CN103517042B - A kind of nursing house old man's hazardous act monitoring method - Google Patents

Abstract

The invention discloses a kind of nursing house old man's hazardous act monitoring method, be intended to overcome in prior art to exist that old man is followed the tracks of accuracy rate is low, behavior analysis weak effect, Activity recognition algorithm calls frequently, system load is big and monitoring effect too relies on the problems such as equipment performance, the method is characterized by use CCTV camera, video frequency collection card, personal PC machine, display, warning audio amplifier is in conjunction with the contour detecting program module in personal PC machine, tracer module, Activity recognition program module, based target is followed the tracks of the method that combines with Activity recognition and the active state in nursing house old man's daily life is monitored and danger and Deviant Behavior to old man is reported to the police.This method monitoring efficiency is high, high to old man's Activity recognition degree, operating cost is cheap, has good practicality and promotion prospect.<!--2-->

Description

A kind of nursing house old man's hazardous act monitoring method

Technical field

A kind of method that the present invention relates to technical field of computer vision, is specifically related to the monitoring method of a kind of nursing house old man's hazardous act.

Background technology

By inquiring about and retrieving, obtain the prior art close with technical field and Problems existing be as follows:

1. Chinese patent publication No. is CN102547216A, and date of publication is on July 4th, 2012, and denomination of invention is a kind of video frequency graphic monitoring system for major hazard source, speaks of a kind of video images detection system that monitoring system is applied to major hazard source in this invention.This patent, by organically being combined by different types of video camera, constitutes a complicated efficient system.But this system has two shortcomings, this system cannot be effectively applied in old man's hazardous act monitoring of nursing house: the investment of (1) this system is excessive, although the too much high performance video camera of quantity can the monitoring of all-dimensional multi-angle, but when nursing house limited fund, common CCTV camera is only optimum selection.(2) full manual monitoring, psychologic research shows, when employee is engaged in monotonous work for a long time, the aprosexia of employee, response speed also follows decline, directly results in employee and ignore the suspicious event occurred on screen, so that security system lost efficacy, lose the chance of very first time prevention crime, also the physical and mental health of employee can be produced totally unfavorable impact simultaneously, additionally, nursing house is for cost reasons, monitoring device generally by nurse on duty on behalf of supervision, nurse is when being busy with one's work, it is easy to ignore monitored picture, make monitoring thrashing.

2. Chinese patent publication No. is CN102764131A, date of publication is on November 7th, 2012, denomination of invention is lived to monitor for a kind of remotely old man and is taken care of system and method, this invention is spoken of a kind of comprehensive by various equipment, including: a series of equipment such as monitoring device, medical treatment sign device constitute the system of remotely monitoring old man life, this system has without in the advantage installing detection equipment with old man, old man is lived noiseless, and can arrange according to the Different Individual of old man or self study arranges different alarm threshold values, make this system more intelligent.This system side overweights monitoring old man's physical condition, for instance the biological informations such as heart rate, blood pressure, blood glucose, blood oxygen, and time of getting up, and takes medicine the rule of life information such as time, finds body abnormality in time, ahead of time diagnosis.But the real-time of this system is not high, too much depends on physical characteristic information, it is easy to ignoring old man's foudroyant disease, cost of investment is too high.

3. Chinese patent publication No. is CN101727570A, date of publication is on June 9th, 2010, denomination of invention is tracking, detecting and tracking process equipment and monitoring system, this invention is spoken of a kind of system utilizing tracking to be monitored, this invention is by predicting the possible position of target, and determine connected region, gained characteristic parameter and original characteristic parameter are compared to judge the position of target, and target characteristic is updated.This invention can solve the target characteristic variation issue being tracked in process to target preferably, and this invention can effectively be applied in monitoring system, improves the accuracy of target following.But the movement locus that this invention is only capable of by target is monitored, and cannot be distinguished by the target different behaviors under same movement locus.

Summary of the invention

Old man is followed the tracks of, in order to overcome, the shortcoming that accuracy rate is low, behavior analysis weak effect, Activity recognition algorithm call frequently, system load is big and monitoring effect too relies on the aspects such as equipment performance by the present invention, it is proposed to a kind of based target follows the tracks of old man's hazardous act monitoring method of the nursing house combined with Activity recognition.

In order to overcome the problems referred to above present invention to adopt the following technical scheme that realization: the step of described a kind of nursing house old man's hazardous act monitoring method is as follows:

Step one, use CCTV camera, video frequency collection card, personal PC machine obtain the video image of the old man entering monitoring region；

The video image obtained in step one is carried out contour detecting by the profile detection module in step 2, use personal PC machine, uses the number of pixels change of old man's contour area of frame difference method calculating gained to judge the state that old man is presently in.Use frame difference method, the old man entering monitoring region is carried out contour detecting, two continuous frames is monitored picture X₁、X₂Subtract each other according to the gray value of its correspondence, i.e. X₁(i,j)-X₂(i, the region of j) >=δ is the profile of old man, X₁(i,j)-X₂(i, j) < δ is the background area of old man, calculates the number of pixel in the connected region corresponding to profile, and the number of pixel is by when becoming many less, enter monitoring region for old man, now profile information need not be passed to tracking module；When the number of pixel does not continue to increase, can determine whether that old man has come into monitoring region, now the profile information of old man is passed to tracking module, this module was run once every 1 second.

The profile information of the old man to obtaining of the tracking module in step 3, personal PC machine is tracked, according to the defined rectangular shaped rim of the profile information of old man as tracking box, and extract local feature description's subcharacter of the ORB feature in tracking box and rotational invariance, and local feature description's subcharacter number of ORB characteristic number and rotational invariance, being input in the fitting function in tracking module by the center position coordinate of continuous several frames, old man's next frame position is predicted by the result drawn by fitting function；Detailed process is as follows:

According to the defined rectangular shaped rim of the profile information of old man as tracking box, and extract ORB feature (English full name: the OrientedFASTandRotatedBRIEF in tracking box, local feature description's subcharacter of rotational invariance) and ORB characteristic number (local feature description's subcharacter number of rotational invariance), the detailed process of the described ORB feature extracted in present frame tracking box is as follows:

First present frame is carried out piecemeal: individually extracted by the image in current tracking box, this picture is represented with I, if the matrix that image is m × n dimension, image is divided the fritter of several k × k pixels, so, whole picture is divided into the individual fritter of (m/k) × (n/k) altogether, takes the pixel value that meansigma methods is this fritter of the pixel of each fritter, (x y) is expressed as xth and arranges the average gray value of fritter corresponding to y row I.

Then the ORB characteristic point of present frame is selected: according to m_pq=∑_x,yx^py^pI (x, y), wherein p, q value respectively be 0 and 1, x ∈ [1 ..., m/k], y ∈ [1 ..., n/k] obtain the gradient θ=argtan (m of characteristic point₀₁,m₁₀). by all characteristic points according to the descending arrangement of θ value, from the individual coordinate points of (m/k) × (n/k), choose maximum for θ 8 as the characteristic point chosen.

Finally calculate the ORB characteristic number of present frame: choose 8 characteristic points be calculated according to following formula:

Wherein (x_i,y_i) respectively selected under step B 8 characteristic points, by formulaObtaining an eight-digit binary number numerical value, the form class of this numerical value is similar to the eight-digit binary number character string of 10011101, by f₁(I)…f₈(I) 10 system numbers corresponding to the character string of composition 8 × 8=64 position are as the ORB characteristic number extracted.

To be extracted after the ORB characteristic number of present frame, the center position coordinate of continuous several frames is input in the fitting function in tracking module, old man's next frame position is predicted by the result drawn by fitting function, it was predicted that detailed process be:

A. being provided with three fitting functions in the tracking module in personal PC machine, 1. fitting function is a beeline y=kx+b, and 2. fitting function is sine curve y=asin (cx-t)+kx+b, and 3. fitting function is parabola y=ax²+ bx+c, by the center position (x by continuous several frames₁,y₁), (x₂,y₂), (x₃,y₃) be brought in three fitting functions as parameter, the parameter in three functions is all solved.Every 10 seconds, automatically update the parameter of once fitting function.

B. the central point of front cross frame is set as (x₁,y₁), (x₂,y₂), then can calculate the old man component v in x-axis_x=| x₁-x₂|/T, wherein T is the interval of two video pictures, and then can calculate next frame abscissa x₃Estimation position be: x₃=2x₂-x₁, by by x₃It is brought into fitting function y₃=f (x₃), it is possible to obtain three pre-judgement coordinatesWhereinFor the vertical coordinate that fitting function is 1. corresponding, whereinFor the vertical coordinate that fitting function is 2. corresponding, whereinFor the vertical coordinate that fitting function is 3. corresponding.

C. for the priority of three coordinates according toPriority >Priority >Priority select, when 1. fitting function is predicted in not time, 1. fitting function is designated as miss, when 2. fitting function is predicted in not time, 2. fitting function is designated as miss, when 3. fitting function is predicted in not time, 3. fitting function is designated as miss, all is marked as miss function, will be no longer participate in selecting, when three functions are all marked as miss time, then return matching miss.

Judge that the method whether fitting function hits is as follows:

A. by the image of next frame to judge centered by coordinate in advance, it is sized to border with tracking box, extract the image in next frame tracking box, and the dividing method taked according to present frame is split, and is divided into the individual fritter of (m/k) × (n/k).

B. 8 characteristic points extracted according to next frame tracking box, use formula

Calculating the ORB characteristic number of next frame, this characteristic number is one 64 string of binary characters being.

C. by the ORB characteristic number (64 strings of binary characters) of the ORB characteristic number (64 strings of binary characters) of present frame Yu next frame, make comparisons, count the number that the numerical value (0 or 1) on correspondence position is identical, if the numerical value default more than one, then meaning to follow the tracks of successfully, matching is hit.

Step 4, use the tracking module in personal PC machine, target old man is tracked in the position of next frame according to the above-mentioned old man predicted, the detailed process followed the tracks of is ORB feature and the ORB characteristic number that the tracking module in personal PC machine calculates previous frame tracking box, and under present frame tracking box in the ORB feature of position of prediction gained and ORB characteristic number, and compare whether the element on same position in the two characteristic number identical, if identical number is be more than or equal to threshold value, show to trace into target, retain current ORB feature and ORB characteristic number, continue with next frame；If identical number is less than threshold value, show that target is miss, then need to re-fetch next possible position according to described approximating method and recalculate correspondence position ORB characteristic number, and compare and contrast with current ORB characteristic number, until matching hit, continue with next frame；Or all possible Fitting Coordinate System is all miss, then this step returning tracking is lost.

If above-mentioned tracking process can trace into old man by continuous 10 frames, then carry out frame-skipping operation, namely every three frames performance objective track algorithm again, so can be greatly improved tracking efficiency；If above-mentioned tracking process returning tracking is lost, then, before needing to return to three frames, reusing above-mentioned tracking process and process, if following the tracks of successfully, continuing next frame；If following the tracks of unsuccessfully, returning tracking is lost.

If returning tracking is lost, then return to previous frame, to carrying out 360 degree of samplings around the old man currently followed the tracks of, and find the maximum tracking box of matching degree, if the matching degree of this tracking box is more than threshold value, show that tracking target picks up after the loss. and huge change occurs in the path of old man, then call the Activity recognition algorithm present case to old man and be identified；If the matching degree of this tracking box is less than threshold value, it was shown that old man follows the tracks of loss, then delete current tracking box, and call frame difference method and again detect the old man's motion conditions in monitoring region.

The behavior of the old man that tracking module is traced into by the Activity recognition module in step 5, personal PC machine is identified, if the behavior classification results identified is normal behaviour, then need not report to the police, if the behavior classification results identified is Deviant Behavior, current recognition result is passed to needs display and warning audio amplifier is reported to the police, difference according to the Deviant Behavior classification results identified, to the different signal that display is corresponding with the offer of warning audio amplifier.

When first use nursing house old man's hazardous act monitoring method of the present invention, it is necessary to using personal PC machine that initial parameter required in Activity recognition process is determined, initial parameter determines that process is: first choose the sample χ={ χ having label₁,χ₂,…,χ_n, wherein χ_iBeing the segment mark video that has label, label substance is the old man that records of this section of video behavior in daily life, for instance C₁For walking label, C₂For label of standing, C₃For label of falling, C₄For label etc. of too bending over, using these videos as training sample, each labeling requirement takes several training sample, the video of each training sample to have three dimensions, image transverse axis, the image longitudinal axis, time shaft, i.e. input data χ under various circumstances_iIt is three rank tensors, then adopts the PCA based on tensor that these training samples are learnt, thus obtaining three transition matrix: U₁,U₂,U₃, then each sample is calculated three rank tensor: Y after dimensionality reduction_i=χ_i×₁U₁×₂U₂×₃U₃, will be belonging respectively to walk, stand, fall, the N kind behavior such as too bend over will be defined as C₁,C₂…C_NClass, the classification center of each class is:Wherein Y_jFor belonging to C_iClass sample χ_jIn tensor after all dimensionality reductions,For belonging to C_iThe number of dvielement.The classification center of obtain three transition matrixes and each class behavior is inputted the Activity recognition module in personal PC machine and compares the parameter of identification as behavior；

The described step based on the PCA of tensor is as follows:

A. for the video segment of input, it is possible to understand that being a three-dimensional matrice, this matrix has three dimensions, i.e. transverse axis x, longitudinal axis y, time shaft t, if each video segment is X_i, then the video segment of all inputs is constituted four bit matrix X=[X₁,X₂,…,X_n]. random initializtion matrix U₁, U₂, U₃

B. matrix D is calculated₁=X ×₂U₂×₃U₃, to matrix D₁×D₁ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₁, calculate matrix D₂=X ×₁U₁×₃U₃, to matrix D₂×D₂ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₂, calculate matrix D₃=X ×₁U₁×₂U₂, to matrix D₃×D₃ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₃, wherein, X ×_iU_iFor the multiplication of tensor, its implication is by high dimension vector X, retains its i-th dimension, other all dimensions is launched successively, forms a two-dimensional matrix X_i, and calculate U_i×X_i, gained matrix according still further to launch order carry out inverse transformation, the higher dimensional matrix of gained be X ×_iU_iValue.

C. double iteration, gained matrix are calculatedWithDifference: WithThree matrixes of gained, by all elements summed square, gained and the variable quantity that is twice iteration, if variable quantity is less than threshold value, then iteration ends, currentlyIt is the result of algorithm gained, otherwise foregoing b process.

Identification process is compared in behavior: first as behavior, the Deviant Behavior that autotracking algorithm in video image finds is compared the sart point in time identifying sample, with current tracking box longest edge for the length of side centered by the central point in current tracking box region, thus obtaining a square area, by the video image in current region, read the monitor video of 3 seconds, as sample χ to be sorted, by formula Y=χ ×₁U₁×₂U₂×₃U₃, wherein U₁,U₂,U₃The transition matrix of process gained is determined for initial parameter, ×_iThe operative symbol that (i=1,2,3) are multiplied for tensor, obtains three rank tensor Y after this sample dimensionality reduction, by the classification center of Y Yu above-mentioned each classThree dimensions calculates its Euclidean distance, the nearest class of chosen distance is as its classification results, thus completing classification, it is recognition result by the result of gained of classifying according to current behavior, if recognition result is normal behaviours such as walking, stand, then need not transmit signal to display and warning impression, if recognition result is for falling, too the Deviant Behavior such as bend over, it is necessary to signal corresponding for current recognition result to be passed to display and warning audio amplifier.

Step 6, display show the picture of current monitor in real time, when the Deviant Behavior classification results receiving old man is reported, the signal that behavior classification results that display provides according to Activity recognition module is corresponding, showing different color boxes around old man and glimmer, warning audio amplifier sends corresponding alarm sound warning simultaneously.

Compared with prior art the invention has the beneficial effects as follows:

1. the present invention make use of target following and the two kinds of methods of Activity recognition advantage in monitoring system fully so that the two has reached the effect having complementary advantages；Target tracking module can provide the information such as movement objective orbit, speed, and can fall substantial amounts of normal behaviour for Activity recognition modular filtration, thus reducing the load of Activity recognition module；Activity recognition module provides concrete analysis result for whole monitoring system, thus enabling a system to the efficient situation distinguishing different behavior movement locus of the same race.

2. the target tracking module of the present invention adopts the target tracking algorism based on ORB feature, improves the accuracy rate of tracking；The pre-determination methods of the behavior that have employed so that each frame only need to be sampled once, reduces the load of target tracking module.

3. the present invention is by adopting tracking, it is possible to the normal regular behavior of effective eliminating, owing in daily life, most of behavior is all normal, regular behavior, therefore, the present invention can efficiently reduce the number of run of Activity recognition algorithm.

Accompanying drawing explanation

Fig. 1. for implementing the schematic diagram of the monitoring system structure composition that a kind of nursing house old man's hazardous act monitoring method of the present invention adopts；

Fig. 2. for the allomeric function module frame figure of the computer program that a kind of nursing house old man's hazardous act monitoring method of the present invention adopts；

Fig. 3. for the functional sequence block diagram of a kind of nursing house old man's hazardous act monitoring method of the present invention；

Fig. 4. for the FB(flow block) of the track algorithm adopted in a kind of nursing house old man's hazardous act monitoring method of the present invention and the pre-discriminatory analysis method of behavior；

Fig. 5. for the schematic diagram of the i.e. a kind of typical moving target movement locus that embodiment one is disclosed in a kind of nursing house old man's hazardous act monitoring method of the present invention；

Fig. 6. for the schematic diagram of the i.e. a kind of track of falling that embodiment two is disclosed in a kind of nursing house old man's hazardous act monitoring method of the present invention；

In figure: 1, CCTV camera, 2, video frequency collection card, 3, personal PC machine, 4, display, 5, warning audio amplifier, 6, No. 1 movement locus, 7, No. 2 movement locus, 8, No. 3 movement locus, 9, No. 4 movement locus.

Detailed description of the invention

Below in conjunction with accompanying drawing, the present invention is explained in detail:

The ultimate principle of nursing house old man's hazardous act monitoring method of the present invention is:

Referring to Fig. 1, the monitoring system that the present invention adopts includes the identical CCTV camera of N number of structure 1, video frequency collection card 2, personal PC machine 3, display 4 and warning audio amplifier 5.Wherein, N is less than or equal to 48 be more than or equal to 10, CCTV camera 1 is placed to the place of the daily process of old man such as nursing house corridor and outdoor indoor square, when when corridor, it is equidistant that video camera is placed to corridor ceiling distance side walls, when being placed in indoor and outdoor square, video camera is placed to the corner on square, about 3.5 meters of ground of distance, and ensure to greatest extent when laying, the monitoring dead angle area of each CCTV camera 1 can be monitored by the CCTV camera 1 that other structures are identical；The bnc interface of the CCTV camera 1 that N number of structure is identical is connected on the bnc interface corresponding to video frequency collection card 2, again the pci interface of video frequency collection card 2 is connected with pci interface corresponding on personal PC machine 3, make personal PC machine 3 can read the video image that all CCTV cameras 1 collect, and by the functional module of computer program of the present invention, video signal is processed；Display 4 is by USB interface (or DVI interface, depending on the interface that PC provides) it is connected with the USB interface (or DVI interface) of personal PC machine 3, the video image that collects for display monitoring video camera 1 and in time finding that old man is abnormal, old man is marked tracking box glimmering with specific color report to the police；Warning audio amplifier 5 is connected with the audio output interface of personal PC machine 3 by the audio interface of 3.5mm plug, for when finding that old man sends sound time abnormal and reports to the police.

The video frequency collection card 2 considering current main flow is 4 tunnels, 8 roads or 16 tunnels, namely a video frequency collection card can connect 4,8 or 16 CCTV cameras 1, one personal PC machine 3 generally has 3 PCI slot, therefore, the method can connect at most 16 × 3=48 platform CCTV camera 1, can meet the demand of middle-size and small-size nursing house completely.

Consult Fig. 2. the functional module construction of the computer program being namely arranged in personal PC machine 3 of the present invention is made up of 5 modules, i.e. image capture module, profile detection module, tracking module, Activity recognition module and display alarm module.

1) image capture module:

Consulting Fig. 1, image capture module is operate in a program module on personal PC machine 3, and for manipulating CCTV camera 1, video frequency collection card 2 reads in the monitoring image monitoring region, and monitoring image is passed to profile detection module.

2) profile detection module:

Profile detection module is operate in a program module on personal PC machine 3, acting as of this profile detection module uses frame difference method to detect the profile entering the old man monitoring region in real time, when finding to have old man to enter into monitoring region, detect the boundary information of old man, and pass it to tracking module, tracking module needs to process in real time the video image that each CCTV camera 1 collects, and even currently has N number of CCTV camera 1, then needs N number of profile detection module.In actual motion, due to the time relative to the every frame period of video camera, human motion speed is slow, therefore in real time every frame of each CCTV camera need not be carried out contour detecting, therefore, the image that each CCTV camera 1 is collected by the present invention carried out a contour detecting every 1 second, in order to disperse the load to system, N number of CCTV camera to current active, the mode adopting rotation processes: namely in current one second, then the 1st profile detection module is run at once, second profile detection module is run after the 1st module postrun 1/N second, n-th profile detection module is run after the 1st module postrun (N-1)/N second.

3) tracking module:

Tracking module is operate in a program module on personal PC machine 3, acting as of this tracking module receives the boundary information that profile detection module collects, and according to these information, moving object in monitoring region is tracked, and according to the action trail of old man, the position that current old man is possible is judged in advance, when finding that following the tracks of result is not inconsistent with the pre-result judged, namely current behavior is considered as questionable conduct, it is necessary to calls Activity recognition module and carries out Activity recognition；Otherwise continue next frame image is tracked.Each tracking module is corresponding with each profile detection module, when a certain profile detection module finds have old man to enter in monitoring region time, corresponding tracking module brings into operation, when a certain profile detection module finds that all old men move out monitoring region, corresponding tracking module is out of service.

4) Activity recognition module

Activity recognition module is operate in a program module on personal PC machine 3, and acting as of this program module receives old man's questionable conduct that tracking module finds, current questionable conduct are identified.This module only exists one in systems, and any tracking module all can call behavior identification module.

5) display alarm module

Display alarm module is operate in a program module on personal PC machine 3.The acting as of this display alarm module shows the operation interface of monitoring image that all CCTV cameras collect and some hommizations on display 4；And the recognition result of Activity recognition module can be carried out classification according to the order of severity；When needs are reported to the police, according to different ranks, tracking box in different colors is being glimmered with warning old man, and the alarm sound sent corresponding to different stage by warning audio amplifier 5 alarm sound is reported to the police.

Consulting Fig. 3, the step of a kind of nursing house old man's hazardous act monitoring method of the present invention is as follows:

Step one, use CCTV camera 1, video frequency collection card 2, personal PC machine 3 obtain the video image of the old man entering monitoring region；

The video image obtained in step one is carried out contour detecting by the profile detection module in step 2, use personal PC machine 3, uses the number of pixels change of old man's contour area of frame difference method calculating gained to judge the state that old man is presently in.Use frame difference method, the old man entering monitoring region is carried out contour detecting, two continuous frames is monitored picture X₁、X₂Subtract each other according to the gray value of its correspondence, i.e. X₁(i,j)-X₂(i, the region of j) >=δ is the profile of old man, X₁(i,j)-X₂(i, j) < δ is the background area of old man, calculates the number of pixel in the connected region corresponding to profile, and the number of pixel is by when becoming many less, enter monitoring region for old man, now profile information need not be passed to tracking module；When the number of pixel does not continue to increase, can determine whether that old man has come into monitoring region, now the profile information of old man is passed to tracking module, this module was run once every 1 second.

The profile information of the old man to obtaining of the tracking module in step 3, personal PC machine 3 is tracked, according to the defined rectangular shaped rim of the profile information of old man as tracking box, and extract local feature description's subcharacter of the ORB feature in tracking box and rotational invariance, and local feature description's subcharacter number of ORB characteristic number and rotational invariance, being input in the fitting function in tracking module by the center position coordinate of continuous several frames, old man's next frame position is predicted by the result drawn by fitting function；Detailed process is as follows:

According to the defined rectangular shaped rim of the profile information of old man as tracking box, and extract ORB feature (English full name: the OrientedFASTandRotatedBRIEF in tracking box, local feature description's subcharacter of rotational invariance) and ORB characteristic number (local feature description's subcharacter number of rotational invariance), the detailed process of the ORB feature in extraction present frame tracking box is as follows:

First present frame is carried out piecemeal: individually extracted by the image in current tracking box, this picture is represented with I, if the matrix that image is m × n dimension, image is divided the fritter of several k × k pixels, so, whole picture is divided into the individual fritter of (m/k) × (n/k) altogether, takes the pixel value that meansigma methods is this fritter of the pixel of each fritter, (x y) is expressed as xth and arranges the average gray value of fritter corresponding to y row I.

Then the ORB characteristic point of present frame is selected: according to m_pq=∑_x,yx^py^pI (x, y), wherein p, q value respectively be 0 and 1, x ∈ [1 ..., m/k], y ∈ [1 ..., n/k] obtain the gradient θ=argtan (m of characteristic point₀₁,m₁₀). by all characteristic points according to the descending arrangement of θ value, from the individual coordinate points of (m/k) × (n/k), choose maximum for θ 8 as the characteristic point chosen.

Finally calculate the ORB characteristic number of present frame: choose 8 characteristic points be calculated according to following formula:

Wherein (x_i,y_i) respectively selected under step B 8 characteristic points, by formulaObtaining an eight-digit binary number numerical value, the form class of this numerical value is similar to the eight-digit binary number character string of 10011101, by f₁(I)…f₈(I) 10 system numbers corresponding to the character string of composition 8 × 8=64 position are as the ORB characteristic number extracted.

To be extracted after the ORB characteristic number of present frame, the center position coordinate of continuous several frames is input in the fitting function in tracking module, old man's next frame position is predicted by the result drawn by fitting function, it was predicted that detailed process be:

A. being provided with three fitting functions in the tracking module in personal PC machine 3,1. fitting function is a beeline y=kx+b, and 2. fitting function is sine curve y=asin (cx-t)+kx+b, and 3. fitting function is parabola y=ax²+ bx+c, by the center position (x by continuous several frames₁,y₁), (x₂,y₂), (x₃,y₃) be brought in three fitting functions as parameter, the parameter in three functions is all solved.Every 10 seconds, automatically update the parameter of once fitting function.

B. the central point of front cross frame is set as (x₁,y₁), (x₂,y₂), then can calculate the old man component v in x-axis_x=| x₁-x₂|/T, wherein T is the interval of two video pictures, and then can calculate next frame abscissa x₃Estimation position be:By by x₃It is brought into fitting function y₃=f (x₃), it is possible to obtain three pre-judgement coordinatesWhereinFor the vertical coordinate that fitting function is 1. corresponding, whereinFor the vertical coordinate that fitting function is 2. corresponding, whereinFor the vertical coordinate that fitting function is 3. corresponding.

C. for the priority of three coordinates according toPriority >Priority >Priority select, when 1. fitting function is predicted in not time, 1. fitting function is designated as miss, when 2. fitting function is predicted in not time, 2. fitting function is designated as miss, when 3. fitting function is predicted in not time, 3. fitting function is designated as miss, all is marked as miss function, will be no longer participate in selecting, when three functions are all marked as miss time, then return matching miss.

Described judges that the method whether fitting function hits is as follows:

A. by the image of next frame to judge centered by coordinate in advance, it is sized to border with tracking box, extract the image in next frame tracking box, and the dividing method taked according to present frame is split, and is divided into the individual fritter of (m/k) × (n/k).

B. 8 characteristic points extracted according to next frame tracking box, use formula

Calculating the ORB characteristic number of next frame, this characteristic number is one 64 string of binary characters being.

C. by the ORB characteristic number (64 strings of binary characters) of the ORB characteristic number (64 strings of binary characters) of present frame Yu next frame, make comparisons, count the number that the numerical value (0 or 1) on correspondence position is identical, if the numerical value default more than one, then meaning to follow the tracks of successfully, matching is hit.

Step 4, use the tracking module in personal PC machine 3, target old man is tracked in the position of next frame according to the above-mentioned old man predicted, the detailed process followed the tracks of is ORB feature and the ORB characteristic number that the tracking module in personal PC machine calculates previous frame tracking box, and under present frame tracking box in the ORB feature of position of prediction gained and ORB characteristic number, and compare whether the element on same position in the two characteristic number identical, if identical number is be more than or equal to threshold value, show to trace into target, retain current ORB feature and ORB characteristic number, continue with next frame；If identical number is less than threshold value, show that target is miss, then need to re-fetch next possible position according to described approximating method and recalculate correspondence position ORB characteristic number, and compare and contrast with current ORB characteristic number, until matching hit, continue with next frame；Or all possible Fitting Coordinate System is all miss, then this step returning tracking is lost.

If above-mentioned tracking process can trace into old man by continuous 10 frames, then carry out frame-skipping operation, namely every three frames performance objective track algorithm again, so can be greatly improved tracking efficiency；If above-mentioned tracking process returning tracking is lost, then, before needing to return to three frames, reusing above-mentioned tracking process and process, if following the tracks of successfully, continuing next frame；If following the tracks of unsuccessfully, returning tracking is lost.

If returning tracking is lost, then return to previous frame, to carrying out 360 degree of samplings around the old man currently followed the tracks of, and find the maximum tracking box of matching degree, if the matching degree of this tracking box is more than threshold value, show that tracking target picks up after the loss. and huge change occurs in the path of old man, then call the Activity recognition algorithm present case to old man and be identified；If the matching degree of this tracking box is less than threshold value, it was shown that old man follows the tracks of loss, then delete current tracking box, and call frame difference method and again detect the old man's motion conditions in monitoring region.

The behavior of the old man that tracking module is traced into by the Activity recognition module in step 5, personal PC machine 3 is identified, if the behavior classification results identified is normal behaviour, then need not report to the police, if the behavior classification results identified is Deviant Behavior, current recognition result is passed to needs display 4 and warning audio amplifier 5 is reported to the police, difference according to the Deviant Behavior classification results identified, provides corresponding different signal to display 4 and warning audio amplifier 5.

When first use this method, it is necessary to using personal PC machine 4 that initial parameter required in Activity recognition process is determined, initial parameter determines that process is: first choose the sample χ={ χ having label₁,χ₂,…,χ_n, wherein χ_iBeing the segment mark video that has label, label substance is the old man that records of this section of video behavior in daily life, for instance C₁For walking label, C₂For label of standing, C₃For label of falling, C₄For label etc. of too bending over, using these videos as training sample, each labeling requirement takes several training sample, the video of each training sample to have three dimensions, image transverse axis, the image longitudinal axis, time shaft, i.e. input data χ under various circumstances_iIt is three rank tensors, then adopts the PCA based on tensor that these training samples are learnt, thus obtaining three transition matrix: U₁,U₂,U₃, then each sample is calculated three rank tensor: Y after dimensionality reduction_i=χ_i×₁U₁×₂U₂×₃U₃, will be belonging respectively to walk, stand, fall, the N kind behavior such as too bend over will be defined as C₁,C₂…C_NClass, the classification center of each class is:Wherein Y_jFor belonging to C_iClass sample χ_jIn tensor after all dimensionality reductions,For belonging to C_iThe number of dvielement.The classification center of obtain three transition matrixes and each class behavior is inputted the Activity recognition module in personal PC machine 3 and compares the parameter of identification as behavior；

The described step based on the PCA of tensor is as follows:

A. for the video segment of input, it is possible to understand that being a three-dimensional matrice, this matrix has three dimensions, i.e. transverse axis x, longitudinal axis y, time shaft t, if each video segment is X_i, then the video segment of all inputs is constituted four bit matrix X=[X₁,X₂,…,X_n]. random initializtion matrix U₁, U₂, U₃

B. matrix D is calculated₁=X ×₂U₂×₃U₃, to matrix D₁×D₁ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₁, calculate matrix D₂=X ×₁U₁×₃U₃, to matrix D₂×D₂ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₂, calculate matrix D₃=X ×₁U₁×₂U₂, to matrix D₃×D₃ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₃, wherein, X ×_iU_iFor the multiplication of tensor, its implication is by high dimension vector X, retains its i-th dimension, other all dimensions is launched successively, forms a two-dimensional matrix X_i, and calculate U_i×X_i, gained matrix according still further to launch order carry out inverse transformation, the higher dimensional matrix of gained be X ×_iU_iValue.

C. double iteration, gained matrix are calculatedWithDifference: WithThree matrixes of gained, by all elements summed square, gained and the variable quantity that is twice iteration, if variable quantity is less than threshold value, then iteration ends, currentlyIt is the result of algorithm gained, otherwise foregoing b process.

Identification process is compared in behavior: first as behavior, the Deviant Behavior that autotracking algorithm in video image finds is compared the sart point in time identifying sample, with current tracking box longest edge for the length of side centered by the central point in current tracking box region, thus obtaining a square area, by the video image in current region, read the monitor video of 3 seconds, as sample χ to be sorted, by formula Y=χ ×₁U₁×₂U₂×₃U₃, wherein U₁,U₂,U₃The transition matrix of process gained is determined for initial parameter, ×_iThe operative symbol that (i=1,2,3) are multiplied for tensor, obtains three rank tensor Y after this sample dimensionality reduction, by the classification center of Y Yu above-mentioned each classThree dimensions calculates its Euclidean distance, the nearest class of chosen distance is as its classification results, thus completing classification, it is recognition result by the result of gained of classifying according to current behavior, if recognition result is normal behaviours such as walking, stand, then need not transmit signal to display 4 and warning audio amplifier 5, if recognition result is for falling, too the Deviant Behavior such as bend over, it is necessary to signal corresponding for current recognition result to be passed to display 4 and warning audio amplifier 5.

Step 6, display 4 show the picture of current monitor in real time, when the Deviant Behavior classification results receiving old man is reported, the signal that behavior classification results that display 4 provides according to Activity recognition module is corresponding, showing different color boxes around old man and glimmer, warning audio amplifier 5 sends corresponding alarm sound warning simultaneously.

Provide below two detailed description of the invention detailed process to old man's hazardous act monitoring method of nursing house of the present invention to illustrate:

One, the present embodiment is use one old man of camera head monitor daily behavior under nursing house environment.In the present embodiment, we simulate the normal walking states of old man, and detailed process is as follows:

Step one, image capture module manipulation CCTV camera acquisition monitoring image:

The monitoring image of senior activity in the monitoring region gathered is passed in video frequency collection card 2 by CCTV camera 1, is transferred to personal PC machine 3 in the profile detection module of operation and processes after video frequency collection card 2 is by compression of images.

Step 2, profile detection module use frame difference method that old man is carried out contour detecting:

The monitoring image that profile detection module place reason image capture module passes over.Use frame difference method, the old man entering monitoring region is carried out contour detecting, two continuous frames is monitored picture X₁、X₂Subtract each other according to the gray value of its correspondence, i.e. X₁(i,j)-X₂(i, the region of j) >=δ is the profile of old man, X₁(i,j)-X₂(i, j) < δ is the background area of old man, δ is a default threshold value, the present embodiment sets δ=50, calculate the number of pixel in the connected region corresponding to profile of old man, the number of pixel is by, when becoming many less, entering monitoring region for old man, now profile information need not be passed to tracking module；When the number of pixel does not continue to increase, can determine whether that old man has come into monitoring region, now the profile information of old man is passed to tracking module, tracking module ran once every 1 second.

Step 3, tracking module predict old man's next frame position:

Consulting Fig. 5, as shown in FIG., old man first directly walks the movement locus that old man goes down at normal row, turns, more directly walks, until walking out monitoring region.Old man is first into monitoring region, and is detected by profile detection module, after tracking module obtains the profile information of old man, starts the movement locus of old man is tracked.First, old man directly walks in monitoring region, and now, the direction that old man advances is straight line, shown in No. 1 movement locus 6 in figure.

First present frame is carried out piecemeal: individually extracted by the image in current tracking box, this picture is represented with I, if the matrix that image is m × n dimension, image is divided the fritter of several k × k pixels, so, whole picture is divided into the individual fritter of (m/k) × (n/k) altogether, takes the pixel value that meansigma methods is this fritter of the pixel of each fritter, (x y) is expressed as xth and arranges the average gray value of fritter corresponding to y row I.

Then the ORB characteristic point of present frame is selected: according to m_pq=∑_x,yx^py^pI (x, y), wherein p, q value respectively be 0 and 1, x ∈ [1 ..., m/k], y ∈ [1 ..., n/k] obtain the gradient θ=argtan (m of characteristic point₀₁,m₁₀). by all characteristic points according to the descending arrangement of θ value, from the individual coordinate points of (m/k) × (n/k), choose maximum for θ 8 as the characteristic point chosen.

Finally calculate the ORB characteristic number of present frame: choose 8 characteristic points be calculated according to following formula:

Wherein (x_i,y_i) respectively selected under step B 8 characteristic points, by formulaObtaining an eight-digit binary number numerical value, the form class of this numerical value is similar to the eight-digit binary number character string of 10011101, by f₁(I)…f₈(I) 10 system numbers corresponding to the character string of composition 8 × 8=64 position are as the ORB characteristic number extracted.

To be extracted after the ORB characteristic number of present frame, the center position coordinate of continuous several frames is input in the fitting function in tracking module, old man's next frame position is predicted by the result drawn by fitting function, it was predicted that detailed process be:

A. being provided with three fitting functions in the tracking module in personal PC machine, 1. fitting function is a beeline y=kx+b, and 2. fitting function is sine curve y=asin (cx-t)+kx+b, and 3. fitting function is parabola y=ax²+ bx+c, by the center position (x by continuous several frames₁,y₁), (x₂,y₂), (x₃,y₃) be brought in three fitting functions as parameter, the parameter in three functions is all solved.Every 10 seconds, automatically update the parameter of once fitting function.

B. the central point of front cross frame is set as (x₁,y₁), (x₂,y₂), then can calculate the old man component v in x-axis_x=| x₁-x₂|/T, wherein T is the interval of two video pictures, and then can calculate next frame abscissa x₃Estimation position be: x₃=2x₂-x₁, by by x₃It is brought into fitting function y₃=f (x₃), it is possible to obtain three pre-judgement coordinatesWhereinFor the vertical coordinate that fitting function is 1. corresponding, whereinFor the vertical coordinate that fitting function is 2. corresponding, whereinFor the vertical coordinate that fitting function is 3. corresponding.

C. for the priority of three coordinates according toPriority >Priority >Priority select, when 1. fitting function is predicted in not time, 1. fitting function is designated as miss, when 2. fitting function is predicted in not time, 2. fitting function is designated as miss, when 3. fitting function is predicted in not time, 3. fitting function is designated as miss, all is marked as miss function, will be no longer participate in selecting, when three functions are all marked as miss time, then return matching miss.

Judge that the method whether fitting function hits is as follows:

A. by the image of next frame to judge centered by coordinate in advance, it is sized to border with tracking box, extract the image in next frame tracking box, and the dividing method taked according to present frame is split, and is divided into the individual fritter of (m/k) × (n/k).

B. 8 characteristic points extracted according to next frame tracking box, use formula

Calculating the ORB characteristic number of next frame, this characteristic number is one 64 string of binary characters being.

C. by the ORB characteristic number (64 strings of binary characters) of the ORB characteristic number (64 strings of binary characters) of present frame Yu next frame, make comparisons, count the number that the numerical value (0 or 1) on correspondence position is identical, if the numerical value default more than one, then meaning to follow the tracks of successfully, matching is hit.

Target old man is tracked by step 4, tracking module according to the position of old man's next frame that step 3 is predicted:

Use the tracking module in personal PC machine 3, target old man is tracked in the position of next frame according to the above-mentioned old man predicted, the detailed process followed the tracks of is ORB feature and the ORB characteristic number that the tracking module in personal PC machine 3 calculates previous frame tracking box, and under present frame tracking box in the ORB feature of position of prediction gained and ORB characteristic number, and compare whether the element on same position in the two characteristic number identical, if identical number is be more than or equal to threshold value, show to trace into target, retain current ORB feature and ORB characteristic number, continue with next frame；If identical number is less than threshold value, show that target is miss, then need to re-fetch next possible position according to described approximating method and recalculate correspondence position ORB characteristic number, and compare and contrast with current ORB characteristic number, until matching hit, continue with next frame；Or all possible Fitting Coordinate System is all miss, then this step returning tracking is lost.Now, owing to old man goes on along straight line, so using fitting function y=kx+b to be just easily found old man, if above-mentioned tracking process can trace into old man by continuous 10 frames, then carry out frame-skipping operation, namely every three frames performance objective track algorithm again, so can be greatly improved tracking efficiency.

When old man has moved to the end of No. 1 movement locus 6 in figure, when being prepared to enter into the initiating terminal of No. 2 movement locus 7, movement locus changes, now, fitting function y=kx+b is miss, this function is labeled as miss, replacing fitting function is: y=asin (cx-t)+kx+b, and judge that whether the frequency of this SIN function is abnormal, when old man turns, the frequency of SIN function is excessive, and what exceed that people normally walks rocks scope, so giving up just profound function, and to change fitting function be y=ax²+ bx+c, it is noted that old man's movement locus when turning is parabola, No. 2 movement locus 7 in figure, therefore, old man in the process of the track shown in movement locus 7, fitting function y=ax²+ bx+c can effectively predict next step position of old man.

When old man moves to the end of No. 2 movement locus 7 in figure, when being prepared to enter into the initiating terminal of No. 3 movement locus 8, movement locus there occurs change, now, and fitting function y=ax²+ bx+c is miss, it is contemplated that now all of fitting function all has been labeled as miss, therefore, using this frame as start frame, calling Activity recognition algorithm and be identified, recognition result is for walking, for normal behaviour, now, three fitting functions are all cancelled labelling.Reselect y=kx+b to be fitted.

Noticing that old man's movement locus in region shown in No. 3 movement locus 8 in the drawings is straight line, therefore, the present invention can effectively dope the old man position at next frame.When old man moves to the border in monitoring region time, the area of tracking box diminishes.Occur following the tracks of situation about losing due to as easy as rolling off a log at this moment, and at this moment old man be tracked and Activity recognition is nonsensical, when occur following the tracks of lose time, this algorithm deletes tracking box, calls frame difference method and detects.When finding the situation that tracking box area reduces and a lateral boundaries is fixing, namely can conclude that this old man is walking out tracked region, frame difference method not using this tracking box as parameter call track algorithm, but continue to monitor, until old man walks out tracing area.

In the present embodiment, the present invention needs only to No. 1 movement locus 6 in the drawings, No. 2 movement locus 7, fitting function is changed when the starting point of No. 3 movement locus 8 occurs some fitting function miss, and owing to all fitting functions are all marked as miss thus needing to call Activity recognition algorithm and being identified in the process entering into movement locus 8 from movement locus 7, visible, this algorithm is when monitoring normal behaviour, most normal behaviour can be got rid of, improve the operational efficiency of system.

In actual use, owing to the movement locus of old man is not likely to be the more complete straight line of ratio, in particular by after the process of three frames performance objective track algorithm again, easily occur following the tracks of Loss, if above-mentioned tracking process returning tracking is lost, before then needing to return to three frames, reusing above-mentioned tracking process and process, if following the tracks of successfully, continuing next frame；If following the tracks of unsuccessfully, returning tracking is lost.

If returning tracking is lost, then return to previous frame, to carrying out 360 degree of samplings around the old man currently followed the tracks of, and find the maximum tracking box of matching degree, if the matching degree of this tracking box is more than threshold value, show that tracking target picks up after the loss. and huge change occurs in the path of old man, then call the Activity recognition algorithm present case to old man and be identified；If the matching degree of this tracking box is less than threshold value, it was shown that old man follows the tracks of loss, then delete current tracking box, and call frame difference method and again detect the old man's motion conditions in monitoring region.

The questionable conduct of old man are identified by step 5, Activity recognition module:

1) Parameter Initialization procedure:

In the present embodiment, behavior is divided into 6 classes: C₁Class is walking, C₂Class for standing, C₃Class for falling, C₄Class for too to bend over, C₅Class is impact mutually, C₆Class is that two people are overlapping and cannot the behavior of identification.

A. by gathering old man's video sample in above-mentioned 6 behaviors in actual life, for the video segment of input, it is possible to understand that being a three-dimensional matrice, this matrix has three dimensions, i.e. transverse axis x, longitudinal axis y, time shaft t, if each video segment is X_i, then the video segment of all inputs is constituted four bit matrix X=[X₁,X₂,…,X_n]. random initializtion matrix U₁, U₂, U₃.

B. matrix D is calculated₁=X ×₂U₂×₃U₃, to matrix D₁×D₁ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₁, calculate matrix D₂=X ×₁U₁×₃U₃, to matrix D₂×D₂ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₂, calculate matrix D₃=X ×₁U₁×₂U₂, to matrix D₃×D₃ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₃, wherein, X ×_iU_iFor the multiplication of tensor, its implication is by high dimension vector X, retains its i-th dimension, other all dimensions is launched successively, forms a two-dimensional matrix X_i, and calculate U_i×X_i, gained matrix according still further to launch order carry out inverse transformation, the higher dimensional matrix of gained be X ×_iU_iValue.

C. double iteration, gained matrix are calculatedWithDifference: WithThree matrixes of gained, by all elements summed square, gained and the variable quantity that is twice iteration, if variable quantity is less than threshold value, then iteration ends, currentlyIt is the result of algorithm gained, otherwise repeats step B.

By said method, obtain transition matrix U₁,U₂,U₃, pass throughCalculate the classification center of each classThe data of Parameter Initialization procedure gained are written to when system is dispatched from the factory in system, call at any time for Activity recognition process.

2) old man's normal row is walked behavior to be identified

When tracking module finds questionable conduct, and when calling Activity recognition module, Activity recognition module is with current time frame for starting point, with current tracking box longest edge for the length of side centered by the central point in current tracking box region, thus obtaining a square area, the video image in current region is read 3 seconds, thus obtaining one section of video image, this video has three dimensions: image transverse axis, the image longitudinal axis, time shaft, i.e. three rank tensor χ, by formula Y=X ×₁U₁×₂U₂×₃U₃(wherein, X is input video, U₁,U₂,U₃For the transition matrix of Parameter Initialization procedure gained, ×_i(i=1,2,3) are operative symbol, obtain the tensor Y after dimensionality reduction, by calculating Y and six classification centerDistance, i.e. Euclidean distance between higher-dimension array, the C that chosen distance is nearest_iClass is as classification results.

In the present embodiment, old man is when moving track significant change, and tracking module calls Activity recognition module and is identified.The classification results of Activity recognition module is for belonging to C_iClass, for normal behaviour, it is not necessary to report to the police, and the result of identification must be returned to tracking module.

The old man's Deviant Behavior identified is reported to the police by step 6, display alarm module

In the present embodiment, old man's behavior is normal behaviour, therefore, shows current monitoring image, do not do any alert process in display 4 screen.

Two, in the present embodiment, we have selected kind of a representative Deviant Behavior of falling, and the implementation process of the present embodiment is as follows:

Step one, use image capture module manipulation CCTV camera acquisition monitoring image:

The monitoring image of senior activity in the monitoring region gathered is passed in video frequency collection card 2 by CCTV camera 1, is transferred to personal PC machine 3 in the profile detection module of operation and processes after video frequency collection card 2 is by compression of images.

Old man is carried out contour detecting by step 2, use profile detection module:

The monitoring image that profile detection module place reason image capture module passes over.Use frame difference method, the old man entering monitoring region is carried out contour detecting, two continuous frames is monitored picture X₁、X₂Subtract each other according to the gray value of its correspondence, i.e. X₁(i,j)-X₂(i, the region of j) >=δ is the profile of old man, X₁(i,j)-X₂(i, j) < δ is the background area of old man, δ is a default threshold value, the present embodiment sets δ=50, calculate the number of pixel in the connected region corresponding to profile of old man, the number of pixel is by, when becoming many less, entering monitoring region for old man, now profile information need not be passed to tracking module；When the number of pixel does not continue to increase, can determine whether that old man has come into monitoring region, now the profile information of old man is passed to tracking module, tracking module ran once every 1 second.

Step 3, use tracking module prediction old man's next frame position:

Old man is in the front and back fallen, its movement locus there occurs obvious change, in reality, old man is at the movement locus fallen, consult shown in the head of the movement locus 9 shown in Fig. 6, it it is a flex point slightly swinging to side, when old man is before falling, running orbit is straight line, then first present frame is carried out piecemeal: individually extracted by the image in current tracking box, this picture is represented with I, if the matrix that image is m × n dimension, image is divided the fritter of several k × k pixels, so, whole picture is divided into the individual fritter of (m/k) × (n/k) altogether, take the pixel value that meansigma methods is this fritter of the pixel of each fritter, I (x, y) it is expressed as xth and arranges the average gray value of fritter corresponding to y row.

Then the ORB characteristic point of present frame is selected: according to m_pq=∑_x,yx^py^pI (x, y), wherein p, q value respectively be 0 and 1, x ∈ [1 ..., m/k], y ∈ [1 ..., n/k] obtain the gradient θ=argtan (m of characteristic point₀₁,m₁₀). by all characteristic points according to the descending arrangement of θ value, from the individual coordinate points of (m/k) × (n/k), choose maximum for θ 8 as the characteristic point chosen.

Finally calculate the ORB characteristic number of present frame: choose 8 characteristic points be calculated according to following formula:

Wherein (x_i,y_i) respectively selected under step B 8 characteristic points, by formulaObtaining an eight-digit binary number numerical value, the form class of this numerical value is similar to the eight-digit binary number character string of 10011101, by f₁(I)…f₈(I) 10 system numbers corresponding to the character string of composition 8 × 8=64 position are as the ORB characteristic number extracted.

To be extracted after the ORB characteristic number of present frame, the center position coordinate of continuous several frames is input in the fitting function in tracking module, old man's next frame position is predicted by the result drawn by fitting function, it was predicted that detailed process be:

A. being provided with three fitting functions in the tracking module in personal PC machine 3,1. fitting function is a beeline y=kx+b, and 2. fitting function is sine curve y=asin (cx-t)+kx+b, and 3. fitting function is parabola y=ax²+ bx+c, by the center position (x by continuous several frames₁,y₁), (x₂,y₂), (x₃,y₃) be brought in three fitting functions as parameter, the parameter in three functions is all solved.Every 10 seconds, automatically update the parameter of once fitting function.

B. the central point of front cross frame is set as (x₁,y₁), (x₂,y₂), then can calculate the old man component v in x-axis_x=| x₁-x₂|/T, wherein T is the interval of two video pictures, and then can calculate next frame abscissa x₃Estimation position be: x₃=2x₂-x₁, by by x₃It is brought into fitting function y₃=f (x₃), it is possible to obtain three pre-judgement coordinatesWhereinFor the vertical coordinate that fitting function is 1. corresponding, whereinFor the vertical coordinate that fitting function is 2. corresponding, whereinFor the vertical coordinate that fitting function is 3. corresponding.

C. for the priority of three coordinates according toPriority >Priority >Priority select, when 1. fitting function is predicted in not time, 1. fitting function is designated as miss, when 2. fitting function is predicted in not time, 2. fitting function is designated as miss, when 3. fitting function is predicted in not time, 3. fitting function is designated as miss, all is marked as miss function, will be no longer participate in selecting, when three functions are all marked as miss time, then return matching miss.

Judge that the method whether fitting function hits is as follows:

A. by the image of next frame to judge centered by coordinate in advance, it is sized to border with tracking box, extract the image in next frame tracking box, and the dividing method taked according to present frame is split, and is divided into the individual fritter of (m/k) × (n/k).

B. 8 characteristic points extracted according to next frame tracking box, use formula

Calculating the ORB characteristic number of next frame, this characteristic number is one 64 string of binary characters being.

C. by the ORB characteristic number (64 strings of binary characters) of the ORB characteristic number (64 strings of binary characters) of present frame Yu next frame, make comparisons, count the number that the numerical value (0 or 1) on correspondence position is identical, if the numerical value default more than one, then meaning to follow the tracks of successfully, matching is hit.

Target old man is tracked by the position of old man's next frame that step 4, use tracking module are predicted according to step 3:

Use the tracking module in personal PC machine 3, target old man is tracked in the position of next frame according to the above-mentioned old man predicted, the detailed process followed the tracks of is ORB feature and the ORB characteristic number that the tracking module in personal PC machine 3 calculates previous frame tracking box, and under present frame tracking box in the ORB feature of position of prediction gained and ORB characteristic number, and compare whether the element on same position in the two characteristic number identical, if identical number is be more than or equal to threshold value, show to trace into target, retain current ORB feature and ORB characteristic number, continue with next frame；If identical number is less than threshold value, show that target is miss, then need to re-fetch next possible position according to described approximating method and recalculate correspondence position ORB characteristic number, and compare and contrast with current ORB characteristic number, until matching hit, continue with next frame；Or all possible Fitting Coordinate System is all miss, then this step returning tracking is lost.Now, owing to old man goes on along straight line, so using fitting function y=kx+b to be just easily found old man, if above-mentioned tracking process can trace into old man by continuous 10 frames, then carry out frame-skipping operation, namely every three frames performance objective track algorithm again, so can be greatly improved tracking efficiency.

Owing to occurring in that speed clearly and direction of motion variation issue in the scene of Falls Among Old People, matching is miss, then need that present image information is passed to Activity recognition module and be identified.

The questionable conduct of old man are identified by step 5, Activity recognition module

1) Parameter Initialization procedure:

In the present embodiment, behavior is divided into 6 classes: C₁Class is walking, C₂Class for standing, C₃Class for falling, C₄Class for too to bend over, C₅Class is impact mutually, C₆Class is that two people are overlapping and cannot the behavior of identification.

A. by gathering old man's video sample in above-mentioned 6 behaviors in actual life, for the video segment of input, it is possible to understand that being a three-dimensional matrice, this matrix has three dimensions, i.e. transverse axis x, longitudinal axis y, time shaft t, if each video segment is X_i, then the video segment of all inputs is constituted four bit matrix X=[X₁,X₂,…,X_n]. random initializtion matrix U₁, U₂, U₃.

B. matrix D is calculated₁=X ×₂U₂×₃U₃, to matrix D₁×D₁ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₁, calculate matrix D₂=X ×₁U₁×₃U₃, to matrix D₂×D₂ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₂, calculate matrix D₃=X ×₁U₁×₂U₂, to matrix D₃×D₃ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₃, wherein, X ×_iU_iFor the multiplication of tensor, its implication is by high dimension vector X, retains its i-th dimension, other all dimensions is launched successively, forms a two-dimensional matrix X_i, and calculate U_i×X_i, gained matrix according still further to launch order carry out inverse transformation, the higher dimensional matrix of gained be X ×_iU_iValue.

C. double iteration, gained matrix are calculatedWithDifference: WithThree matrixes of gained, by all elements summed square, gained and the variable quantity that is twice iteration, if variable quantity is less than threshold value, then iteration ends, currentlyIt is the result of algorithm gained, otherwise repeats b step.

By said method, obtain transition matrix U₁,U₂,U₃, pass throughCalculate the classification center of each classThe data of Parameter Initialization procedure gained are written to when system is dispatched from the factory in system, call at any time for Activity recognition process.

2) Falls Among Old People situation is carried out Activity recognition

When tracking module finds questionable conduct, and when calling Activity recognition module, Activity recognition module is with current time frame for starting point, with current tracking box longest edge for the length of side centered by the central point in current tracking box region, thus obtaining a square area, the video image in current region is read 3 seconds, thus obtaining one section of video image, this video has three dimensions: image transverse axis, the image longitudinal axis, time shaft, i.e. three rank tensor χ, by formula Y=X ×₁U₁×₂U₂×₃U₃(wherein, X is input video, U₁,U₂,U₃For the transition matrix of Parameter Initialization procedure gained, ×_i(i=1,2,3) are operative symbol, obtain the tensor Y after dimensionality reduction, by calculating Y and six classification centerDistance, i.e. Euclidean distance between higher-dimension array, the C that chosen distance is nearest_iClass is as classification results.

In the present embodiment, old man is when moving track significant change, and tracking module calls Activity recognition module and is identified.The classification results of Activity recognition module is for belonging to C₃Class is the behavior of falling, it is necessary to report to the police, and the result of identification must be returned to tracking module.

The old man's Deviant Behavior identified is reported to the police by step 6, display alarm module

In the present embodiment, the recognition result of old man's behavior is for belonging to C₃Class is the behavior of falling, it is necessary to reports to the police, then reports to the police according to the order of severity set in advance, warning adopts the mode that image and sound combine, tracking box shows on the screen of display 4, and changes with corresponding warning color and glimmer, and warning audio amplifier 5 sends the sound of correspondence.

Claims (5)

1. nursing house old man's hazardous act monitoring method, it is characterised in that the step of described a kind of nursing house old man's hazardous act monitoring method is as follows:

The monitoring image of old man the monitoring image of collection passes to video frequency collection card (2) in step one, CCTV camera (1) acquisition monitoring region, video frequency collection card (2) will be transferred to after Surveillant Image Compression and process at personal PC machine (3)；

The monitoring image gathered is carried out contour detecting by the profile detection module in step 2, personal PC machine (3), uses the number of pixels change that frame difference method calculates old man's contour area to judge old man's state in which；

The profile information of the old man to obtaining of the tracking module in step 3, personal PC machine (3) is tracked, according to the defined rectangular shaped rim of the profile information of old man as tracking box, and extract local feature description's subcharacter of the ORB feature in tracking box and rotational invariance, and local feature description's subcharacter number of ORB characteristic number and rotational invariance, being input in the fitting function in tracing program by the center position coordinate of continuous several frames, old man's next frame position is predicted by the result drawn by fitting function；

Tracking module in step 4, personal PC machine (3), is tracked target old man in the position of next frame according to predicted old man；

The behavior of the old man that tracking module is followed the tracks of is identified by the Activity recognition module in step 5, personal PC machine (3), if the behavior classification results identified is normal behaviour, then need not report to the police, if the behavior classification results identified is Deviant Behavior, current recognition result is passed to needs display (4) and warning audio amplifier (5) is reported to the police, and the difference according to the Deviant Behavior classification results identified, provide different signals to display (4) and warning audio amplifier (5)；

The upper picture showing current monitor in real time of step 6, display (4), when the Deviant Behavior classification results receiving old man is reported, the signal that behavior classification results that display (4) provides according to Activity recognition module is corresponding, showing different color boxes around old man and glimmer, audio amplifier of simultaneously reporting to the police (5) sends corresponding alarm sound and reports to the police.

2. nursing house old man's hazardous act monitoring method according to claim 1, it is characterized in that judging that old man is presently in the detailed process of state and is described in step 2: use frame difference method, the old man entering monitoring region is carried out contour detecting, two continuous frames is monitored picture X₁、X₂Subtract each other according to the gray value of its correspondence, i.e. X₁(i,j)-X₂(i, the region of j) >=δ is the profile of old man, X₁(i,j)-X₂(i, j) < δ is the background area of old man, δ is a default threshold value, calculate the number of pixel in the connected region corresponding to profile, the number of pixel is by when becoming many less, enter monitoring region for old man, now profile information need not be passed to tracking module；When the number of pixel does not continue to increase, can determine whether that old man has come into monitoring region, now the profile information of old man is passed to tracking module, this module was run once every 1 second.

3. nursing house old man's hazardous act monitoring method according to claim 1, it is characterised in that the process of the prediction old man's next frame position described in step 3 is:

According to the defined rectangular shaped rim of the profile information of old man as tracking box, and extracting the ORB feature in tracking box and ORB characteristic number, the detailed process extracting the ORB feature in present frame tracking box and ORB characteristic number is as follows:

First present frame is carried out piecemeal: individually extracted by the image in current tracking box, this picture is represented with I, if the matrix that image is m × n dimension, image is divided the fritter of several k × k pixels, so, whole picture is divided into the individual fritter of (m/k) × (n/k) altogether, takes the pixel value that meansigma methods is this fritter of the pixel of each fritter, (x y) is expressed as xth and arranges the average gray value of fritter corresponding to y row I；

Then the ORB characteristic point of present frame is selected: according to m_pq=Σ_x,yx^py^pI (x, y), wherein p, q value respectively be 0 and 1, x ∈ [1 ..., m/k], y ∈ [1 ..., n/k] obtain the gradient θ=argtan (m of characteristic point₀₁,m₁₀). by all characteristic points according to the descending arrangement of θ value, from the individual coordinate points of (m/k) × (n/k), choose maximum for θ 8 as the characteristic point chosen；

Finally calculate the ORB characteristic number of present frame: choose 8 characteristic points be calculated according to following formula:

$\mathrm{\&tau;}(I,x_1,y_1,x_2,y_2)=\left\{\begin{array}{cc}1& :I(x_1,y_1)<I(x_2,y_2)\\ 0& :I(x_1,y_1,)\&GreaterEqual;I(x_2,y_2)\end{array}\right.$

Wherein (x_i,y_i) respectively selected under step B 8 characteristic points, by formulaObtaining an eight-digit binary number numerical value, the form class of this numerical value is similar to the eight-digit binary number character string of 10011101, by f₁(I) ..., f₈(I) 10 system numbers corresponding to the character string of composition 8 × 8=64 position are as the ORB characteristic number extracted；

To be extracted after the ORB characteristic number of present frame, the center position coordinate of continuous several frames is input in the fitting function in tracing program, old man's next frame position is predicted by the result drawn by fitting function, it was predicted that detailed process be:

A. being provided with three fitting functions in the tracing program in personal PC machine (3), 1. fitting function is a beeline y=kx+b, and 2. fitting function is sine curve y=asin (cx-t)+kx+b, and 3. fitting function is parabola y=ax²+ bx+c, by the center position (x by continuous several frames₁,y₁), (x₂,y₂), (x₃,y₃) be brought in three fitting functions as parameter, the parameter in three functions is all solved, every 10 seconds, automatically updates the parameter of once fitting function；

B. the central point of front cross frame is set as (x₁,y₁), (x₂,y₂), then can calculate the old man component v in x-axis_x=| x₁-x₂|/T, wherein T is the interval of two video pictures, and then can calculate next frame abscissa x₃Estimation position be: x₃=2x₂-x₁, by by x₃It is brought into fitting function y₃=f (x₃), it is possible to obtain three pre-judgement coordinatesWhereinFor the vertical coordinate that fitting function is 1. corresponding, whereinFor the vertical coordinate that fitting function is 2. corresponding, whereinFor the vertical coordinate that fitting function is 3. corresponding；

C. for the priority of three coordinates according to Select, when 1. fitting function is predicted in not time, 1. fitting function is designated as miss, when 2. fitting function is predicted in not time, 2. fitting function is designated as miss, when 3. fitting function is predicted in not time, 3. fitting function is designated as miss, all is marked as miss function, will be no longer participate in selecting, when three functions are all marked as miss time, then return matching miss；

Judge that the method whether fitting function hits is as follows:

First by the image of next frame to judge centered by coordinate in advance, it is sized to border with tracking box, extract the image in next frame tracking box, and the dividing method taked according to present frame is split, and is divided into the individual fritter of (m/k) × (n/k)；

Then 8 characteristic points extracted according to next frame tracking box, use formulaCalculating the ORB characteristic number of next frame, this characteristic number is one 64 string of binary characters being；

Finally by the ORB characteristic number (64 strings of binary characters) of the ORB characteristic number (64 strings of binary characters) of present frame Yu next frame, make comparisons, count the number that the numerical value (0 or 1) on correspondence position is identical, if the numerical value default more than one, then meaning to follow the tracks of successfully, matching is hit.

4. nursing house old man's hazardous act monitoring method according to claim 1, it is characterized in that the detailed process that target old man is tracked by the old man that the foundation described in step 4 is predicted in the position of next frame is, use the tracing program in personal PC machine (3), target old man is tracked in the position of next frame according to the above-mentioned old man predicted, the detailed process followed the tracks of is ORB feature and the ORB characteristic number that the tracing program in personal PC machine (3) calculates previous frame tracking box, and under present frame tracking box in the ORB feature of position of prediction gained and ORB characteristic number, and compare whether the element on same position in the two characteristic number identical, if identical number is be more than or equal to threshold value, show to trace into target, retain current ORB feature and ORB characteristic number, continue with next frame；If identical number is less than threshold value, show that target is miss, then need to re-fetch next possible position according to described approximating method and recalculate correspondence position ORB characteristic number, and compare and contrast with current ORB characteristic number, until matching hit, continue with next frame；Or all possible Fitting Coordinate System is all miss, then this step returning tracking is lost；

If above-mentioned tracking process can trace into old man by continuous 10 frames, then carry out frame-skipping operation, namely every three frames performance objective track algorithm again, so can be greatly improved tracking efficiency；If above-mentioned tracking process returning tracking is lost, then, before needing to return to three frames, reusing above-mentioned tracking process and process, if following the tracks of successfully, continuing next frame；If following the tracks of unsuccessfully, returning tracking is lost；

If returning tracking is lost, then return to previous frame, to carrying out 360 degree of samplings around the old man currently followed the tracks of, and find the maximum tracking box of matching degree, if the matching degree of this tracking box is more than threshold value, show that tracking target picks up after the loss. and huge change occurs in the path of old man, then call the Activity recognition algorithm present case to old man and be identified；If the matching degree of this tracking box is less than threshold value, it was shown that old man follows the tracks of loss, then delete current tracking box, and call frame difference method and again detect the old man's motion conditions in monitoring region.

5. nursing house old man's hazardous act monitoring method according to claim 1, it is characterised in that the process that the behavior of the old man that tracing program is traced into by the Activity recognition program in the use personal PC machine (3) described in step 5 is identified includes using initial parameter during this method determine process and use behavior during this method to compare identification process every time for the first time；

Described initial parameter during this method that uses for the first time determines that process is: first choose the sample χ={ χ having label₁,χ₂,…,χ_n, wherein χ_iBeing the segment mark video that has label, label substance is the old man that records of this section of video behavior in daily life, for instance C₁For walking label, C₂For label of standing, C₃For label of falling, C₄For label of too bending over, using these videos as training sample, each labeling requirement takes several training sample, the video of each training sample to have three dimensions, image transverse axis, the image longitudinal axis, time shaft, i.e. input data χ under various circumstances_iIt is three rank tensors, then adopts and based on incremental differentiation tensor and canonical correlation analysis algorithm, these training samples are learnt, thus obtaining three transition matrix: U₁,U₂,U₃, then each sample is calculated three rank tensor: Y after dimensionality reduction_i=χ_i×₁U₁×₂U₂×₃U₃, will be belonging respectively to walk, stand, fall, N kind behavior of too bending over will be defined as C₁,C₂…C_NClass, the classification center of each class is:Wherein Y_jFor belonging to C_iClass sample χ_jIn tensor after all dimensionality reductions,For belonging to C_iThe number of dvielement, inputs the Activity recognition program in personal PC machine (3) using the classification center of obtain three transition matrixes and each class behavior and compares the parameter of identification as behavior；

The described step based on the PCA of tensor is as follows:

A. for the video segment of input, it is possible to understand that being a three-dimensional matrice, this matrix has three dimensions, i.e. transverse axis x, longitudinal axis y, time shaft t, if each video segment is X_i, then the video segment of all inputs is constituted four bit matrix X=[X₁,X₂,…,X_n]. random initializtion matrix U₁, U₂, U₃

B. matrix D is calculated₁=X ×₂U₂×₃U₃, to matrix D₁×D₁ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₁, calculate matrix D₂=X ×₁U₁×₃U₃, to matrix D₂×D₂ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₂, calculate matrix D₃=X ×₁U₁×₂U₂, to matrix D₃×D₃ ^TSeek its eigenvalue λ_iAnd characteristic vector v_i, select eigenvalue λ_iThe set that characteristic vector corresponding to > 0 is constituted, is assigned to U by the value of this set₃, wherein, X ×_iU_iFor the multiplication of tensor, its implication is by high dimension vector X, retains its i-th dimension, other all dimensions is launched successively, forms a two-dimensional matrix X_i, and calculate U_i×X_i, gained matrix according still further to launch order carry out inverse transformation, the higher dimensional matrix of gained be X ×_iU_iValue；

C. double iteration, gained matrix are calculatedWithDifference: WithThree matrixes of gained, by all elements summed square, gained and the variable quantity that is twice iteration, if variable quantity is less than threshold value, then iteration ends, currentlyIt is the result of algorithm gained, otherwise repeats foregoing b process；

Described use behavior during this method to compare identification process to be: first as behavior, the Deviant Behavior that autotracking algorithm in video image finds is compared the sart point in time identifying sample every time, with current tracking box longest edge for the length of side centered by the central point in current tracking box region, thus obtaining a square area, by the video image in current region, read the monitor video of 3 seconds, as sample X to be sorted, by formula Y=χ ×₁U₁×₂U₂×₃U₃, wherein U₁,U₂,U₃The transition matrix of process gained is determined for initial parameter, ×_iThe operative symbol that (i=1,2,3) are multiplied for tensor, obtains three rank tensor Y after this sample dimensionality reduction, by the classification center of Y Yu above-mentioned each classThree dimensions calculates its Euclidean distance, the nearest class of chosen distance is as its classification results, thus completing classification, it is recognition result by the result of gained of classifying according to current behavior, if recognition result is walking, normal behaviour of standing, then need not transmit signal to display (4) and warning audio amplifier (5), if recognition result is for falling, too bend over Deviant Behavior, it is necessary to signal corresponding for current recognition result to be passed to display (4) and warning audio amplifier (5).