CN103955699A - Method for detecting tumble event in real time based on surveillance videos - Google Patents

Abstract

The invention discloses a method for detecting a tumble event in real time based on surveillance videos. A plurality of cameras which face the same target area and have different shooting angles are installed in a detection scene and shoot the target area continuously. The method comprises the following steps that the cameras respectively shoot a video of the target area at the same time; foreground images, representing a target, of each image frame are extracted from the multiple videos shot by the cameras respectively in the same time period; the shape and position characteristics of the foreground images, in the corresponding images shot by the cameras, of the same object at the same moment are extracted, and target posture categories at the moments corresponding to all the image frames respectively are determined through an RVM categorizer; the target posture categories of all the image frames are input into a HMM evaluator as a target posture value sequence to obtain the posterior probability of change of the target posture categories, wherein the change of the target posture categories indicates that the target tumbles down; if the posterior probability is larger than a preset threshold value, it is determined that the tumble event happens.

Description

A kind ofly fall down in real time event detecting method based on monitor video

Technical field

The present invention relates to image model identification field, more specifically, relate to and fall down in real time detection method based on RVM and HMM.

Background technology

Have higher theory significance and practical value owing to falling down the research of detection, existing relevant achievement in research and product come out both at home and abroad.According to the method for falling down detection and adopting, detection technique can be divided three classes: Worn type instrument detection method, environmental device detection method, monitor video detection method, wherein first two technology is sensor-based method, and rear a kind of technology is the method based on image processing.

Wear instrumentation style detect in, user need to wear with oneself some instruments that sensor or other apparatuss are housed and help system acquisition user's action message and the movable information of health, system, by Information Monitoring being classified to detect the event of falling down, judges that by operating speed and acceleration transducer the unexpected static of human action detects the object of falling down to reach in document [1].

Worn type instrument detection method is simple and easy to execute, but the subject matter existing is: because sensor correlation parameter threshold value is to set according to the accurate relative position relation between instrument and wearer, once this relation destroyed (in fact often occurring), as strenuous exercise or wear off clothes, can produce a large amount of flase drops; In addition, due to needs, user wears instrument, will bring the larger inconvenience that do not accommodate for user.

Environment supervision instrument method mainly gathers the every data of the human body relevant to user by the multiple sensor being placed in environment, occurs by the data analysis event that judged whether to fall down.The people such as Alwan operative installations vibratory sensation receiver on floor in document [2] judges to be fallen down; Technical Solutions Australia system [3] is collected user's applied pressure information by " leaving the bed " alarm, floor foot pad alarm etc., and the data that gather by analysis are differentiated user's attitude.

Identical with Worn type instrument detection method, this kind of method is also easily subject to other interference in environment and produces flase drop; Although exempted user and wear the trouble of instrument, owing to having increased a large amount of sensors, the complicacy of system rises to some extent.

Monitor video detects, i.e. Computer Vision Detection occurs to have judged whether the event of falling down by the video data in real-time analysis monitoring environment.This kind of method can be further subdivided into again three kinds: (1) static detection.Conventionally the people after falling down can lie in ground the preceding paragraph time still, based on this hypothesis, Nait-Charif and McKenna[4] obtain user's movement locus, the unexpected termination of track while falling down to detect with the wide-angle lens that is erected at user's overhead.(2) body shape changes detection.In the process of falling down, can there is obvious variation in the person's of falling down body shape conventionally, as changed into and lie low from standing.Based on this principle, the people such as Ganapathy [5] end user's body boundary rectangle the ratio of width to height, boundary rectangle inclination angle be as posture feature, and judge the variation of body shape by the variation of analytical characteristic value, then detects the event of whether falling down and occur.(3) head movement/position probing.In the method, researcher is by human body head, and follows the tracks of the movement locus of head or the relative distance on location head and ground and detect the generation of the event of falling down.The people such as Shoaib [6] carry out human body head by ellipse fitting, and utilize the scene terrestrial information of simulation Gaussian distribution calculate head with respect to the distance on ground and judge whether to fall down.

In monitor video detection method, most research only adopts single motion feature or posture feature, therefore easily causes a large amount of flase drops.On the other hand, pertinent literature is not considered to process the event of falling down along video camera direction of illumination, and in this type of situation, the person's of falling down shape is similar with the person's of standing shape, is difficult to two kinds of attitudes to distinguish mutually near simple external appearance characteristic.

Above mentioned list of references list

[1]Almeida,O.,M.Zhang,and?J.C.Liu.Dynamic?fall?detection?and?pace?measurement?in?walking?sticks.IEEE?Joint?Workshop?on?High?Confidence?Medical?Devices,Software,and?Systems?and?Medical?Device?Plug-and-Play?Interoperability,2007.

[2]Alwan,M.,et?al.A?smart?and?passive?floor-vibration?based?fall?detector?for?elderly.IEEE2nd?Conf.on?Information?and?Communication?Technologies,2006.

[3]http://www.tecsol.com.au/

[4]Nait-Charif,H.and?S.J.McKenna.Activity?summarisation?and?fall?detection?in?a?supportive?home?environment.IEEE17th?Conf.on?Pattern?Recognition,2004.

[5]V.Vaidehi?et?al.Video?based?automatic?fall?detection?in?indoor?environment.IEEE?International?Conference?on?Recent?Trends?in?Information?Technology,2011.

[6]Shoaib,Muhammad,Dragon,R.,Ostermann,J.View-invariant?fall?detection?for?elderly?in?real?home?environment.4th?Pacific-Rim?Symposium?on?Image?and?Video?Technology,2010.

Summary of the invention

Present inventor considers the above-mentioned situation of prior art and has made the present invention.The present invention proposes a kind of detection method of falling down based on Multi-angle camera, can detect the event of falling down of different spaces direction, and there is processing capability in real time, possess compared with high practicability.For example, in indoor home environment, can detect in time the contingent event of falling down of stay alone old man, patient etc. (object of observation), alleviate to a great extent the injury that the event of falling down is brought.

Conventionally, in the time falling down, can there is variation by a relatively large margin in the person's of falling down attitude, based on this principle, the different attitudes in the process of falling down are divided into four classes by the present invention, extracted outward appearance, the scene characteristic of target and utilized Method Using Relevance Vector Machine (RVM) to carry out rapid posture identification to moving target by the camera video image of two different angles.Utilize Hidden Markov Model (HMM) (HMM) to change and carry out modeling attitude in the process of falling down, and remove to assess each section of motion process in monitor video with this model, thereby judged whether to fall down event generation.

According to embodiments of the invention, provide a kind of and fallen down in real time event detecting method based on monitor video, wherein, multiple video cameras of be provided with towards same target area in detection scene, shooting angle is different, the continuous photographic subjects of described multiple video camera region, said method comprising the steps of: step 1, described multiple video cameras one section of video in photographic subjects region simultaneously; Multiple videos of step 2, same period of taking separately from described multiple video cameras, extract respectively the foreground area of the representative target of each frame picture of video; Shape separately and the position feature of the described foreground area of step 3, the same target of extraction synchronization in the picture of being taken by described multiple video cameras, and use RVM sorter, determine the targeted attitude classification in the moment that each frame picture is corresponding; Step 4, be input to the targeted attitude classification of each obtained frame picture as targeted attitude value sequence HMM assessment-, obtain the posterior probability that targeted attitude classification changes, wherein, described targeted attitude classification change procedure indicating target is fallen down the generation of event; And if the described posterior probability of step 5 is greater than predetermined threshold, determine that target falls down the generation of event.

The present invention has adopted the mode of RVM and HMM combination to carry out video pictures pattern-recognition innovatively, not only can from video, identify the object of observation attitude of any time, can also identify the attitude change procedure of appointing in a period of time, like this, for the situation that attitude variation occurs within a period of time as falling down, can detect in real time.

The present invention is mainly used at home in monitor video scene, monitor also and alarm, thereby effectively ensure monitored person's personal safety for (improper droping to the ground) event of falling down that may occur monitor video.Beneficial effect mainly contains: (1) carries out round-the-clock Real-Time Monitoring to solitary empty nest old man, behavior and state to the elderly are analyzed, automatically filter out garbage, and old man's event of falling down that may occur is made to quick judgement warning, to succour in time, fundamentally ensure old solitary people safety.(2) patient of needs supervision is carried out to the analysis of condition, in the time falling down, can report to the police from trend operator on duty, prompting medical personnel process in time.Can reduce medical personnel's work load on the one hand, on the other hand also for patient's timely rescue provides the valuable time.

Brief description of the drawings

Fig. 1 is the poor schematic diagram of angle that the external oval drift angle of human body in two orthogonal shooting directions is according to an embodiment of the invention shown;

Fig. 2 illustrates the schematic diagram of the scene information of human body place background according to an embodiment of the invention;

Fig. 3 is the structural representation that 3 RVM sorters that train are according to an embodiment of the invention shown;

Fig. 4 is the figure that illustrates that in one section of video according to an embodiment of the invention, under training pattern, the log posterior probability log (P (O| λ)) of attitude sequence changes with frame number.

Embodiment

Below, by reference to the accompanying drawings the enforcement of technical scheme is described in further detail.

First, sketch principle of the present invention.

According to embodiments of the invention, the detection method of falling down in real time based on HMM and RVM mainly comprised the following steps in the model training stage: 1) feature extraction, is used for extracting in the training video frame of the video camera of two different angles multiple features that the attitude of reflection human body changes; 2) attitude classification, utilizes the above-mentioned features training of extracting to go out sorter, and obtains the attitude classification of each training video frame by sorter; 3) utilize Hidden Markov Model (HMM) (HMM) to change and carry out modeling (generating HMM model) the attitude in the process of falling down.

According to embodiments of the invention, the detection method of falling down in real time based on RVM and HMM mainly comprised the following steps in the event detection stage: 1) feature extraction, is used for extracting in the test video frame of the video camera of two different angles multiple features that the attitude of reflection human body changes; 2) attitude classification, utilizes the sorter training to obtain the attitude classification of each test video frame; 3) utilize the HMM model generating in the above-mentioned training stage, each section of motion process (process that attitude classification changes) in assessment monitor video, thus the event that judged whether to fall down occurs.

According to order above, the specific implementation process of falling down in real time detection method based on RVM and HMM of the present invention is described respectively below.Those skilled in the art will appreciate that following some step/operation is present in training stage and test phase simultaneously, for brevity, do not carry out repeat specification.

1. feature extraction

The process of falling down is decomposed, and the attitude change procedure of human body can be summed up as stands-tilts-lie (to ground).Based on this, adopt the feature of human geometry's outward appearance, scene information composition as the input of RVM sorter, carry out the judgement of attitude, the attitude of human body in the video of house roughly can be divided into 4 classes:

1) stand;

2) tilt;

3) lie on (only for ground);

4) other, comprise seat, squat, lie bed first-class.

Above-mentioned classification is only example, also those skilled in the art will appreciate that and can according to actual needs, human body attitude be divided into the classification of the arbitrary number different from above-mentioned 4 classes.

The human geometry's external appearance characteristic adopting has:

1) the ratio of width to height of human body boundary rectangle.For stance, this ratio is less, and for lateral attitude, its boundary rectangle shape approaches square, and the ratio of width to height is close to 1;

2) angle of the external oval drift angle of human body in two shooting directions is poor.In order to analyze human body attitude from multiple camera angle, need to consider the angle of ellipse fitting.Due to the perspective transform of video camera, there is certain information dropout when three-dimensional scenic is mapped in two dimensional image, many attitudes are not easily distinguishable, as stand, along lying of camera direction etc.Contour and the orthogonal video camera of two sight lines can be adopted, message complementary sense can be effectively carried out.For stance, because target is perpendicular to ground level, therefore in two camera views, the major axis of the external ellipse of human body and transverse axis angle are all about 90 °, therefore the difference of two angles is about 0 °; For lateral attitude, target and ground level are in a certain angle, and poor (absolute value) of the angle angle in two shooting directions, within the scope of 0 °～90 °, is generally significantly higher than this differential seat angle (approximately 0 °) under stance; For lying posture (only for ground), because target is parallel to ground level, therefore the differential seat angle (absolute value) of the angle (considering positive and negative) in two vertical shooting directions is about 90 °.Above-mentioned principle can be explained by Fig. 1 and (be followed successively by from top to bottom and stand, tilt, lie; First classifies actual scene schematic diagram as, and the two or three row are respectively two video camera shooting picture).Contour and the vertical video camera pendulum position of above-mentioned two sight lines is only example, those skilled in the art will appreciate that in fact can also there be alternate manner the pendulum position of two video cameras, as long as the variation of above-mentioned angle under different attitudes can present certain rule.Certainly, also can use more video camera, thereby obtain more accurate, sophisticated category result.

The scene information feature adopting has:

1) the scene information histogram of human body place background.According to house video properties, manually scene areas is carried out to artificial mark in advance, main mark is bed/sofa/chair region, wall, ground, as shown in Figure 2 (wherein grey represents wall, and black represents bed/sofa/chair, and white represents ground).By three regions, with different value representations, the scene information of statistics human body region, calculates three kinds of gray-scale value proportions, the scene information histogram of a 3bin of composition.

For original video, utilize foreground segmentation algorithm to extract moving region (can referring to document [7]), for above three kinds of features, all from above-mentioned two camera pictures, extract simultaneously, form so altogether 2 × 1+1+3 × 2=9 dimensional feature vector.,, for moment corresponding to each frame of video of monitor video, all extract above-mentioned 9 dimensional feature vectors.

2. attitude classification

Next, can carry out attitude classification.Attitude sorter adopt RVM(for example, can use disclosed RVM in document [8], because of its test speed very fast).Because RVM is mainly used in the situation of 2 classification, therefore need to train multiple 2 sorters, to successively classify.Analyze above-mentioned scene information feature and the human appearance feature extracted, due to this two category features difference, therefore adopt the taxonomic structure of decision tree that this two category feature is separately considered, each selected part feature is done to classify, successively judgement.Choose successively scene histogram, boundary rectangle the ratio of width to height and three kinds of features of external oval differential seat angle, train 32 sorters, the taxonomic structure obtaining is illustrated in fig. 3 shown below.Particularly, for example, first sorter RVM1 can be used to distinguish above-mentioned the 4th class attitude and other 3 class attitude, and second sorter RVM2 is used for distinguishing above-mentioned the 3rd class attitude and the 1st, 2 class attitudes, and the 3rd sorter RVM3 is used for distinguishing above-mentioned the 2nd class attitude and the 1st class attitude.

Like this, above-mentioned 9 dimensional features that extract in each frame of taking separately for two video cameras, all send in these 3 sorters successively, obtain the result of attitude classification.According to the attitude of personnel in every frame, obtain corresponding attitude type number, thereby produce attitude sequence, this is the observation sequence in the HMM model of using in HMM assessment below, and wherein status number N is 4, has above-mentioned 4 kinds of possible output states.

The prediction assorting process of RVM sorter can be summarized as follows (specifically can list of references [8]):

1) the eigenmatrix X ∈ R of known participation training ^{n × m}, the new feature vector x * ∈ R that obtains in test sample book ^{1 × n}and the RVM model vector p ∈ R that obtains of training ^{m × 1}, wherein n is intrinsic dimensionality, m is the number of samples that participates in training;

2) utilize x* and X to calculate base vector b ∈ R ^{1 × m};

3) base vector and model are multiplied each other and obtain numerical value y=b*p, if y > 0.5 is predicted as positive class, otherwise be predicted as negative class.

The training process of described RVM sorter comprises:

1) select suitable kernel function, proper vector is mapped to higher dimensional space.Conventional several kernel functions comprise RBF kernel function, Laplace kernel function, and polynomial kernel functions etc., adopt RBF core in the present invention;

2) initialization RVM parameter;

3) from four kinds of attitudes (stand, lie ground, tilt, other) training sample extract posture feature, the feature composition characteristic matrix X of all samples, the attitude label composition of vector Y that all samples are corresponding;

4) according to bayesian criterion, with weight distribution and the distribution parameter of training characteristics and label iterative training sample optimum;

5) output RVM parameter, the model that training obtains.

3.HMM assessment

In each frame, the attitude classification of record object, in order to utilize HMM, targeted attitude is represented by discrete value, i.e. (0,1,2,3), like this in continuous a period of time, obtaining one group of length is the frame number of T(corresponding to video) targeted attitude value sequence, i.e. observation sequence O ₁o ₂... O _t.In the training stage, according to HMM problem concerning study, utilize and fall down the observation sequence O extracting in process ₁o ₂... O _tlearn, find that a group model parameter lambda={ B} makes P (O| λ) maximum for π, A, and this is the parameter that HMM falls down model.

The training process of described HMM model comprises:

1) collect the video of falling down of the multistage difference person of falling down different directions;

2) extract every section of feature of falling down in video, and utilize RVM to carry out attitude classification, in sliding window of a time, the attitude of each frame output is numbered as HMM observation sequence;

3) utilize Baum-Welch training algorithm based on many observation sequences training HMM model λ, traditional Baum-Welch algorithm steps following (can referring to document [9]):

3-1 is that model parameter is composed an initial value λ ₀;

3-2 utilizes forward-backward algorithm method (can referring to document [10]), calculates the posterior probability of observation sequence O under this model, i.e. P (O| λ ₀);

3-3 is based on observation sequence O and "current" model parameter, Renewal model parameter lambda, its more new formula be:

$\begin{array}{cc}{\stackrel{\‾}{a}}_{\mathrm{ij}}=\frac{\underset{t=1}{\overset{T-1}{\mathrm{\Σ}}}{\mathrm{\ξ}}_t(i,j)}{\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\γ}}_t\left(i\right)},1\≤i\≤N,1\≤j\≤N& {\stackrel{\‾}{b}}_j\left(k\right)=\frac{\underset{t=1,O_t=v_k}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\γ}}_t\left(j\right)}{\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\γ}}_t\left(j\right)},1\≤j\≤N,1\≤k\≤M\end{array}$

${\stackrel{\‾}{\mathrm{\π}}}_i={\mathrm{\γ}}_1\left(i\right),1\≤i\≤N,$

In formula, N is the number of hidden state, and its implication is implicit targeted attitude information (for example, falls down attitude, the appearance of standing, lateral attitude), gets empirical value 3 in this patent; M is the number of observed reading, is also four kinds of state numberings of RVM output; a _ijbe the transition probability between state, represent to be forwarded to by state i the probability of state j; b _j(k) be emission probability, while being illustrated in state j, export the probability that observed reading is k; π _iit is the probability distribution of original state; v _krepresent k observed reading; O _trepresent the observed reading in t moment.ξ _t(i, j) and γ _t(i) be the auxiliary variable that obtained by "current" model calculation of parameter (can referring to document [9]), T be training video frame length.

4, calculate the posterior probability P (O| λ) of observation sequence under new model

If 5 logP (O| λ)-logP is (O| λ ₀) < Delta(Delta is a very little number, conventionally gets 1e-6 left and right), illustrate to train to produce a desired effect, algorithm finishes, and exports current model λ, otherwise, make λ ₀=λ, turns back to the 3rd step and works on.

Because traditional training algorithm is only based on single observation sequence, the model universality training is not strong, introduces the training algorithm based on many observation sequences in this patent.

Consider a series of observation sequence O={O under same pattern ⁽¹⁾, O ⁽²⁾... O ^(K), be respectively single observation sequence (1≤k≤K), owing to being independent of each other between every section of sequence, therefore suppose that they are for mutually independently, based on this hypothesis, the more new formula of model parameter is changed to (can referring to document [9]):

$\begin{array}{cc}{\stackrel{\&OverBar;}{a}}_{\mathrm{mn}}=\frac{\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T_k-1}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_t^{\left(k\right)}(m,n)}{\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T_k-1}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_t^{\left(k\right)}\left(m\right)},1\&le;m\&le;N,1\&le;n\&le;N& {\stackrel{\&OverBar;}{b}}_n\left(m\right)=\frac{\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{t=1,O_t^{\left(k\right)}=v_m}{\overset{T_k}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_t^{\left(k\right)}\left(n\right)}{\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T_k}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_t^{\left(k\right)}\left(n\right)},1\&le;n\&le;N,1\&le;m\&le;M\end{array}$

${\stackrel{\&OverBar;}{\mathrm{\&pi;}}}_n=\frac{1}{K}{\mathrm{\&gamma;}}_1^{\left(k\right)}\left(n\right),1\&le;n\&le;N,$

The model parameter of HMM can be summarized by following table:

Parameter	Implication
		T	Observation sequence length
O＝o 1,o 2,…o T	The sequence of observations
		Q＝q 1q 2…q T	Hidden status switch
N	Hidden state number
		S＝{s 1,s 2,…s N}	The possible value set of hidden state
M	The possible value number of observed reading
		V＝{v 1,v 2,…v M}	The possible value set of observed reading
A＝{a ij|a ij＝P r(q t+1＝s j|q t＝s i)}	Hidden state transition probability matrix with time-independent
		B＝{b j(k)|b j(k)＝P r(v k|q t＝s j)}	Observed reading probability distribution under given hidden state
π＝{π i|π i＝P r(q 1＝s i)}	Initial hidden distributions

At test phase, according to HMM evaluation problem, utilize the observation sequence O=O extracting in existing parameter and test video ₁o ₂... O _t, calculate the probability P (O| λ) (can computing formula in list of references [10]) of this sequence, thereby determine whether to fall down.

Fig. 4 is the figure that in one section of video, under training pattern, the log posterior probability logP of attitude sequence (O| λ) changes with frame number, wherein corresponding six sections of different events of falling down between six sections of peak region, during falling down, the log posterior probability of attitude output sequence reaches extreme point, has stronger difference with the non-posture of falling down.

Above mentioned list of references

[7]Yuyang?Chen,Yanyun?Zhao,Anni?Cai,A?robust?moving?object?segmentation?algorithm?using?integrated?mask-based?background?maintenance,3rd?IEEE?International?Conference?on?Network?Infrastructure?and?Digital?Content?(IC-NIDC),2012.

[8]Michael?E.Tipping,Sparse?Bayesian?Learning?and?the?Relevance?Vector?Machine,Journal?of?Machine?Learning?Research1(2001)211-244.

[9]L.E.Baum,T.Petrie,G.Soules,and?N.Weiss, ^aA?Maximization?Technique?Occurring?in?the?Statistical?Analysis?of?Probabilistic?Functions?of?Markov?Chains, ^oAnnals?of?Math.Statistics,vol.41,no.1,pp.164-171,1970

[10]Xiaolin?Li,Parizeau,M.,Plamondon,Rejean,“Training?hidden?Markov?models?with?multiple?observations-a?combinatory?method”,in?IEEE?Transactions?on?Pattern?Analysis?and?Machine?Intelligence,vol.22,no.4,pp.371-377,April,2000.

Be limited to miscellaneous for fear of the description that makes this instructions, in description in this manual, may the processing such as omission, simplification, accommodation have been carried out to the part ins and outs that can obtain in above-mentioned list of references or other prior art data, this is understandable for a person skilled in the art, and this can not affect the open adequacy of this instructions.At this, above-mentioned list of references is herein incorporated by reference of text.

In sum, those skilled in the art will appreciate that the above embodiment of the present invention can be made various amendments, modification and be replaced, it all falls into the protection scope of the present invention limiting as claims.

Claims (6)

1. fall down in real time event detecting method based on monitor video for one kind, wherein, multiple video cameras of be provided with towards same target area in detection scene, shooting angle is different, the continuous photographic subjects of described multiple video cameras region, said method comprising the steps of:

Step 1, described multiple video cameras one section of video in photographic subjects region simultaneously;

Multiple videos of step 2, same period of taking separately from described multiple video cameras, extract respectively the foreground area of the representative target of each frame picture of video;

Shape separately and the position feature of the described foreground area of step 3, the same target of extraction synchronization in the picture of being taken by described multiple video cameras, and use RVM sorter, determine the targeted attitude classification in the moment that each frame picture is corresponding;

Step 4, be input to the targeted attitude classification of each obtained frame picture as targeted attitude value sequence HMM assessment-, obtain the posterior probability that targeted attitude classification changes, wherein, described targeted attitude classification change procedure indicating target is fallen down the generation of event; And

If the described posterior probability of step 5 is greater than predetermined threshold, definite target is fallen down the generation of event.

2. the event detecting method of falling down in real time according to claim 1, wherein,

Described targeted attitude classification comprises following four classifications: 1) stand; 2) tilt; 3) lie in ground; 4) other,

Described targeted attitude classification changes and represents following variation: stands → tilts → lie in ground,

Wherein, described multiple video cameras are orthogonal two video cameras in shooting visual angle, and, in the target area of taking at each video camera in advance, mark multiclass subregion,

Described shape and position feature comprise following 3 features: 1) the ratio of width to height of the boundary rectangle of target; 2) angle between major axis and the horizontal line of the external ellipse of the target in two video cameras shooting direction separately is poor; 3) position of target is in which class subregion.

3. the event detecting method of falling down in real time according to claim 2, wherein, described RVM sorter comprises 32 sorters, wherein, first 2 sorter is distinguished first three classification and the 4th classification in described four classifications, second 2 sorter distinguished the first two classification and the 3rd classification in described four classifications, and the 3rd 2 sorters are distinguished second classification and first classification in described four classifications.

4. the event detecting method of falling down in real time according to claim 3, wherein, the training process of described RVM sorter comprises the following steps:

Step 11, from belong to respectively the training sample video of described four classifications, extract described shape and position feature, composition characteristic matrix X, the described targeted attitude classification composition of vector Y that all training sample videos are corresponding;

Step 12, employing RBF kernel function, be mapped to higher dimensional space by described eigenmatrix X;

Step 13, according to bayesian criterion, solve optimum weight distribution and distribution parameter, as RVM parameter.

5. the event detecting method of falling down in real time according to claim 4, wherein, the training process of described HMM model comprises the following steps:

What step 21, input K section comprised the event of falling down falls down sample video;

Step 22, extract described shape and position feature from every section of each frame picture of falling down sample video, and use RVM sorter, determine the targeted attitude classification in the moment that each frame picture is corresponding, by targeted attitude classification numbering composition HMM observation sequence set O={O ⁽¹⁾, O ⁽²⁾... O ^(K);

Step 23, the utilization Baum-Welch training algorithm training based on many observation sequences obtains HMM model.

6. the event detecting method of falling down in real time according to claim 5, wherein, described step 23 comprises the following steps:

Step 23-1, give initial value λ for "current" model parameter lambda ₀, wherein λ={ π, A, B};

Step 23-2, utilize forward-backward algorithm method, the initial value λ calculating in "current" model parameter ₀under posterior probability P (the O| λ of observation sequence set O ₀);

Step 23-3, based on following formula, Renewal model parameter lambda:

$\begin{array}{cc}{\stackrel{\&OverBar;}{a}}_{\mathrm{ij}}=\frac{\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T_k-1}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_t^{\left(k\right)}(i,j)}{\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T_k-1}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_t^{\left(k\right)}\left(i\right)},1\&le;i\&le;N,1\&le;j\&le;N& {\stackrel{\&OverBar;}{b}}_j\left(i\right)=\frac{\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{t=1,O_t^{\left(k\right)}=v_i}{\overset{T_k}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_t^{\left(k\right)}\left(j\right)}{\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T_k}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_t^{\left(k\right)}\left(j\right)},1\&le;j\&le;N,1\&le;i\&le;M\end{array}$

${\stackrel{\&OverBar;}{\mathrm{\&pi;}}}_i=\frac{1}{K}{\mathrm{\&gamma;}}_1^{\left(k\right)}\left(i\right),1\&le;i\&le;N,$

Wherein, N is the number of hidden state; M is the number of observed reading kind, M=4; a _ijbe the transition probability between hidden state, represent to be forwarded to by hidden state i the probability of hidden state j; b _j(k) be emission probability, while being illustrated in state j, export the probability that observed reading is k; π _iit is the probability distribution of original state; v _krepresent k observed reading; O _trepresent the observed quantity in the t moment that frame of video picture is corresponding, ξ _t(i, j) and γ _t(i) auxiliary variable for being obtained by "current" model calculation of parameter, T is training video frame number,

The posterior probability P (O| λ) of observation sequence set O under 23-4, the "current" model parameter lambda of calculating after renewal;

If 23-5 is logP (O| λ)-logP (O| λ ₀) < Delta, export current model parameter λ, otherwise, λ made ₀=λ, turns back to step 23-3, and wherein, Delta is the constant of 1E-6 left and right,

Other model parameter of HMM model is as defined in following table:

Parameter Implication Q＝q ₁q ₂…q _T Hidden status switch S＝{s ₁,s ₂,…s _N} The possible value set of hidden state V＝{v ₁,v ₂,…v _M} The possible value set of observed reading A＝{a _ij|a _ij＝P _r(q _t+1＝s _j|q _t＝s _i)} Hidden state transition probability matrix with time-independent B＝{b _j(k)|b _j(k)＝P _r(v _k|q _t＝s _j)} Observed reading probability distribution under given hidden state π＝{π _i|π _i＝P _r(q ₁＝s _i)} Initial hidden distributions