The content of the invention
The purpose of the embodiment of the present invention is to provide a kind of tracking of low signal-to-noise ratio moving small target with knowing method for distinguishing, purport
There is the problem of tracking real-time speed is compared with slow, tracking or poor recognition effect solving prior art.
The embodiment of the present invention is achieved in that a kind of tracking of low signal-to-noise ratio moving small target with knowing method for distinguishing, its
It is characterised by, the tracking of the low signal-to-noise ratio moving small target stage by stage, is according to target provided not respectively with knowing method for distinguishing with one kind
Same implementation, is easily blocked or is flooded by other objects or noise in complex background to moving small target in video image
Situation, it is proposed that the elimination of opening and closing conversion or the algorithm for weakening background and noise;To the small and weak characteristic of Small object, it is proposed that
The adaptive neural network competitive model of line study, the multidimensional characteristic parameter of Weak target is extracted using the active unit of competition;
For the kinetic characteristic of video Small Target, using the mutation of gray scale, Small object motion state model and forecast model are given;
Detection real-time to moving small target and tracking, employ fuzzy pushdown automata chain and carry out track identification and tracking, under fuzzy
Push away automatic chain depth and carry out track judgement for threshold value;
Specifically include following steps:
Step one, the extraction mesh calibration method from the single-frame images of video sequence is provided, weakens or eliminates background and noise
Influence;
Step 2, provides Weak target extraction of motion information and status predication modeling;
Step 3, sets up the incidence matrix of two inter frame image moving small targets;
Step 4, the information fusion being superimposed based on multiple image, is carried using fuzzy pushdown automata chain bullet stack recursive operation
Track algorithm and recognition methods of the large-scale image with video image motion Small object under network environment are gone out.
Further, in step one, the extraction mesh calibration method from the single-frame images of video sequence is provided, weakens or eliminates
During the influence of background and noise, concrete methods of realizing is:
Local maximum and minimum are asked for using mathematical morphology combinatorial operation, mitigates the amount of calculation of subsequent treatment, to the greatest extent
Amount reduces false alarm rate and counted, and region growing, minimum point are carried out to each Local modulus maxima and implements to weaken or eliminates, to possible
Target selected;
G=f-f ο B or closed operation conversion g=fB-f are converted using opening operation, single-frame images is filtered, detected
The place changed greatly in image, that is, high fdrequency component, while it is comparatively gentle to filter off gray-value variation in image
Place, equivalent to low frequency component, just a width single-frame images is filtered using this conversion, low-frequency component is filtered, quite
In filtering extended background, the HFS including leaving comprising Small object;In formula, f is gradation of image frame, and B is structure, ο
Opening operation is represented, closed operation is represented;
To accurately identify target or track, carry out the suppression of Small object enhancing and interference, due to Small object point each frame it
Between move, multiframe superposition can be carried out to video image, in the last frame of superposition, Small object point shows as correlation very strong rail
Mark point, but noise, it is possible to flooding Small object track, proposition multi-frame difference superposition algorithm, selection exists comprising moving small target point
Interior image sequence, the superposition value of odd-numbered frame and each n frames difference of even frame, i.e.,:
In formula, fiFor the i-th frame in image sequence, fzFor last superposition frame;
Through the stack plus frame takes thresholding to handle, method is as follows:
δ is threshold value, is takenM, N are the size of superposition two field picture.
Further, in step 2, when providing Weak target extraction of motion information and status predication modeling, the side of implementing
Method is:
The adaptive neural network competitive model of on-line study is constructed first, extracts weak using the active unit of its competition
The multidimensional characteristic parameter of Small object:
The first step, initializes network:The dimension of fixed output nerve network grid is N × M, and input layer is quadravalence network,
And random initializtion input neuron and the weight of output neuron connection, make t represent algorithm iteration number of times, put t=0;
Second step, selects victor:The gray scale of each frame Small object image, colourity, motion parameter X={ x1,x2,…,
xdThe input neuron in network is input to, to each input neuron value xj, the output of the node i in competition layerForG is an activation primitive, is such as takenα > 0 are constant, the slope of controlling curve;μi(t)
It is p dimensional input vectors xjWith p dimensional weights vector ωji(t) Euclidean distance between | | xj-ωji(t) | | and, i.e.,ωji(t) it is from input layer node j to competition layer node i connection weight vector, j in t
∈ J, J={ 1 ..., d }, i ∈ I, I={ 1 ..., NiBe competition layer certain regional area;
Select the output neuron i won*, in competition layer, correspondenceMinimum node will win, if that is,The node so won in competition layer is i*, then with i*The weight of association and and i*The power of the neighbouring point association of point
Weight can be all adjusted;
3rd step, updates weight:N(i*) it is triumph output neuroni*Neighbour, the distance between output neuron is specific
Specify, to each output neuron i ∈ { N (i*),i*, adjust renewal according to the following formula:
η (t)=η has determined that in advance;This rule only updates the neighbour of triumph output neuron;
4th step, standardized weight:Update to standardization after weight, so as to be consistent with input measurement standard;
5th step, is continued cycling through:The first step is repeated to the 4th step, the number of times of iteration is set to t=t+1, stopped until meeting
Machine criterion, shutting down criterion is | | xj-ωji(t) | | < ε, take ε=0.5, or untill having exceeded the cycle-index of maximum.
Further, in step 3, the implementation method for setting up the incidence matrix of two inter frame image moving small targets is:
M × n object matching matrix M are set up, here, m is the number of the moving small target of present frame, and n is the fortune of previous frame
The number of dynamic Small object, the value of element M (i, j) is given by:
R in formulaiFor the radius of the i-th target;rjFor the radius of jth target;(xi,yi) be the i-th target center-of-mass coordinate;
(xj,yj) be jth target center-of-mass coordinate;∞ represents a very big numerical value;
First, selective value is minimum in matrix M and is not ∞ element, and the row and column corresponding to the element is current respectively
The numbering of moving small target and previous frame moving small target, the corresponding moving small target of row moving small target phase corresponding with row
Match somebody with somebody, all elements value for the row and column for completing matching is then changed into ∞;Minimum value is found in continuation in matrix M, completes motion
The matching of Small object, until all values in matrix are all changed into ∞;After search terminates, the row for not finding matching target is represented
There is the appearance of new moving small target in present image, the row for not finding matching target represent certain small mesh of motion in present image
Mark disappears.
Further, in step 4, the information fusion being superimposed based on multiple image utilizes fuzzy pushdown automata chain bullet stack
The track algorithm that recursive operation proposes large-scale image and video image motion Small object under network environment is with recognition methods:
The first step, each fuzzy pushdown automata recognizes fusion in time to moving small target:
fipAnd Ξ (t)i(t) represent that t is belonged to by obscuring the identified moving small target that pushdown automata i is measured respectively
The fuzzy membership and Fuzzy Distribution of pth class,Represent untill the l moment to be merged by i-th of fuzzy pushdown automata accumulation
To identified target belong to the fuzzy membership of pth class,Represent untill the l moment by i-th of fuzzy pushdown automata product
The tired Fuzzy Distribution for merging obtained identified target, here, l=1,2 ..., t, i.e.,
With
op(p ∈ U) is moving small target, by the accumulation fusion Fuzzy Distribution and the measurement Fuzzy Distribution of t at t-1 moment
Merged, obtain target identification accumulation fusion Fuzzy Distribution of i-th of fuzzy pushdown automata untill tFor:
Wherein,S2It is Fuzzy Integration Function, usual S2Remove formula:
Now, with Fuzzy DistributionThe motion state of corresponding Small object isIt is the small mesh of current t
Target state estimation:
Fi(t) it is the motion state transfer matrix from last moment to current time, selects metastasis model,For upper a period of time
The state estimation at quarter,For the state estimation at current time;
Second step, the Space integration that fuzzy pushdown automata is recognized to Small object:
Obtaining the accumulation Fuzzy Distribution of each fuzzy pushdown automata target identification of tAfterwards, i=1 ... here,
N, is merged using Fuzzy Integration Function to this N number of Fuzzy Distribution, just obtained untill t to target identification when-
Fuzzy Distribution is merged in sky accumulation:
It is theoretical using Fuzzy Integration Function, it can obtain
SNAlso illustrate that Fuzzy Integration Function;If
Now, with Fuzzy Distribution ΞtThe motion state of corresponding Small object isIt is all small of current kth frame
The state estimation of target, motion state fusion results are: It is by the motion of the 1st frame to kth frame
The state estimation of i-th of Small object of the current kth frame of information prediction,It is the motion of all Small objects of current kth frame
State estimation,It is the fuzzy membership of model;The key of tracking be by measurement be carved into from the outset the k-1 moment predict ought
The fuzzy membership of preceding i-th of tracking system model of k momentHereIt is the fuzzy membership of k-1 moment tracking system models,Be it is known,
πji=Pr { mk=m(i)|mk-1=m(j)It is from model mk-1To model mkState transfer fuzzy membership.
Further, after multiframe is merged, Small object point is further enhanced, and most of noise spot is cut, and is filtered out
Random noise disturbance is made decisions by fuzzy pushdown automata chain length, in order to subtract as far as possible while Small object point is retained
Pushdown automata chain length is obscured less, thresholding processing is taken to fusion frame, and method is as follows:
Threshold value T is the length of fuzzy pushdown automata chain, but this chain must assure that containing Small object point, fAFor total fusion
Frame, on last fusion frame, Small object point shows as a correlation very strong track;
Each component f of the characteristic vector of moving small target to be identified or track is calculated according to blending algorithmAFuzzy membership
Spend μij, that is, the characteristic vector for obtaining Unknown Motion Small object or track is Ui=[μi1,μi2,…,μik]T;It with trained
Good known i-th0The multi-Dimensional parameters characteristic vector U of classificationi0Compare, and if only ifWhen, adjudicate fortune to be identified
Dynamic Small object or track belong to i-th0Class;EvenSo thatThen judge the small mesh of motion to be identified
Mark or track belong to i-th0Class;Here, δ is threshold value, and B is the index set of target or track class.
The tracking of the low signal-to-noise ratio moving small target that the present invention is provided is with knowing method for distinguishing, it is proposed that one kind based on single frames with
The small target tracking algorithm of the time-space domain fused filtering of multiframe and recognition methods, are easily answered moving small target in video image
The situation that other objects or noise in miscellaneous background are blocked or flooded, it is proposed that the elimination of opening and closing conversion weakens background and noise
Algorithm;To the small and weak characteristic of Small object, it is proposed that the adaptive neural network competitive model of on-line study, its competition is utilized
Active unit extracts the multidimensional characteristic parameter of Weak target;For the kinetic characteristic of video Small Target, using the mutation of gray scale,
Give Small object motion state model and forecast model;Detection real-time to moving small target and tracking, employ and are pushed away under obscuring
Automatic chain carries out track identification and tracking, track judgement is carried out using fuzzy pushdown automata chain depth as threshold value, so as to propose
A kind of moving small target track algorithm and recognition methods based under complex environment, the present invention contribute to target identification and image
Treatment people understands the characteristics of motion, active degree and its influence to other targets of detection target, so as to provide corresponding determine
Plan, seeks to suppress or eliminates influence of the undesirable element to itself or other important goal to be all very important;Will be to the military, people
The development of all target recognition and trackings based on video system such as thing, public security system, road traffic play important reference and
Reference role.The average correct recognition rata highest of the feature extraction method of identification of the present invention, in experiment, with increasing for sample number,
Average correct recognition rata constantly increases, and reaches that increasing sample curve during certain sample size again gradually tends to be steady.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
Below in conjunction with the accompanying drawings and specific embodiment to the present invention application principle be further described.
As shown in figure 1, the tracking of the low signal-to-noise ratio moving small target of the embodiment of the present invention is with knowing method for distinguishing including following
Step:
Step S101, provides the extraction mesh calibration method from the single-frame images of video sequence, weakens or eliminates background with making an uproar
The influence of sound;
Step S102, provides Weak target extraction of motion information and status predication modeling;
Step S103, sets up the incidence matrix of two inter frame image moving small targets;
Step S104, the information fusion being superimposed based on multiple image, utilizes fuzzy pushdown automata chain bullet stack recursive operation
Propose track algorithm and recognition methods of the large-scale image with video image motion Small object under network environment.
In embodiments of the present invention, in step S101, the side that target is extracted from the single-frame images of video sequence is provided
Method, during the influence of decrease or elimination background and noise, concrete methods of realizing is:
On the basis of based on mathematical morphology, it is proposed that the image object detection algorithm of opening and closing conversion.The master of the algorithm
It is to ask for local maximum and minimum using mathematical morphology combinatorial operation to want thinking, mitigates the amount of calculation of subsequent treatment, to the greatest extent
Amount reduces false alarm rate points.Carry out region growing, minimum point to each Local modulus maxima to implement to weaken or eliminate, to possible
Target selected.
Make to reach that it shows as the zero of large area by the filtered image of opening and closing conversion for each frame using this algorithm
High fdrequency component including background and comprising pinpoint target, and random spotted noise.Then moved according between consecutive frame
The correlation of target, carries out difference multiframe superposition, and Small object is due to its motility, and it is very strong on superposition frame to show as correlation
Tracing point, and high frequency spotted noise, then because its randomness is cancelled out each other a part, unmatched part is showed in geometric area
For the random distribution noise spot that correlation is very poor.
For example, using a kind of simple opening operation conversion g=f-f ο B, being filtered to single-frame images, figure can be detected
The place changed greatly as in, that is, high fdrequency component, while it is comparatively gentle to filter off gray-value variation in image
Place, equivalent to low frequency component.A width single-frame images can be just filtered using this conversion, filter low-frequency component, phase
When in filtering extended background, the HFS including leaving comprising Small object.In formula, f is gradation of image frame, and B is structure,
ο represents opening operation.Using opening and closing conversion, the wherein frame result to the video image of moving small target is as shown in Figure 2.
In embodiments of the present invention, in step s 102, Weak target extraction of motion information and status predication modeling are provided
When, concrete methods of realizing is:
Target is detected and tracked, a reference template is initially set up as standard form.
Given known moving target, i.e., in the two field picture containing this target under the conditions of certain video, detect
Its movement velocity, current location, direction of motion o, height h, gray average μ and the variances sigma of imaging.It assign the image of this frame as ginseng
Examine template.
Set up reference template as follows:Using target location central point as the center of circle, to clap frame time and target in imaging plane
Projection speed product for radius preceding half-circle area in, scan for.Using the change frequency of gray scale as thresholding, threshold is set
Value.Search the change frequency of gray scale respectively along image width and image height direction, if the change frequency of both direction is both less than or equal to 2,
Frame time is clapped in extension, until the grey scale change number of times at least one direction is more than 3 or 3 times.Then, using the center of circle as starting point,
Calculate the distance between continuous three adjacent grey scale changes d1And d2, can obtain the motion amplitude of target and target a frame into
Image width degree, Target Motion Character parameter s during using this amplitude and width as current location1And s2.It can be positioned by the saltus step of gray scale
The position of targetAnd the direction of motion, speed and height of imaging etc. are calculated with it.The gray average of target in half-circle area
It is another two characteristic parameter with variance, obtains a principal eigenvector V used in successive image processing special as standard
Levy vector.
By the movable information of Small object, will provide motion state forecast model is
Here, Fi(k) it is the motion state transfer matrix from previous frame to present frame, suitable turn must be selected it
Shifting formwork type,For the state estimation of previous frame,For the state estimation of present frame.
In embodiments of the present invention, in step s 103, the reality of the incidence matrix of two inter frame image moving small targets is set up
Now method is:
Set up m × n object matching matrixes M.Here, m is the number of the moving small target of present frame, and n is the fortune of previous frame
The number of dynamic Small object.The value of element M (i, j) is given by.
R in formulaiFor the radius of the i-th target;rjFor the radius of jth target;(xi,yi) be the i-th target center-of-mass coordinate;
(xj,yj) be jth target center-of-mass coordinate;∞ represents a very big numerical value.
Of the moving small target and moving small target in previous frame image in present image is realized using matrix is matched
Match somebody with somebody.First, selective value is minimum in matrix M and is not ∞ element, and the row and column corresponding to the element is current kinetic respectively
The numbering of Small object and previous frame moving small target, the corresponding moving small target of such row moving small target phase corresponding with row
Match somebody with somebody.Then all elements value for the row and column for completing matching is changed into ∞.Minimum value is found in continuation in matrix M, completes motion
The matching of Small object, until all values in matrix are all changed into ∞.After search terminates, the row for not finding matching target is represented
There is the appearance of new moving small target in present image, the row for not finding matching target represent certain small mesh of motion in present image
Mark disappears.
In embodiments of the present invention, in step S104, the information fusion being superimposed based on multiple image is pushed away using under fuzzy
Automatic chain bullet stack recursive operation proposes large-scale image and the track algorithm of video image motion Small object under network environment
It is with recognition methods:
Based on multiple features fusion, it is proposed that feature and the connection of motion state based on present frame and previous frame moving small target
Close probability and coarseness data association algorithm, hypothesis testing method.Estimated using joint probability and coarseness data association algorithm
Count multimode small target tracking algorithm.The information updating in rear frame is utilized to assume state using multihypothesis test method.Proposing
While these algorithms and method, provide suitable " thresholding ", only calculate the pass of the observation being located in " thresholding " and Small object
Join probability and roughness, this greatly reduces amount of calculation, solve Small object with the distance of predicted position with observation position and conflict
Problem.
The moving small target tracking and identification of image are realized, a kind of suitable recognizer must be just proposed, you can recognize
Moving small target characteristic vector for known class that is to be identified and having trained all is fuzzy number vector, by its feature to
Amount compares, defined feature membership function, degree of membership and matched rule, provides fuzzy object recognizer.It is embodied as:
To implement to track to the track of moving small target in image or video, the present invention proposes the motion shape of multiframe superposition
State blending algorithm:According to the motion state of moving small target in image or video, a kind of detect and tracking system will be designed
Blending algorithm, to realize the tracking to moving small target.This algorithm is made up of three parts:(a) mapping tracking system, which enters, has N
In the neutral net of individual feedback neural member;(b) N number of feedback neural member completion tracking system i.e. in multiframe fusion in time is allowed
State fusion;(c) M non-feedback output neurons are made to complete spatially the state of the tracking system in i.e. single frames fusion again
Fusion, finally calculates the Space Time accumulation fusion of tracking system state, come complete tracking system to the fusion of moving small target with
Track algorithm.
Designing tracking system state fusion algorithm is:Here,It is by the 1st frame to kth frame
The state estimation of i-th of Small object of the current kth frame of motion information prediction,It is all Small objects of current kth frame
State estimation,It is the fuzzy membership of model.The key of tracking is that to be carved into the k-1 moment from the outset pre- by measurement
Survey the fuzzy membership of current i-th of tracking system model of k momentHereIt is the fuzzy membership of k-1 moment tracking system models,Be it is known,
πji=Pr { mk=m(i)|mk-1=m(j)It is from model mk-1To model mkState transfer fuzzy membership.The tracking system of foundation
State fusion of uniting is as shown in Figure 3 to the tracking process of moving small target.
On last superposition frame, Small object point shows as a correlation very strong track.But identify Small object point
The still non-easy thing in track, mainly some discrete noise spots it is possible in the presence of.In addition it is also possible that tracing point and noise spot
It is interlaced.Therefore it can only proceed from the situation as a whole to be made decisions according to overall relevancy under certain hypothesis.Therefore, proposing to be based on
Fuzzy pushdown automata chain, track judgement is carried out by thresholding of the correlation length of track using the method for playing stack Yu popping.Tracking
As a result it is as shown in Figure 4.From fig. 4, it can be seen that explanation fuzzy pushdown automata D-chain trace algorithm proposed by the present invention is known in pattern
Larger success is achieved in other system.
The fuzzy membership of each component of the characteristic vector of moving small target to be identified or track is calculated according to blending algorithm
μij, you can the characteristic vector for obtaining Unknown Motion Small object or track is Ui=[μi1,μi2,…,μik]T.It with trained
Good known i-th0The multi-Dimensional parameters characteristic vector U of classificationi0Compare, and if only ifWhen, adjudicate fortune to be identified
Dynamic Small object or track belong to i-th0Class.EvenSo thatThen judge the small mesh of motion to be identified
Mark or track belong to i-th0Class.Here, δ is threshold value, and B is the index set of target or track class.
The present invention's concretely comprises the following steps:
Step one, the extraction mesh calibration method from the single-frame images of video sequence is provided, weakens or eliminates background and noise
Influence;
Step 2, provides Weak target extraction of motion information and status predication modeling;
Step 3, sets up the incidence matrix of two inter frame image moving small targets;
Step 4, the information fusion being superimposed based on multiple image, is carried using fuzzy pushdown automata chain bullet stack recursive operation
Track algorithm and recognition methods of the large-scale image with video image motion Small object under network environment are gone out.
Further, in step one, the extraction mesh calibration method from the single-frame images of video sequence is provided, weakens or eliminates
During the influence of background and noise, concrete methods of realizing is:
Because complexity, the degree of stability of scene affect the effect of target following.For example, target image by illumination not
Uniformly, influence of the variation of object etc. in background.Will from image effective Ground Split, extract target, it is necessary to propose that one kind subtracts
Algorithm that is weak or eliminating background or influence of noise.The present invention will propose opening and closing conversion on the basis of based on mathematical morphology
Image object detection algorithm.The main thought of the algorithm be using mathematical morphology combinatorial operation ask for local maximum with it is minimum
Value, mitigates the amount of calculation of subsequent treatment, and false alarm rate points are reduced as far as possible.Region growing, pole are carried out to each Local modulus maxima
Small value point is implemented to weaken or eliminated, and possible target is selected.
Make to reach that it shows as the zero of large area by the filtered image of opening and closing conversion for each frame using this algorithm
High fdrequency component including background and comprising pinpoint target, and random spotted noise.Then moved according between consecutive frame
The correlation of target, carries out difference multiframe superposition, and Small object is due to its motility, and it is very strong on superposition frame to show as correlation
Tracing point, and high frequency spotted noise, then because its randomness is cancelled out each other a part, unmatched part is showed in geometric area
For the random distribution noise spot that correlation is very poor.
For example, using a kind of simple opening operation conversion g=f-f ο B, being filtered to single-frame images, figure can be detected
The place changed greatly as in, that is, high fdrequency component, while it is comparatively gentle to filter off gray-value variation in image
Place, equivalent to low frequency component.A width single-frame images can be just filtered using this conversion, filter low-frequency component, phase
When in filtering extended background, the HFS including leaving comprising Small object.In formula, f is gradation of image frame, and B is structure,
ο represents opening operation.Using opening and closing conversion, the wherein frame result to the video image of moving small target is as shown in Figure 2.
Because opening and closing conversion filtering is relevant with structure size, the size of structure size decides high-pass filtering effect.
Structure size is smaller, filters out that low frequency background is more thorough, and the target size that can retain is just smaller.
Further, in step 2, Weak target extraction of motion information and status predication modeling, concrete methods of realizing are implemented
For:
Target is detected and tracked, a reference template is initially set up as standard form.
Given known moving target, i.e., in the two field picture containing this target under the conditions of certain video, detect
Its movement velocity, current location, direction of motion o, height h, gray average μ and the variances sigma of imaging.It assign the image of this frame as ginseng
Examine template.
Set up reference template as follows:Using target location central point as the center of circle, to clap frame time and target in imaging plane
Projection speed product for radius preceding half-circle area in, scan for.Using the change frequency of gray scale as thresholding, threshold is set
Value.Search the change frequency of gray scale respectively along image width and image height direction, if the change frequency of both direction is both less than or equal to 2,
Frame time is clapped in extension, until the grey scale change number of times at least one direction is more than 3 or 3 times.Then, using the center of circle as starting point,
Calculate the distance between continuous three adjacent grey scale changes d1And d2, can obtain the motion amplitude of target and target a frame into
Image width degree, Target Motion Character parameter s during using this amplitude and width as current location1And s2.It can be positioned by the saltus step of gray scale
The position of targetAnd the direction of motion, speed and height of imaging etc. are calculated with it.The gray average of target in half-circle area
It is another two characteristic parameter with variance, obtains a principal eigenvector V used in successive image processing special as standard
Levy vector.
By the movable information of Small object, will provide motion state forecast model is
Here, Fi(k) it is the motion state transfer matrix from previous frame to present frame, suitable turn must be selected it
Shifting formwork type,For the state estimation of previous frame,For the state estimation of present frame.
Further, in step 3, the implementation method for setting up the incidence matrix of two inter frame image moving small targets is:
Set up m × n object matching matrixes M.Here, m is the number of the moving small target of present frame, and n is the fortune of previous frame
The number of dynamic Small object.The value of element M (i, j) is given by.
R in formulaiFor the radius of the i-th target;rjFor the radius of jth target;(xi,yi) be the i-th target center-of-mass coordinate;
(xj,yj) be jth target center-of-mass coordinate;∞ represents a very big numerical value.
Of the moving small target and moving small target in previous frame image in present image is realized using matrix is matched
Match somebody with somebody.First, selective value is minimum in matrix M and is not ∞ element, and the row and column corresponding to the element is current kinetic respectively
The numbering of Small object and previous frame moving small target, the corresponding moving small target of such row moving small target phase corresponding with row
Match somebody with somebody.Then all elements value for the row and column for completing matching is changed into ∞.Minimum value is found in continuation in matrix M, completes motion
The matching of Small object, until all values in matrix are all changed into ∞.After search terminates, the row for not finding matching target is represented
There is the appearance of new moving small target in present image, the row for not finding matching target represent certain small mesh of motion in present image
Mark disappears.
Further, in step 4, the information fusion being superimposed based on multiple image utilizes fuzzy pushdown automata chain bullet stack
The track algorithm that recursive operation proposes large-scale image and video image motion Small object under network environment is with recognition methods:
Based on multiple features fusion, the present invention proposes description moving small target and calculated with tracking the data correlation of its movement locus
Method.Feature and the joint probability and coarseness data of motion state based on present frame Yu previous frame moving small target are proposed simultaneously
Association algorithm, hypothesis testing method.Utilize joint probability and coarseness data association algorithm pre-estimation multimode small target tracking
Algorithm.The information updating in rear frame is utilized to assume state using multihypothesis test method.Proposing the same of these algorithms and method
When, provide suitable " thresholding ", only calculate the association probability and roughness of the observation being located in " thresholding " and Small object, this is just
Amount of calculation is greatly reduced, Small object collision problem is solved with the distance of observation position and predicted position.
The moving small target tracking and identification of image are realized, a kind of suitable recognizer must be just proposed, you can recognize
Moving small target characteristic vector for known class that is to be identified and having trained all is fuzzy number vector, by its feature to
Amount compares, defined feature membership function, degree of membership and matched rule, provides fuzzy object recognizer.It is embodied as:
To implement to track to the track of moving small target in image or video, the present invention proposes the motion shape of multiframe superposition
State blending algorithm:According to the motion state of moving small target in image or video, a kind of detect and tracking system will be designed
Blending algorithm, to realize the tracking to moving small target.This algorithm is made up of three parts:(a) mapping tracking system, which enters, has N
In the neutral net of individual feedback neural member;(b) N number of feedback neural member completion tracking system i.e. in multiframe fusion in time is allowed
State fusion;(c) M non-feedback output neurons are made to complete spatially the state of the tracking system in i.e. single frames fusion again
Fusion, finally calculates the Space Time accumulation fusion of tracking system state, come complete tracking system to the fusion of moving small target with
Track algorithm.
Designing tracking system state fusion algorithm is:Here,It is by the 1st frame to kth frame
The current kth frame of motion information prediction i-th of Small object state estimation,It is all Small objects of current kth frame
State estimation,It is the fuzzy membership of model.The key of tracking is to be carved into the k-1 moment from the outset by measurement
Predict the fuzzy membership of current i-th of tracking system model of k momentHereIt is the fuzzy membership of k-1 moment tracking system models,Be it is known,
πji=Pr { mk=m(i)|mk-1=m(j)It is from model mk-1To model mkState transfer fuzzy membership.The tracking system of foundation
State fusion of uniting is as shown in Figure 3 to the tracking process of moving small target.
On last superposition frame, Small object point shows as a correlation very strong track.But identify Small object point
The still non-easy thing in track, mainly some discrete noise spots it is possible in the presence of.In addition it is also possible that tracing point and noise spot
It is interlaced.Therefore it can only proceed from the situation as a whole to be made decisions according to overall relevancy under certain hypothesis.Therefore, proposing to be based on
Fuzzy pushdown automata chain, track judgement is carried out by thresholding of the correlation length of track using the method for playing stack Yu popping.Tracking
As a result it is as shown in Figure 4.From fig. 4, it can be seen that explanation fuzzy pushdown automata D-chain trace algorithm proposed by the present invention is known in pattern
Larger success is achieved in other system.
The fuzzy membership of each component of the characteristic vector of moving small target to be identified or track is calculated according to blending algorithm
μij, you can the characteristic vector for obtaining Unknown Motion Small object or track is Ui=[μi1,μi2,…,μik]T.It with trained
Good known i-th0The multi-Dimensional parameters characteristic vector U of classificationi0Compare, and if only ifWhen, adjudicate fortune to be identified
Dynamic Small object or track belong to i-th0Class.EvenSo thatThen judge the small mesh of motion to be identified
Mark or track belong to i-th0Class.Here, δ is threshold value, and B is the index set of target or track class.
The tracking of moving small target under the low signal-to-noise ratio that the present invention is provided and identification problem, existing document also has similar
Research, but all there is the weakness such as tracking real-time speed is poor compared with slow, tracking or recognition effect in these literature methods.However,
The present invention with it is a kind of stage by stage, according to target provide different implementations respectively, it is proposed that it is a kind of based on single frames and multiframe when
The small target tracking algorithm of spatial domain fused filtering and recognition methods.To moving small target in video image easily by complex background
Other objects or the noise situation blocking or flood, it is proposed that the elimination of opening and closing conversion or the algorithm for weakening background and noise;
It is single using the activity of its competition to the small and weak characteristic of Small object, it is proposed that the adaptive neural network competitive model of on-line study
Member extracts the multidimensional characteristic parameter of Weak target;For the kinetic characteristic of video Small Target, using the mutation of gray scale, give
Small object motion state model and forecast model;Detection real-time to moving small target and tracking, employ fuzzy pushdown automata
Chain carries out track identification and tracking, track judgement is carried out using fuzzy pushdown automata chain depth as threshold value, so as to propose one kind
Based on the moving small target track algorithm under complex environment and recognition methods.The present invention these research contribute to target identification with
Image procossing personnel understand the characteristics of motion, active degree and its influence to other targets of detection target, so as to provide corresponding
Decision-making, seek to suppress or eliminate influence of the undesirable element to itself or other important goal to be all very important.This will be to army
Important borrow is played in the development of all target recognition and trackings based on video system such as thing, civil, public security system, road traffic
Mirror and reference role.
This moving small target tracking proposed by the present invention is with recognition methods compared with existing recognition methods, and emulation is tied
Fruit is as shown in Fig. 5 and table 1.
To every width moving small target image in video, selected in emulation to being dimensioned so as to 28Wavelet basis function according to this hair
The method of bright proposition carries out 10 repetitions and tested, and experiment every time takes the sample of different numbers respectively.The inventive method and text at present
That dedicates to out is compared with more two kinds of moving small targets tracking and method of identification, correct average when emulating 500 times
Discrimination is respectively 95.14%, 92.45%, 88.17%.Moreover, the recognition speed of the inventive method is also recognized compared with mesh first two
Method is very fast, as a result as shown in Fig. 5 and table 1.
Know from Fig. 5, the average correct recognition rata highest of the feature extraction method of identification based on the present invention.In experiment, with sample
This number increases, and average correct recognition rata constantly increases, and reaches that increasing sample curve during certain sample size again gradually tends to be flat
Surely.
In order to evaluate the combination property of each algorithm, we are according to calculating speed, amount of storage, the traffic, correct recognition rata etc.
Several aspects, using the method for fixed guantity combining with fixed quality, Integrated comparative.It has rated the quality of different personal recognition methods.Table 1 is given
Comprehensive comparison is gone out.
The Integrated comparative of the different motion Small object method of identification of table 1
Algorithm |
Correct recognition rata average |
Calculating speed |
Amount of storage |
The traffic |
Existing method 1 |
0.8817 |
0.82s |
In |
In |
Existing method 2 |
0.9245 |
0.49s |
It is relatively low |
It is relatively low |
Proposition method of the present invention |
0.9514 |
0.426s |
It is low |
It is low |
Calculating speed in table 1 is that algorithm often walks the average computations all calculated used in 10 repetition experiments under simulated environment
Time, simply the calculating time of algorithm in itself.It is the internal memory to 4,2G, programming language used to emulate the computer used
It is MATLAB.Storage and traffic demands in table 1 are the calculating process according to various algorithms, what complexity was substantially estimated.
Can be seen that amount of storage and the traffic from the result of table is closely linked system.Correct average recognition rate is that each algorithm exists in table
Under given emulation experiment ambient conditions, after 500 emulation experiments are averaged, then the average value of 10 time steps is taken.It is real
They are correct recognition rata being averaged on room and time on border, thus are the population means of correct recognition rata.
Know that faster processing speed, relatively low is not only had based on the method for identification that feature of present invention is extracted from simulation result
Amount of storage and the traffic, and also preferable recognition effect.
The implementation of the present invention is as follows:
(1) the extraction mesh calibration method from the single-frame images of video sequence is provided, weakens or eliminate the shadow of background and noise
Ring on the basis of based on mathematical morphology, it is proposed that the image object detection algorithm of opening and closing conversion.The main thought of the algorithm
It is to ask for local maximum and minimum using mathematical morphology combinatorial operation, mitigates the amount of calculation of subsequent treatment, reduce as far as possible
False alarm rate is counted.Carry out region growing, minimum point to each Local modulus maxima to implement to weaken or eliminate, to possible target
Selected.
Make to reach that it shows as the zero of large area by the filtered image of opening and closing conversion for each frame using this algorithm
High fdrequency component including background and comprising pinpoint target, and random spotted noise.Then moved according between consecutive frame
The correlation of target, carries out difference multiframe superposition, and Small object is due to its motility, and it is very strong on superposition frame to show as correlation
Tracing point, and high frequency spotted noise, then because its randomness is cancelled out each other a part, unmatched part is showed in geometric area
For the random distribution noise spot that correlation is very poor.
For example, using a kind of simple opening operation conversion g=f-f ο B, being filtered to single-frame images, figure can be detected
The place changed greatly as in, that is, high fdrequency component, while it is comparatively gentle to filter off gray-value variation in image
Place, equivalent to low frequency component.A width single-frame images can be just filtered using this conversion, filter low-frequency component, phase
When in filtering extended background, the HFS including leaving comprising Small object.In formula, f is gradation of image frame, and B is structure,
ο represents opening operation.Using opening and closing conversion, the wherein frame result to the video image of moving small target is as shown in Figure 2.
(2) Weak target extraction of motion information and status predication modeling are provided
Target is detected and tracked, a reference template is initially set up as standard form.
Given known moving target, i.e., in the two field picture containing this target under the conditions of certain video, detect
Its movement velocity, current location, direction of motion o, height h, gray average μ and the variances sigma of imaging.It assign the image of this frame as ginseng
Examine template.
Set up reference template as follows:Using target location central point as the center of circle, to clap frame time and target in imaging plane
Projection speed product for radius preceding half-circle area in, scan for.Using the change frequency of gray scale as thresholding, threshold is set
Value.Search the change frequency of gray scale respectively along image width and image height direction, if the change frequency of both direction is both less than or equal to 2,
Frame time is clapped in extension, until the grey scale change number of times at least one direction is more than 3 or 3 times.Then, using the center of circle as starting point,
Calculate the distance between continuous three adjacent grey scale changes d1And d2, can obtain the motion amplitude of target and target a frame into
Image width degree, Target Motion Character parameter s during using this amplitude and width as current location1And s2.It can be positioned by the saltus step of gray scale
The position of targetAnd the direction of motion, speed and height of imaging etc. are calculated with it.The gray average of target in half-circle area
It is another two characteristic parameter with variance, obtains a principal eigenvector V used in successive image processing special as standard
Levy vector.
By the movable information of Small object, will provide motion state forecast model is
Here, Fi(k) it is the motion state transfer matrix from previous frame to present frame, suitable turn must be selected it
Shifting formwork type,For the state estimation of previous frame,For the state estimation of present frame.
(3) incidence matrix of two inter frame image moving small targets is set up
Set up m × n object matching matrixes M.Here, m is the number of the moving small target of present frame, and n is the fortune of previous frame
The number of dynamic Small object.The value of element M (i, j) is given by.
R in formulaiFor the radius of the i-th target;rjFor the radius of jth target;(xi,yi) be the i-th target center-of-mass coordinate;(xj,
yj) be jth target center-of-mass coordinate;∞ represents a very big numerical value.
Of the moving small target and moving small target in previous frame image in present image is realized using matrix is matched
Match somebody with somebody.First, selective value is minimum in matrix M and is not ∞ element, and the row and column corresponding to the element is current kinetic respectively
The numbering of Small object and previous frame moving small target, the corresponding moving small target of such row moving small target phase corresponding with row
Match somebody with somebody.Then all elements value for the row and column for completing matching is changed into ∞.Minimum value is found in continuation in matrix M, completes motion
The matching of Small object, until all values in matrix are all changed into ∞.After search terminates, the row for not finding matching target is represented
There is the appearance of new moving small target in present image, the row for not finding matching target represent certain small mesh of motion in present image
Mark disappears.
(4) information fusion being superimposed based on multiple image, is proposed using fuzzy pushdown automata chain bullet stack recursive operation
Track algorithm and recognition methods of the large-scale image with video image motion Small object under network environment
Based on multiple features fusion, the present invention proposes description moving small target and calculated with tracking the data correlation of its movement locus
Method.Propose feature and the joint probability and the coarse number of degrees of motion state based on present frame Yu previous frame moving small target simultaneously
According to association algorithm, hypothesis testing method.Using joint probability and coarseness data association algorithm pre-estimation multimode Small object with
Track algorithm.The information updating in rear frame is utilized to assume state using multihypothesis test method.Proposing these algorithms and method
Meanwhile, provide suitable " thresholding ", only calculate the association probability and roughness of the observation being located in " thresholding " and Small object, this
Amount of calculation is considerably reduced, Small object collision problem is solved with the distance of observation position and predicted position.
The moving small target tracking and identification of image are realized, a kind of suitable recognizer must be just proposed, you can recognize
Moving small target characteristic vector for known class that is to be identified and having trained all is fuzzy number vector, by its feature to
Amount compares, defined feature membership function, degree of membership and matched rule, provides fuzzy object recognizer.It is embodied as:
To implement to track to the track of moving small target in image or video, the present invention proposes the motion shape of multiframe superposition
State blending algorithm:According to the motion state of moving small target in image or video, a kind of detect and tracking system will be designed
Blending algorithm, to realize the tracking to moving small target.This algorithm is made up of three parts:(a) mapping tracking system, which enters, has N
In the neutral net of individual feedback neural member;(b) N number of feedback neural member completion tracking system i.e. in multiframe fusion in time is allowed
State fusion;(c) M non-feedback output neurons are made to complete spatially the state of the tracking system in i.e. single frames fusion again
Fusion, finally calculates the Space Time accumulation fusion of tracking system state, come complete tracking system to the fusion of moving small target with
Track algorithm.
Designing tracking system state fusion algorithm is:Here,It is by the 1st frame to kth frame
The state estimation of i-th of Small object of the current kth frame of motion information prediction,It is all Small objects of current kth frame
State estimation,It is the fuzzy membership of model.The key of tracking is that to be carved into the k-1 moment from the outset pre- by measurement
Survey the fuzzy membership of current i-th of tracking system model of k momentHereIt is the fuzzy membership of k-1 moment tracking system models,It is known, πji
=Pr { mk=m(i)|mk-1=m(j)It is from model mk-1To model mkState transfer fuzzy membership.The tracking system shape of foundation
State merges as shown in Figure 3 to the tracking process of moving small target.
On last superposition frame, Small object point shows as a correlation very strong track.But identify Small object point
The still non-easy thing in track, mainly some discrete noise spots it is possible in the presence of.In addition it is also possible that tracing point and noise spot
It is interlaced.Therefore it can only proceed from the situation as a whole to be made decisions according to overall relevancy under certain hypothesis.Therefore, proposing to be based on
Fuzzy pushdown automata chain, track judgement is carried out by thresholding of the correlation length of track using the method for playing stack Yu popping.Tracking
As a result it is as shown in Figure 4.From fig. 4, it can be seen that explanation fuzzy pushdown automata D-chain trace algorithm proposed by the present invention is known in pattern
Larger success is achieved in other system.
The fuzzy membership of each component of the characteristic vector of moving small target to be identified or track is calculated according to blending algorithm
μij, you can the characteristic vector for obtaining Unknown Motion Small object or track is Ui=[μi1,μi2,…,μik]T.It with trained
Good known i-th0The multi-Dimensional parameters characteristic vector U of classificationi0Compare, and if only ifWhen, adjudicate fortune to be identified
Dynamic Small object or track belong to i-th0Class.EvenSo thatThen judge the small mesh of motion to be identified
Mark or track belong to i-th0Class.Here, δ is threshold value, and B is the index set of target or track class.
The innovative point of the present invention is as follows:
(1) how to set up data correlation to two inter frame image Small objects in video sequence is moving small target track algorithm
Major Difficulties.The complexity of background under network environment, ratio that Small object is occupied in the picture is small, color of object and background color
Similarity degree, the degree of stability of background, the interaction of multiple target and the generation of various special circumstances all can to moving small target with
Track brings difficulty.The information such as Small object external appearance characteristic such as target shape and texture, because the generation for the process of blocking is several in the picture
It is submerged, and the uncertainty that Small object is moved, cause the loss of Small object information, it is easy to occur tracking failure.Such as
What effective processing is blocked, and particularly serious blocks, and is always a difficult point in moving small target tracking.In monitor video
In, the Small object outward appearance in each frame is often closely similar, how to choose suitable feature and is transported with preferably distinguishing different Small objects
Dynamic state realizes accurate data correlation, is the key issue that this project will be studied.And at present both at home and abroad to this
The research of problem is almost a blank.
Discussion to this problem, the imaging characteristicses and movable information of present invention analysis Small object in the picture are right first
The small and weak and kinetic characteristic of target, removes background using opening and closing conversion, eliminates or reduces the processing of noise scheduling algorithm.Then according to phase
The motion relevance of moving small target between adjacent frame, carries out difference multiframe superposition.
(2) using moving small target characteristic in itself, movable information vector is set up, to extract the motion of Small object multi-characteristic points small
It is the intractable and urgently to be resolved hurrily problem of present many departments that the multi-Dimensional parameters of target, which are extracted,.Present invention utilizes moving small target sheet
The characteristic of body, excavates the characteristic point of each characteristic parameter of moving small target and characteristic parameter, sets up kinematics character vector, gives
The multi-Dimensional parameters of moving small target are extracted.
(3) track is carried out using the information fusion and fuzzy pushdown automata chain recursive operation that are superimposed based on multiple image to sentence
Certainly, set up model algorithm and feature based chooses parallel mechanism and carries out image recognition.
With science and technology development and the mankind security protection consciousness lifting, video monitoring system is in every field under network environment
Increasingly it is widely applied., can be with and the identification of the moving target based on video monitoring is a particularly useful job
Apply in space flight, military affairs, the tracking of guided missile track identification, the various fields such as break in traffic rules and regulations detection.But in for example infrared system of some occasions
In leading, it is desirable to be able to intercept and capture as soon as possible and locking tracking target.So to the accurate detection and tracking of moving small target, in army
The application of the every field such as thing, civil seems more and more important, also more and more urgent.It is low under the conditions of the strong clutter background of network environment
The test problems of signal to noise ratio moving small target directly determine the operating distance and detection performance of detection system, and it is solved for improving
Detection system performance has very important practical significance.
For the research of these problems, the present invention analyzes the various features of polymorphic target, utilizes tracking system state
Fusion, fuzzy pushdown automata chain bullet stack recursive operation carry out track following to the moving small target of video image, provide motion
Small object recognizer.
The specific embodiment of the present invention:
The tracking of moving small target under the low signal-to-noise ratio that the present invention is provided and identification problem, existing document also has similar
Research, but all there is the weakness such as tracking real-time speed is poor compared with slow, tracking or recognition effect in these methods.However, this hair
It is bright with it is a kind of stage by stage, according to target provide different implementations respectively, it is proposed that a kind of time-space domain based on single frames and multiframe
The small target tracking algorithm of fused filtering and recognition methods (being based on such as Fig. 3).1. to the easy quilt of moving small target in video image
The situation that other objects or noise in complex background are blocked or flooded, it is proposed that the elimination of opening and closing conversion weakens background with making an uproar
The algorithm of sound;2. it is competing using its to the small and weak characteristic of Small object, it is proposed that the adaptive neural network competitive model of on-line study
The active unit striven extracts the multidimensional characteristic parameter of Weak target;3. for the kinetic characteristic of video Small Target, gray scale is utilized
Mutation, give Small object motion state model and forecast model;4. detection real-time to moving small target and tracking, are employed
Fuzzy pushdown automata chain carries out track identification and tracking, and track judgement is carried out using fuzzy pushdown automata chain depth as threshold value,
So as to propose a kind of moving small target track algorithm based under complex environment and recognition methods.
It is to the implementation of technical scheme 1.:
G=f-f ο B or closed operation conversion g=fB-f are converted using opening operation, single-frame images is filtered, detected
The place changed greatly in image, that is, high fdrequency component, while it is comparatively gentle to filter off gray-value variation in image
Place, equivalent to low frequency component, just a width single-frame images is filtered using this conversion, low-frequency component is filtered, quite
In filtering extended background, the HFS including leaving comprising Small object;In formula, f is gradation of image frame, and B is structure, ο
Opening operation is represented, closed operation is represented.
It is inadequate only to single frames processing, it is necessary to carry out Small object enhancing and interference to accurately identify target or track
Suppress, because Small object point is moved between each frame, with very strong correlation, therefore it is folded that multiframe can be carried out to video image
Plus.In the last frame of superposition, Small object point shows as correlation very strong tracing point, but noise is it is possible to flood Small object
Track, therefore propose multi-frame difference superposition algorithm.Its algorithm is to choose the image sequence comprising including moving small target point, and its is strange
Number frame and the superposition value of each n frames difference of even frame.I.e.
In formula, fiFor the i-th frame in image sequence, fzFor last superposition frame.
So after multi-frame difference is superimposed, Small object point is further enhanced, and extended background can also be obtained simultaneously
Suppress to further, but noise spot can also may be strengthened, and therefore, it can through the stack plus frame takes thresholding to handle.
Its method is as follows
Here, δ is threshold value, is taken according to many experimentsM, N are the chi of superposition two field picture
It is very little.
After so handling, Small object point is further enhanced, and random noise point further can be suppressed or be disappeared
Remove.
It is to the implementation of technical scheme 2.
The adaptive neural network competitive model of on-line study is constructed first, extracts weak using the active unit of its competition
The multidimensional characteristic parameter of Small object.Its method is as follows:
P1. network is initialized:The dimension of fixed output nerve network grid is N × M, and input layer is quadravalence network, and
Random initializtion inputs the weight of neuron and output neuron connection.Make t represent algorithm iteration number of times, put t=0.
P2. victor is selected:The parameter value X={ x such as gray scale, colourity, motion each frame Small object image1,x2, …,
xdThe input neuron in network is input to, to each input neuron value xj, the output of the node i in competition layer
ForHere G is an activation primitive, is such as takenα > 0 are constants, can be with controlling curve
Slope.μi (t) it is p dimensional input vectors xjWith p dimensional weights vector ωji (t) Euclidean distance between | | xj-ωji (t)
| | and, i.e.,ωji (t) it is from input layer node j to competition layer node i connection weight in t
Weight is vectorial, here j ∈ J, J={ 1 ..., d }, i ∈ I, I={ 1 ..., NiBe competition layer certain regional area.
Select the output neuron i won* .In competition layer, correspondenceMinimum node will win, if that is,The node so won in competition layer is i*, then with i* The weight of association and and i* The neighbouring point association of point
Weight can be all adjusted.
P3. weight is updated:If N (i*) it is triumph output neuron i*Neighbour, the distance between output neuron is specific
Specify.To each output neuron i ∈ { N (i*),i*, weight adjusts renewal according to the following formula.
Here η (t)=η has determined that in advance.This rule only updates the neighbour of triumph output neuron.
P4. standardized weight:It is standardized after updating weight, so that they and input measurement standard are consistent.
P5. continue cycling through:Repeat step P1 to P4, the number of times of iteration is set to t=t+1, and criterion is shut down until meeting,
Here shutting down criterion is | | xj-ωji (t) | | < ε, take ε=0.5, or untill having exceeded the cycle-index of maximum.
It is to the implementation of technical scheme 3.:
Gray scale mutation algorithm is as follows:
If being bianry image original image pretreatment, the threshold value utilized is 1, calculates transition times.To horizontally and vertically side
The image Small object separation calculation implemented to Gray Level Jump is as follows:
It is determined by experiment the high m and width n of image.On the one hand, in the case of a certain fixing point of image height, i.e. height=mi
When, look for transition times c along image width directioni, i=1 ..., m.Known according to the experiment to Small object and background area, Small object
The transition times in region than it is larger the characteristics of, by transition times ci, give a threshold value M1.Any transition times are more than this threshold
Value, i.e. ci> M1, Gao Yukuan at this moment is recorded, y is designated as respectivelyk1And xk1, k1∈N .Obtain ordered pair collection
On the other hand, in the case of a certain fixing point of image width, i.e. width=njWhen, look for transition times along image height direction
dj, j=1 ..., n.By transition times dj, give a threshold value M2.Any transition times are more than this threshold value, i.e. dj> M2
, width and height at this moment is recorded, is designated as respectivelyWithk2∈N .Obtain ordered pair collection
By (I) and (II), ordered pair collection I=I can be tried to achieve1 ∪I2={ (xk,yk),k ∈N}.According to (xk,yk), can
Obtain xkAnd ykThe curve of satisfaction, implements the modeling of Small object curve movement, can obtain Small object i curve movements si.Claim this motion bent
The algorithm of line modeling is Gray Level Jump algorithm.
Using the mutation of gray scale, Small object motion state model s is giveni, while can also provide forecast model ViIt is as follows:
The characteristics of present invention is proposed using gray value saltus step, in the horizontal direction, vertical direction, 1/4 arc direction, is determined fortune
Direction, distance and the motion feature of dynamic Small object.Using pixel i as summit, the window that a size is r × r is opened in its neighborhood
Mouthful.Calculate the gray average μ of this windowi With variances sigmai, it is used as pixel i two characteristic parameters.Using this pixel i as circle
The heart, r is that radius makees a quarter turn, labeled as Θi (1/4) .Along image width radial direction, image height radial direction, 1/4 camber line side
To the change frequency c for finding out gray scale respectivelyh、cv、c1/4, these three parameters are defined as pixel i direction character vector, are designated as
diri=(cih,civ,ci,1/4) .Again along 1/4 arc direction, the position Q of Gray Level Jump is markedkl(xk,yl), k, l ∈ N are calculated
The distance of adjacent two trip pointK, l=1,2 ..., and the gray value at any point between adjacent two trip point is calculated simultaneously
G, G '.Calculate the close distance of gray valueAverage valueIn conjunction with motion state parameterses si, then picture can be obtained
Vegetarian refreshments i characteristic vector is
It is to the implementation of technical scheme 4.:
Tracking system is mapped to enter in the neutral net of the fuzzy pushdown automata with N number of feedback neural member;Make N individual
Feedback neural member is the state fusion that tracking system is completed in multiframe fusion in time;Make M non-feedback output neurons again
To complete spatially the state fusion of the tracking system in i.e. single frames fusion, the Space Time for finally calculating tracking system state is tired out
Product fusion, to complete fusion tracking algorithm of the tracking system to moving small target:
1. each fuzzy pushdown automata recognizes fusion in time to moving small target
If fipAnd Ξ (t)i (t) represent that t is belonged to by obscuring the identified moving small target that pushdown automata i is measured respectively
In the fuzzy membership and Fuzzy Distribution of pth class,Represent untill the l moment by i-th of fuzzy pushdown automata accumulation fusion
Obtained identified target belongs to the fuzzy membership of pth class,Represent untill the l moment by i-th of fuzzy pushdown automata
The Fuzzy Distribution for the identified target that accumulation fusion is obtained, here, l=1,2 ..., t, i.e.,
With
Here, op(p ∈ U) is moving small target.By the accumulation fusion Fuzzy Distribution at t-1 moment and the measurement at t moment
Fuzzy Distribution is merged, and can obtain target identification accumulation fusion Fuzzy Distribution of i-th of fuzzy pushdown automata untill tFor:
Wherein,S2It is Fuzzy Integration Function, usual S2Remove formula.
Now, with Fuzzy DistributionThe motion state of corresponding Small object isIt is the small mesh of current t
Target state estimation.
Here, Fi(t) it is the motion state transfer matrix from last moment to current time, must be suitable to selecting
Metastasis model,For the state estimation of last moment,For the state estimation at current time.
2. the Space integration that pushdown automata is recognized to Small object is obscured
Obtaining the accumulation Fuzzy Distribution of t moment each fuzzy pushdown automata target identificationAfterwards, i=1 here,
..., N is merged to this N number of Fuzzy Distribution using Fuzzy Integration Function, just obtained untill the t moment to target identification
Fuzzy Distribution is merged in Space Time accumulation
It is theoretical using Fuzzy Integration Function, it can obtain
Here, SNAlso illustrate that Fuzzy Integration Function.IfP=1 ..., M.
Now, with Fuzzy Distribution Ξt The motion state of corresponding Small object isIt is all small of current kth frame
The state estimation of target.Motion state fusion results are:Here,It is by the 1st frame to kth frame
The current kth frame of motion information prediction i-th of Small object state estimation,It is all small mesh of current kth frame
Target state estimation,It is the fuzzy membership of model;The key of tracking is when being carved into k-1 from the outset by measurement
Carve the fuzzy membership of current i-th of the tracking system model of k moment of predictionHereIt is the fuzzy membership of k-1 moment tracking system models,Be it is known,
πji=Pr { mk=m(i)|mk-1=m(j)It is from model mk-1To model mkState transfer fuzzy membership.
So after multiframe is merged, Small object point is further enhanced, and most of noise spot is cut, but some with
Machine noise jamming may be strengthened, in order to filter out these noises, can be sentenced by fuzzy pushdown automata chain length
Certainly.In order to reduce fuzzy pushdown automata chain length as far as possible while Small object point is retained, threshold value is taken to fusion frame
Change is handled.Its method is as follows:
Threshold value T is the length of fuzzy pushdown automata chain, but this chain must assure that containing Small object point.Here fATo be total
Merge frame.On last fusion frame, Small object point shows as a correlation very strong track.
Each component f of the characteristic vector of moving small target to be identified or track is calculated according to blending algorithmAFuzzy membership
Spend μij, that is, the characteristic vector for obtaining Unknown Motion Small object or track is Ui=[μi1 , μi2 , …, μik ]T;It with
Trained good known i-th0The multi-Dimensional parameters characteristic vector of classificationCompare, and if only ifWhen, judgement is treated
Identification moving small target or track belong to i-th0Class;EvenSo thatThen judge fortune to be identified
Dynamic Small object or track belong to i-th0Class;Here, δ is threshold value, and B is the index set of target or track class.
The tracking of moving small target under the low signal-to-noise ratio that the present invention is provided and identification problem, it is proposed that one kind is based on single frames
Small target tracking algorithm and recognition methods with the time-space domain fused filtering of multiframe.To the easy quilt of moving small target in video image
The situation that other objects or noise in complex background are blocked or flooded, it is proposed that the elimination of opening and closing conversion weakens background with making an uproar
The algorithm of sound;To the small and weak characteristic of Small object, it is proposed that the adaptive neural network competitive model of on-line study, its competition is utilized
Active unit extract Weak target multidimensional characteristic parameter;For the kinetic characteristic of video Small Target, the prominent of gray scale is utilized
Become, give Small object motion state model and forecast model;Detection real-time to moving small target and tracking, are employed under obscuring
Push away automatic chain and carry out track identification and tracking, track judgement is carried out using fuzzy pushdown automata chain depth as threshold value, so as to carry
A kind of moving small target track algorithm based under complex environment and recognition methods are gone out.These researchs of the present invention contribute to mesh
Mark does not understand the characteristics of motion, active degree and its influence to other targets of detection target with image procossing personnel, so that
Corresponding decision-making is provided, seeks to suppress or eliminate influence of the undesirable element to itself or other important goal to be all very important.
This plays the development to all target recognition and trackings based on video system such as military, civil, public security system, road traffic
Important reference and reference role.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention
Any modifications, equivalent substitutions and improvements made within refreshing and principle etc., should be included in the scope of the protection.