken line graph represents the number of crowd within a period of time, comprises peak value number and average number; Supervisor also can select in 1 hour as required, in 1 day, in 7 days, equal time granularity is checked.Described crowd be evenly distributed thermodynamic chart with seclected time section for dimension, the situation that is evenly distributed of crowd is illustrated, according to different color regions, the region that people's clustering is disturbed easily occurs under can being easy to navigate to scene.

Further, the red area that described crowd is evenly distributed thermodynamic chart shows herein for region is held together in people's clustering; Color along with the reduction of number in a period of time by redness to blue gradual change.

In one embodiment, check to carry out statistics to various crowd's anomalous event, to find rule, after described step S200, described method also comprises:

S500, statistical phenomeon: add up in the temporal projection of periodicity based on a certain crowd's anomalous event nearest W week Frequency, occur in periodically temporal rule with the event of checking; Wherein W is the self-defining value that preset value maybe can be revised.Such as W is 4.

Preferably, described step S500, before statistics, also comprises step:

S50001, query event: in each monitor video event inquire about.By inquiry, can in each monitor video event review and locate, as by select video channel, event type, generation time period, event can be inquired.Here inquiry, except being applied to statistics, when method is embodied as system, also may be used for outwards providing search function, for retrieving all groups' event.

Preferably, described Query Result shows with the form of the sectional drawing of historical events in conjunction with the Word message of corresponding event.In the displaying result of an embodiment, except showing the sectional drawing of historical events, also key message can be shown, as time, gathering number etc. that event occurs.

Optionally, described Query Result at least comprises key frame sectional drawing, and described key frame obtains by carrying out tracking to the whole forming process of described event.In one embodiment, when the event selecting certain to inquire, the details of corresponding event can be opened in new window.Details are made a summary by key message, key frame sectional drawing forms, and wherein key frame carries out tracking by event judge module to whole event forming process and obtains.Monitor staff, by checking key frame, understands the whole generating process of initial, climax that event occurs, end, has basic understanding to the overall condition of event.

In one embodiment, described step S200 also comprises after judgement crowd event generic:

S201, provide early warning: corresponding early warning is provided to the crowd's event judged.

Optionally, the form of described early warning comprises the array configuration of following a kind of or any various ways: static text, pattern or dynamically word, dynamic pattern, sound.Event that such as the crowd is dense, by calculating crowd saturated come reaction density situation.Crowd is wherein saturated can according to the ratio r=N/T of real-time scene number N and scene number capacity T: be divided into 5 grades: sparse (crowd's saturation degree is 0 ~ 30%), normal (crowd's saturation degree is 30 ~ 60%), higher (crowd's saturation degree is 60 ~ 80%), exceed standard (crowd's saturation degree is 80 ~ 120%), supersaturation (crowd's saturation degree is higher than 120%).When crowd's saturation degree lower than 60% time, be green; When crowd's saturation degree is yellow 60 ~ 80%, when crowd's saturation degree is shown in red higher than 80%, show saturation degree in real time simultaneously.Three Estate can be divided into according to the number of current aggregator: normal (below 20 people), merit attention (below 50 people), abnormal aggregation (more than 50 people), the crowd extremely event of running can be divided into 5 grades according to the severe degree of running for crowd massing event, except color tips, show grade of running simultaneously.Delay event can be divided into three different degree of meriting attention according to crowd from the time length of delay, shows hold-up time length in real time simultaneously.Retrograde event only has two state-events: occur and do not occur.Contribute to monitor staff like this and can take the corresponding measure intervention according to the order of severity of event.

In one embodiment, before step S100, described method also comprises:

S000, to camera shooting video image gather and decode.

In one embodiment, for helping monitor staff to find 5 class anomalous events fast, in time, the video image of described camera shooting represents with the form of real-time panel; Described panel shows content in the matrix form, and wherein each is classified as a video channel, and the picture of the first behavior camera shooting of matrix, remaining each behavior one class crowd event.Shown by the picture that camera is taken, contribute to the actual conditions intuitively seeing scene in real time.

In one embodiment, panel shows according to crowd's event class.Described crowd's event class divides and comprises panel and directly show saturation degree, panel and directly show crowd's incident duration, according to the number divided rank of group of participants's event with without associated ratings.Such as in the crowd is dense event, event class is that panel directly shows saturation degree, is shown as orange early warning more than 80%, is red alarm event more than 100%.In crowd massing event, event class is show in panel to assemble the duration.In crowd's delay event, event class is that Display panel is detained incident duration.Run in event crowd, event class is the relation according to feature point number on image and threshold value T2, will run and be divided into 5 grades, and described T2 is number.Drive in the wrong direction in event crowd, event class is without associated ratings, and when there is this crowd's event, panel directly identifies by the color can distinguished when occurring without crowd's event.

Optionally, described panel carries out split screen displaying by the video channel number arranging display in a screen automatically.This is convenient to monitor while multi-channel video, and simultaneously by this mode, it is extending transversely that system can effectively realize video channel number.In addition, optionally, also support to use by the full frame mode fully supporting whole display, facilitate monitor staff to find key message fast.

Optionally, the monitoring parameter that described video channel can carry out scene is arranged, and the monitoring parameter of described scene arranges the Initialize installation comprising scene and optimum configurations related with early warning; The Initialize installation of described scene comprises the overall situation response hot-zone specified in a certain video channel scene, and mode is in schematic picture, choose the region of a polygon as population analysis by mouse; Described be included in schematic picture with the related optimum configurations of early warning choose one or more polygonal region to carry out monitor and/or arrange a direction for responder to or for not responder to.

Optionally, the Initialize installation of described scene also comprises to be demarcated scene, and its mode is: choose at least 2 about height 170cm adults; If without the pedestrian satisfied condition in scene, then need to reselect the image for demarcating.In one embodiment, the adult choosing 2 height 170cm demarcates.In one embodiment, the adult choosing 2 height 175cm demarcates.

Optionally, the Initialize installation of described scene also comprises to be demarcated scene, and its mode is: according to the different scale size of people under the far and near scene of difference, estimate under same level coordinate, in actual scene, 100cm is mapped to the pixel count in image.

In one embodiment, by mouse mark rectangle will demarcate in the people that chooses just frame choosing, and according to the different scale size of people in scene distance, estimate under same level coordinate, in actual scene, one meter is mapped to the pixel count in image.

By demarcating, the parameters such as the real area of scene, degree of tilt, suggestion galleryful can be estimated.

In one embodiment, after demarcation, generate perspective matrix, described perspective matrix is mainly used to the size of the different far and near crowd of normalization scene.From the single-frame images of scene, select at least 2 adults randomly, if be less than 2, then change a uncalibrated image, and suppose that everyone average height is 175cm, use a linear regression model (LRM) to estimate a perspective matrix M.Pixel value M (p) in perspective matrix M to represent in the actual scene of pixel p position 1 meter of pixel quantity shared in the picture.

In one embodiment, video channel can be increased, and by inserting monitor video address at the interface of adding video channel, can complete the connection to new video passage.

Optionally, all existing video channels are shown with tabular form.

In one embodiment, described step S200 also comprises after judgement crowd event generic: S202, by the data write into Databasce of crowd's event.The execution of this step can, before providing early warning, can, after providing early warning, can also be also parallel work-flow.

Below in conjunction with accompanying drawing, the disclosure is set forth.

In one embodiment, as shown in Figure 3, this figure once trains computation process or target identifying to the schematic network structure of the complete convolutional neural networks used in the process.In the drawings, ground floor P0 is former figure, and the second layer is the sub regions P0 ' in former figure P0, and be then 3 layers of convolutional layer, parameter 7*7 wherein, 5*5 represent the size of convolution kernel.The dimension of the binary file blob of the subgraph P0 ' obtained from P0 is 72 × 72 × 3, the dimension of the binary file blob obtained after first convolutional layer conv1 process is a result 72 × 72 × 32 for 3 dimensions, a result 36 × 36 × 32 for 3 dimensions is obtained through second convolutional layer conv2 process, a result 18 × 18 × 64 for 3 dimensions is obtained through the 3rd convolutional layer conv3 process, here pond layer has been lain in convolutional layer, can see that image size that convolutional layer obtains is all the half of front one deck.Two-layer full linking layer is entered after this layer, fc4 and fc5 is two-layer full linking layer, fc4 is the result of 1000 × 1, fc5 is the result of 400 × 1, fc6 is that regression target exports, fc7 is that classifying type target exports, last DenistyMap and Gobalnumber is the calibration result of original image, be manually provide mark and generated, fc6 (324 × 1) result that network generates can arrange by image is rows of the density image becoming (18 × 18), and the density_map counting loss concentrated with demarcation, for training network; Fc7 be 1 × 1 result and gobalnumber contrast and lost, also may be used for training network.Can switch between two of a training pattern convergence target as required, a convergence target loss loss tolerance, the computing method of described loss are as follows:

Result-the globanumber of loss=fc7

Another convergence target loses Eucldieanloss tolerance with Euler, and the computing method that described Euler loses Eucldieanloss are as follows:

Result-the densitymap of Eucldieanloss=fc6

When loss does not meet the demands, this process of iteration.

Be described in detail the disclosure above, apply specific case herein and set forth principle of the present disclosure and embodiment, the explanation of above embodiment just understands system of the present disclosure and core concept thereof for helping; Meanwhile, for those skilled in the art, according to thought of the present disclosure, all will change in specific embodiments and applications, in sum, this description should not be construed as restriction of the present disclosure.

Claims (44)

1. a large-scale crowd video analytic system, is characterized in that:

Described system comprises crowd density computing module, crowd's foreground segmentation module, crowd's tracking module, crowd state analysis module, event determination module; Wherein:

Described crowd density computing module covers crowd's quantity within the vision for obtaining certain two field picture in monitor video view data;

Described crowd's foreground segmentation module for obtain certain two field picture in monitor video view data cover prospect within the vision and background segment after, crowd region in prospect;

Described crowd's tracking module covers crowd movement direction within the vision and speed for obtaining certain two field picture in monitor video view data;

Described crowd state analysis module carries out Treatment Analysis based on the crowd's quantity obtained, crowd region, crowd movement direction and speed, and analysis result is sent into event determination module;

Described event determination module is for judging that whether crowd's event is abnormal;

Described crowd's event refers in guarded region, and certain population characteristic sexual behaviour appears in the crowd reaching certain scale.

2. system according to claim 1, is characterized in that:

Preferably, described crowd density computing module adopts degree of depth convolutional neural networks (DCNN, DeepConvolutionalNeuralNetwork) model carry out crowd density estimation and then obtain crowd's quantity.

3. system according to claim 1, is characterized in that:

Described crowd's foreground segmentation module adopts complete convolutional neural networks (FCNN, FullyConvolutionalNeuralNetwork) model to calculate crowd region.

4. system according to claim 1, is characterized in that:

Described crowd's tracking module adopts KLT algorithm (Kanade-Lucas-Tomasi) to obtain crowd movement direction and speed.

5. system according to claim 2, is characterized in that:

Described crowd density computing module comprises crowd density image mapped unit, and the expression formula of described crowd density image mapped unit is:

$D_i\left(p\right)=\underset{P\∈P_i}{\Σ}\frac{1}{\left|\right|Z\left|\right|}\left(N_h\right(p;P_n,{\mathrm{\σ}}_h\left)N_b\right(p;P_b,\mathrm{\Σ}\left)\right)$

Wherein:

D _ip () is the some p around i-th people in density image D;

Z is normalized parameter;

P _ibe the people of i-th mark;

N _hit is the normalized 2D gaussian kernel function as head model;

P _hfor the position of head;

σ _hfor N _hvariance;

N _bit is the two-variable normal distribution as body model;

P _bfor the position of the person;

∑ is N _bcovariance matrix.

6. system according to claim 1, is characterized in that:

Described population characteristic sexual behaviour comprises that the crowd is dense, crowd massing, crowd is detained, crowd runs and crowd drives in the wrong direction;

It is described that the crowd is dense refers to the parameter threshold number mean value in T1 guarded region second in past being exceeded to setting;

The described crowd massing number referring to stop in guarded region in the crowd region that flocks together exceedes and arranges threshold value;

Described crowd is detained and refers to follow the tracks of the crowd characteristic point in monitoring image, the move distance of described crowd characteristic point in time T2 lower than the threshold value arranged;

Described crowd runs and refers to that the quantity that movement velocity in guarded region is greater than the people of threshold speed V exceedes amount threshold N1;

The quantity of the people moved along direction of driving in the wrong direction in the entry region that described crowd drives in the wrong direction in finger guarded region exceedes threshold value N2;

Wherein, T1, T2, V, N1 and N2 are the self-defining value that preset value maybe can be revised.

7., according to the arbitrary described system of claim 1 ~ 6, it is characterized in that:

Described system also comprises real-time monitoring module, and described real-time monitoring module is for reflecting current point in time, and in scene corresponding to the video of a certain road, the situation of crowd, shows the dense degree of crowd by illustrated mode.

8., according to the arbitrary described system of claim 1 ~ 6, it is characterized in that:

Described system also comprises trend and finds module, and described trend finds that module locates by illustrated mode the region that scene servant clustering holds together.

9., according to the arbitrary described system of claim 1 ~ 6, it is characterized in that:

Described system also comprises event statistics module, and described event statistics module in periodically temporal projection, occurs in periodically temporal rule with the event of checking based on a certain crowd's anomalous event nearest W week Frequency;

Wherein W is the self-defining value that preset value maybe can be revised.

10., according to the arbitrary described system of claim 1 ~ 6, it is characterized in that:

Described event determination module, when determining crowd's event anomalies, provides early warning.

11., according to the arbitrary described system of claim 1 ~ 6, is characterized in that:

Described system also comprises video acquisition and decoder module, and the video image for taking camera gathers and decodes.

12., according to the arbitrary described system of claim 1 ~ 6, is characterized in that:

Described event determination module is also for the data write into Databasce by crowd's event.

13. systems according to claim 7, is characterized in that:

The figure used in described real-time monitoring module comprises crowd and stops colored graph and Crowds Distribute thermodynamic chart;

Described crowd stops colored graph according to residence time length, carries out Jet dyeing to the crowd in scene;

Described Crowds Distribute thermodynamic chart, according to crowd's density, represents the real-time distribution situation of crowd in scene with the form of thermodynamic chart.

14. systems according to claim 8, is characterized in that:

Described trend finds that the figure used in module comprises crowd's quantity broken line graph and crowd is evenly distributed thermodynamic chart;

Described crowd's quantity broken line graph represents the number of crowd within a period of time, comprises peak value number and average number;

Described crowd be evenly distributed thermodynamic chart with seclected time section for dimension, the situation that is evenly distributed of crowd is illustrated.

15. systems according to claim 9, is characterized in that:

Described event statistics module comprises event query unit, described event query unit can in each monitor video event review and locate.

16. systems according to claim 10, is characterized in that:

The form of described early warning comprises the array configuration of following a kind of or any various ways: static text, pattern or dynamically word, dynamic pattern, sound.

17. systems according to claim 11, is characterized in that:

The video image of described camera shooting represents with the form of real-time panel;

Described panel shows content in the matrix form, and wherein each is classified as a video channel, and the picture of the first behavior camera shooting of matrix, remaining each behavior one class crowd event.

18. systems according to claim 15, is characterized in that:

The Query Result of described event query unit at least comprises key frame sectional drawing, and described key frame obtains by carrying out tracking to the whole forming process of described event.

19. systems according to claim 17, is characterized in that:

Described panel carries out split screen displaying by the video channel number arranging display in a screen automatically.

20. systems according to claim 19, is characterized in that:

The monitoring parameter that described video channel can carry out scene is arranged, and the monitoring parameter of described scene arranges the Initialize installation comprising scene and optimum configurations related with early warning;

The Initialize installation of described scene comprises the overall situation response hot-zone specified in a certain video channel scene, and mode is in schematic picture, choose the region of a polygon as population analysis by mouse;

Described be included in schematic picture with the related optimum configurations of early warning choose one or more polygonal region to carry out monitor and/or arrange a direction for responder to or for not responder to.

21. systems according to claim 20, is characterized in that:

The Initialize installation of described scene also comprises to be demarcated scene, and its mode is: choose at least 2 about height 170cm adults; If without the pedestrian satisfied condition in scene, then need to reselect the image for demarcating.

22. systems according to claim 20, is characterized in that:

The Initialize installation of described scene also comprises to be demarcated scene, and its mode is: according to the different scale size of people under the far and near scene of difference, estimate under same level coordinate, in actual scene, 100cm is mapped to the pixel count in image.

23. 1 kinds of large-scale crowd methods of video analyses, is characterized in that, described method comprises the steps:

S100, calculating parameter crowd quantity, crowd region, crowd movement direction and speed:

Calculate certain two field picture in monitor video view data and cover crowd's quantity within the vision;

Calculate certain two field picture in monitor video view data cover prospect within the vision and background segment after, crowd region in prospect;

Calculate certain two field picture in monitor video view data and cover crowd movement direction within the vision and speed;

S200, analysis judgement crowd event category: carry out Treatment Analysis based on the crowd's quantity obtained, crowd region, crowd movement direction and speed, and judge crowd's event generic;

Described crowd's event refers in guarded region, and certain population characteristic sexual behaviour appears in the crowd reaching certain scale.

24. methods according to claim 23, is characterized in that:

Described crowd's quantity calculates further on the basis obtaining crowd density, and described crowd density is estimated by adopting degree of depth convolutional neural networks (DCNN, DeepConvolutionalNeuralNetwork) model.

25. methods according to claim 23, is characterized in that:

Described crowd region calculates by adopting complete convolutional neural networks (FCNN, FullyConvolutionalNeuralNetwork) model.

26. methods according to claim 23, is characterized in that:

Described crowd movement direction and speed are by adopting KLT algorithm (Kanade-Lucas-Tomasi) to calculate.

27. methods according to claim 24, is characterized in that:

Described crowd density is reflected by crowd density image, and the expression formula of described crowd density image is:

$D_i\left(p\right)=\underset{P\∈P_i}{\Σ}\frac{1}{\left|\right|Z\left|\right|}\left(N_h\right(p;P_n,{\mathrm{\σ}}_h\left)N_b\right(p;P_b,\mathrm{\Σ}\left)\right)$

Wherein:

D _ip () is the some p around i-th people in density image D;

Z is normalized parameter;

P _ibe the people of i-th mark;

N _hit is the normalized 2D gaussian kernel function as head model;

P _hfor the position of head;

σ _hfor N _hvariance;

N _bit is the two-variable normal distribution as body model;

P _bfor the position of the person;

∑ is N _bcovariance matrix.

28. methods according to claim 23, is characterized in that:

Described population characteristic sexual behaviour comprises that the crowd is dense, crowd massing, crowd is detained, crowd runs and crowd drives in the wrong direction;

It is described that the crowd is dense refers to the parameter threshold number mean value in T1 guarded region second in past being exceeded to setting;

The described crowd massing number referring to stop in guarded region in the crowd region that flocks together exceedes and arranges threshold value;

Described crowd is detained and refers to follow the tracks of the crowd characteristic point in monitoring image, the move distance of described crowd characteristic point in time T2 lower than the threshold value arranged;

Described crowd runs and refers to that the quantity that movement velocity in guarded region is greater than the people of threshold speed V exceedes amount threshold N1;

The quantity of the people moved along direction of driving in the wrong direction in the entry region that described crowd drives in the wrong direction in finger guarded region exceedes threshold value N2;

Wherein, T1, T2, V, N1 and N2 are the self-defining value that preset value maybe can be revised.

29. according to the arbitrary described method of claim 23 ~ 28, and it is characterized in that, after step S200, described method also comprises:

The dense degree of S300, in real time display crowd: the situation of crowd in scene corresponding to a certain road video being reflected current point in time by illustrated mode.

30. according to the arbitrary described method of claim 23 ~ 28, and it is characterized in that, after step S200, described method also comprises:

The region that S400, location scene servant clustering are held together: locate the region that scene servant clustering holds together by illustrated mode.

31. according to the arbitrary described method of claim 23 ~ 28, and it is characterized in that, after described step S200, described method also comprises:

S500, statistical phenomeon: add up in the temporal projection of periodicity based on a certain crowd's anomalous event nearest W week Frequency, occur in periodically temporal rule with the event of checking;

Wherein W is the self-defining value that preset value maybe can be revised.

32., according to the arbitrary described method of claim 23 ~ 28, is characterized in that, described step S200 also comprises after judgement crowd event generic:

S201, provide early warning: corresponding early warning is provided to the crowd's event judged.

33. according to the arbitrary described method of claim 23 ~ 28, and it is characterized in that, before step S100, described method also comprises:

S000, to camera shooting video image gather and decode.

34., according to the arbitrary described method of claim 23 ~ 28, is characterized in that, described step S200 also comprises after judgement crowd event generic:

S202, by the data write into Databasce of crowd's event.

35. methods according to claim 29, is characterized in that:

The figure used in the dense degree of described real-time display crowd comprises crowd and stops colored graph and Crowds Distribute thermodynamic chart;

Described crowd stops colored graph according to residence time length, carries out Jet dyeing to the crowd in scene;

Described Crowds Distribute thermodynamic chart, according to crowd's density, represents the real-time distribution situation of crowd in scene with the form of thermodynamic chart.

36. methods according to claim 30, is characterized in that:

The figure used in the region that the scene servant clustering of described location is held together comprises crowd's quantity broken line graph and crowd is evenly distributed thermodynamic chart;

Described crowd's quantity broken line graph represents the number of crowd within a period of time, comprises peak value number and average number;

Described crowd be evenly distributed thermodynamic chart with seclected time section for dimension, the situation that is evenly distributed of crowd is illustrated.

37. methods according to claim 31, is characterized in that, described step S500, before statistics, also comprises step:

S50001, query event: in each monitor video event inquire about.

38. methods according to claim 32, is characterized in that:

The form of described early warning comprises the array configuration of following a kind of or any various ways: static text, pattern or dynamically word, dynamic pattern, sound.

39. methods according to claim 33, is characterized in that:

The video image of described camera shooting represents with the form of real-time panel;

Described panel shows content in the matrix form, and wherein each is classified as a video channel, and the picture of the first behavior camera shooting of matrix, remaining each behavior one class crowd event.

40., according to method according to claim 37, is characterized in that:

The Query Result of described query event at least comprises key frame sectional drawing, and described key frame obtains by carrying out tracking to the whole forming process of described event.

41., according to method according to claim 39, is characterized in that:

Described panel carries out split screen displaying by the video channel number arranging display in a screen automatically.

42. methods according to claim 41, is characterized in that:

The monitoring parameter that described video channel can carry out scene is arranged, and the monitoring parameter of described scene arranges the Initialize installation comprising scene and optimum configurations related with early warning;

The Initialize installation of described scene comprises the overall situation response hot-zone specified in a certain video channel scene, and mode is in schematic picture, choose the region of a polygon as population analysis by mouse;

Described be included in schematic picture with the related optimum configurations of early warning choose one or more polygonal region to carry out monitor and/or arrange a direction for responder to or for not responder to.

43. methods according to claim 42, is characterized in that:

The Initialize installation of described scene also comprises to be demarcated scene, and its mode is: choose at least 2 about height 170cm adults; If without the pedestrian satisfied condition in scene, then need to reselect the image for demarcating.

44. methods according to claim 42, is characterized in that:

The Initialize installation of described scene also comprises to be demarcated scene, and its mode is: according to the different scale size of people under the far and near scene of difference, estimate under same level coordinate, in actual scene, 100cm is mapped to the pixel count in image.