CN104835178A - Low SNR(Signal to Noise Ratio) motion small target tracking and identification method - Google Patents

Abstract

The invention discloses a low SNR(Signal to Noise Ratio) motion small target tracking and identification method. The method comprising the following steps: providing a method of extracting a target from a single frame image of a video sequence, and influences of backgrounds and noises can be reduced or eliminated; providing weak target motion information extracting and state predicting modeling; establishing an incidence matrix of the image motion small target between two frames; based on information fusion of overlapped multi-frame images, providing a large-scale image and video image motion small target tracking algorithm and an identification method by using fuzzy push-down automation chain slot stack recursion calculation. The low SNR(Signal to Noise Ratio) motion small target tracking and identification method is advantageous in that the target identification and image processing personnel can be aware of the target motion law, the active level, and the influence on other targets conveniently, and then can give the corresponding decisions, and therefore the searching, the inhibiting, and the eliminating of the influences of the bad factors on other important targets can be very necessary, and in addition, the important experiences and the important references can be provided for the target identification and tracking in the military field, the civil field, the public security system, and the road traffic field based on the video system.

Description

A kind of method of recognition and tracking of low signal-to-noise ratio moving small target

Technical field

The invention belongs to pattern recognition and classification technical field, particularly relate to a kind of method of recognition and tracking of low signal-to-noise ratio moving small target.

Background technology

Along with the lifting that the development of science and technology and the security protection of the mankind are realized, under network environment, video monitoring system obtains applying more and more widely in every field, as military affairs, traffic, bank, factory, community etc.And be a very useful job based on the identification of the moving target of video monitoring, space flight can be applied in, military affairs, guided missile track identification follow the tracks of, the various fields such as break in traffic rules and regulations detection.But in some occasion if countries in the world are in the environmental surveillance of surrounding area, needs can be intercepted and captured and locking tracking target as soon as possible.So to the accurate detection and tracking of moving small target, the application in the every field such as military, civil is seemed more and more important, also more and more urgent.Under the strong clutter background condition of network environment, the test problems of low signal-to-noise ratio Moving Small Targets directly determines operating distance and the detection perform of detection system, and it solves has very important practical significance for raising detection system performance.

At present the denoising of video image under complex background, the multi-Dimensional parameters of moving target are extracted, the process of the problem such as the recognition and tracking of Computer Vision and moving target lacks solution, these problems have become a hot issue of image processing field, and this is also the thorny and problem demanding prompt solution of present many departments.

Because different aerial video acquisition systems, different physical phenomenons are as many-sided reasons such as illumination can not be uniformly distributed completely, make the image border intensity of acquisition different.And, in practical matter, view data toward contact pollute by noise.Scenery characteristic mixes and explanation subsequently can be made to become very difficult simultaneously.Realize accurately understanding aerial image picture intention, need research can detect the noncontinuity of image object intensity, the target identification method of their exact position can be determined again simultaneously.

Under low signal-to-noise ratio, moving small target acquisition of signal is that current various advanced seeker system needs one of key technical problem solved with extracting, now the image of target only occupies the area of one or several pixel, and due to background environment complexity, the impact of the unevenness of atmosphere radiation, internal noise of detector etc. factor, target is almost submerged in clutter varying background, not there is shape and structural information, sometimes even possible lose objects, this just brings very large difficulty to Dim targets detection.Under strong clutter background condition, the test problems of low signal-to-noise ratio Moving Small Targets directly determines operating distance and the detection perform of detection system, and it solves has very important practical significance for raising detection system performance.Also important reference and reference role is played by the development of all target recognition and trackings based on video system such as military, civil, public security system, road traffic.

Summary of the invention

The object of the embodiment of the present invention is the method for the recognition and tracking providing a kind of low signal-to-noise ratio moving small target, be intended to solve prior art exist follow the tracks of real-time speed comparatively slow, follow the tracks of or problem that recognition effect is poor.

The embodiment of the present invention realizes like this, a kind of method of recognition and tracking of low signal-to-noise ratio moving small target, it is characterized in that, the method of the recognition and tracking of this low signal-to-noise ratio moving small target provides different implementation methods stage by stage, according to target respectively with a kind of, to moving small target in video image easily by the situation that other objects in complex background or noise block or flood, propose the elimination of opening and closing conversion or weaken the algorithm of background and noise; To the small and weak characteristic of Small object, propose the adaptive neural network competitive model of on-line study, utilize the active unit of competition to extract the multidimensional characteristic parameter of Weak target; For the kinetic characteristic of video Small Target, utilize the sudden change of gray scale, give Small object motion state model and forecast model; To the real-time detection and tracking of moving small target, have employed fuzzy pushdown automata chain and carry out track identification and tracking, with the fuzzy pushdown automata chain degree of depth for threshold value carries out track judgement;

Specifically comprise the following steps:

Step one, provides and extract order calibration method from the single-frame images of video sequence, weakens or eliminate the impact of background and noise;

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 target;

Step 4, based on the information fusion of multiple image superposition, utilizes fuzzy pushdown automata chain bullet stack recursive operation to propose track algorithm and the recognition methods of large-scale image and video image motion Small object under network environment.

Further, in step one, provide and extract order calibration method from the single-frame images of video sequence, when weakening or eliminate the affecting of background and noise, concrete methods of realizing is:

Mathematical morphology combinatorial operation is utilized to ask for local maximum and minimal value, alleviate the calculated amount of subsequent treatment, reduce false alarm rate to count as far as possible, region growing is carried out to each Local modulus maxima, minimum point enforcement weakens or eliminates, possible target is selected;

Use opening operation conversion g=f-f ο B or closed operation conversion g=f B-f, filtering is carried out to single-frame images, the place changed greatly in image detected, namely be equivalent to high fdrequency component, the place that in the image of energy elimination simultaneously, gray-value variation is comparatively mild, is equivalent to low frequency component, this conversion is utilized just to carry out filtering to a width single-frame images, filter low-frequency component, be equivalent to filter extended background, leave the HFS comprising Small object; In formula, f is gradation of image frame, and B is structure, and ο represents opening operation, and represents closed operation;

For accurately identifying target or track, carry out the suppression of Small object enhancing and interference, because Small object point moves between each frame, multiframe superposition can be carried out to video image, on the last frame of superposition, Small object point shows as the very strong tracing point of correlativity, but noise still likely floods Small object track, propose multi-frame difference superposition algorithm, choose the image sequence comprising moving small target point, the superposition value of odd-numbered frame and even frame each n frame difference, that is:

$f_z=\underset{i=0}{\overset{n}{\mathrm{\Σ}}}(f_{2i+1}-f_{2i})$

In formula, f _ifor the i-th frame in image sequence, f _zfor last superposition frame;

Add frame through the stack and take thresholding process, method is as follows:

$f_z(i,j)=\left\{\begin{array}{cc}{f}_z(i,j),& f_z(i,j)\&GreaterEqual;\mathrm{\&delta;}\\ 0,& f_z(i,j)<\mathrm{\&delta;}\end{array}\right.$

δ is threshold value, gets m, 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, concrete methods of realizing is:

First construct the adaptive neural network competitive model of on-line study, the active unit utilizing it to compete extracts the multidimensional characteristic parameter of Weak target:

The first step, initialization network: the dimension of fixing output nerve network grid is N × M, and input layer is quadravalence network, and the weight that random initializtion input neuron is connected with output neuron, make t represent algorithm iteration number of times, put t=0;

Second step, selects victor: gray scale, colourity, the motion parameter X={x of each frame Small object image ₁, x ₂..., x _dbe input to input neuron in network, to each input neuron value x _j, the output of the node i in competition layer for $S_i^t=G\left({\mathrm{\&mu;}}_i\left(t\right)\right),$ G is an activation function, as got $G\left(x\right)=\frac{1}{1+e^{-\mathrm{\&alpha;x}}},$ α > 0 is constant, the slope of controlling curve; μ _it () is that p ties up input vector x _jwith p dimensional weight vector ω _jieuclidean distance between (t) || x _j-ω _ji(t) || with, namely ω _jit () is in t from input layer node j to the connection weight of competition layer node i vector, j ∈ J, J={1 ..., d}, i ∈ I, I={1 ..., N _iit is certain regional area of competition layer;

Select the output neuron i won ^*, in competition layer, corresponding minimum node will be won, if namely the node of so winning in competition layer is i ^*, then with i ^*the weight of association and and i ^*the weight of the some association that point is contiguous all can be adjusted;

3rd step, upgrades weight: N (i ^*) be triumph output neuron ^i*neighbour, specifically specified by the distance between output neuron, to each output neuron i ∈ { N (i ^*), i ^*, adjust renewal according to the following formula:

${\mathrm{\&omega;}}_{\mathrm{kj}}(t+1)=\left\{\begin{array}{cc}{\mathrm{\&omega;}}_{\mathrm{kj}}\left(t\right)+\mathrm{\&eta;}\left(t\right)(x_k-{\mathrm{\&omega;}}_{\mathrm{kj}}\left(t\right)),\mathrm{if}& j\&Element;N_i\\ 0& ,\mathrm{otherwise}\end{array}\right.$

η (t)=η determines in advance; This rule only upgrades the neighbour of triumph output neuron;

4th step, standardized weight: to standardization after renewal weight, thus be consistent with input measurement standard;

5th step, continues circulation: repeat the first step to the 4th step, the number of times of iteration is set to t=t+1, shuts down criterion until meet, and shutting down criterion is || x _j-ω _ji(t) || < ε, get ε=0.5, or till having exceeded maximum cycle index.

Further, in step 3, the implementation method setting up the incidence matrix of two inter frame image moving small target is:

Set up m × n object matching matrix M, here, m is the number of the moving small target of present frame, and n is the number of the moving small target of previous frame, and the value of element M (i, j) is given by following formula:

$M(i,j)=\left\{\begin{array}{ccc}|r_i-r_j|,& \mathrm{if}& r_i+r_j>\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}\\ \&infin;,& & \mathrm{else}\end{array}\right.$

R in formula _iit is the radius of the i-th target; r _jfor the radius of jth target; (x _i, y _i) be the center-of-mass coordinate of the i-th target; (x _j, y _j) be the center-of-mass coordinate of jth target; ∞ represents a very large numerical value;

First, in matrix M, selective value is minimum and be not the element of ∞, row and column corresponding to this element is the numbering of current kinetic Small object and previous frame moving small target respectively, the moving small target that the moving small target of row correspondence is corresponding with row matches, and then all elements value of the row and row that complete coupling is become ∞; Continue to find minimum value in matrix M, complete the coupling of moving small target, until all values in matrix all becomes ∞; After search terminates, do not find the row representative of coupling target in present image, have the appearance of new moving small target, do not find row representative certain moving small target in present image of coupling target to disappear.

Further, in step 4, based on the information fusion of multiple image superposition, fuzzy pushdown automata chain bullet stack recursive operation is utilized to propose the track algorithm of large-scale image and video image motion Small object under network environment and recognition methods is:

The first step, each fuzzy pushdown automata is identified in temporal fusion to moving small target:

F _ip(t) and Ξ _iwhat t () represented that t records by fuzzy pushdown automata i respectively is identified fuzzy membership and the Fuzzy Distribution that moving small target belongs to p class, represent to the l moment, accumulate that fusion obtains by i-th fuzzy pushdown automata be identified the fuzzy membership that target belongs to p class, represent the Fuzzy Distribution being identified target obtained by i-th fuzzy pushdown automata accumulation fusion to the l moment, here, l=1,2 ..., t, namely

${\mathrm{\&Xi;}}_i\left(t\right)=\underset{p\&Element;U}{\mathrm{\&Sigma;}}{f}_{\mathrm{ip}}\left(t\right)/o_p$

With

${\mathrm{\&Xi;}}_i^{t-1}=\underset{p\&Element;U}{\mathrm{\&Sigma;}}{f}_{\mathrm{ip}}^{(t-1)}/o_p$

O _p(p ∈ U) is moving small target, the measurement Fuzzy Distribution that Fuzzy Distribution and t are merged in the accumulation in t-1 moment is merged, and obtains the target identification accumulation of i-th fuzzy pushdown automata to t and merges Fuzzy Distribution for:

${\mathrm{\&Xi;}}_i^t=\underset{p\&Element;U}{\mathrm{\&Sigma;}}{f}_{\mathrm{ip}}^{\left(t\right)}/o_p$

Wherein, s ₂fuzzy Integration Function, usual S ₂take off formula:

$S_t\left(M_{\mathrm{ip}}\left(t\right)\right)={\left(\frac{1}{t}\underset{l=1}{\overset{t}{\mathrm{\&Sigma;}}}{f}_{\mathrm{ip}}^q\left(l\right)\right)}^{\frac{1}{q}},q>0$

Now, with Fuzzy Distribution the motion state of corresponding Small object is the state estimation of the Small object of current t:

${\stackrel{\mathrm{\&Lambda;}}{x}}_{t+1|t}^{\left(i\right)}={\stackrel{\mathrm{\&Lambda;}}{x}}^{\left(i\right)}(t+1|t)=F_i\left(t\right){\stackrel{\mathrm{\&Lambda;}}{x}}^{\left(i\right)}\left(t\right|t)$

F _it (), for being carved into the motion state transition matrix of current time from upper a period of time, selects metastasis model, for the state estimation in a upper moment, for the state estimation of current time;

Second step, fuzzy pushdown automata is to the Space integration of Small object identification:

Obtaining the accumulation Fuzzy Distribution of t each fuzzy pushdown automata target identification afterwards, i=1 here ..., N, utilizes Fuzzy Integration Function to merge this N number of Fuzzy Distribution, merges Fuzzy Distribution to be till just obtaining t to the Space Time accumulation of target identification:

${\mathrm{\&Xi;}}^t=\underset{p\&Element;U}{\mathrm{\&Sigma;}}{f}_p^{\left(t\right)}/o_p$

Utilize Fuzzy Integration Function theoretical, can obtain

$f_p^{\left(t\right)}=S_N[f_{1p}^{\left(t\right)},f_{2p}^{\left(t\right)},...,f_{\mathrm{Np}}^{\left(t\right)}]$

S _nalso Fuzzy Integration Function is represented; If

Now, with Fuzzy Distribution Ξ ^tthe motion state of corresponding Small object is be the state estimation of all Small object of current kth frame, motion state fusion results is: by the state estimation of the 1st frame to i-th Small object of the current kth frame of motion information prediction of kth frame, the state estimation of all Small object of current kth frame, it is the fuzzy membership of model; The key of following the tracks of is carved into by measuring the fuzzy membership that the k-1 moment predicts current k moment i-th tracker model from the outset ${\mathrm{\&mu;}}_{k|k-1}^{\left(i\right)}\stackrel{\mathrm{\&Delta;}}{=}\mathrm{Pr}\left\{m_k^{\left(i\right)}\right|z^{k-1}\}=\underset{j}{\mathrm{\&Sigma;}}{\mathrm{\&pi;}}_{\mathrm{ji}}{\mathrm{\&mu;}}_{k-1}^{\left(j\right)},$ Here ${\mathrm{\&mu;}}_{k-1}^{\left(j\right)}=\mathrm{Pr}\left\{m_{k-1}^{\left(j\right)}\right|z^{k-1}\}$ The fuzzy membership of k-1 moment tracker model, known, π _ji=Pr{m _k=m ⁽ⁱ⁾| m _k-1=m ^(j)from model m _k-1to model m _kstate transfer fuzzy membership.

Further, after multiframe merges, Small object point is further enhanced, major part noise spot is cut, filtering random noise disturbance is adjudicated by fuzzy pushdown automata chain length, in order to reduce fuzzy pushdown automata chain length as far as possible while reservation Small object point, take thresholding process to fusion frame, method is as follows:

$f_A(i,j)=\left\{\begin{array}{cc}{f}_A(i,j),& f_A(i,j)\&GreaterEqual;T\\ 0,& f_A(i,j)<T\end{array}\right.$

Threshold value T is the length of fuzzy pushdown automata chain, but this chain must ensure containing Small object point, f _afor total fusion frame, on last fusion frame, Small object point shows as the very strong track of a correlativity;

The each component f of proper vector of moving small target to be identified or track is calculated according to blending algorithm _afuzzy membership μ _ij, the proper vector namely obtaining Unknown Motion Small object or track is U _i=[μ _i1, μ _i2..., μ _ik] ^t; It with trained known i-th ₀the multi-Dimensional parameters proper vector U of classification _i0compare, and if only if time, adjudicate moving small target to be identified or track belongs to i-th ₀class; Even make then judge that moving small target to be identified or track belong to i-th ₀class; Here, δ is threshold value, and B is the index set of target or track class.

The method of the recognition and tracking of low signal-to-noise ratio moving small target provided by the invention, propose a kind of small target tracking algorithm and recognition methods of the time-space domain fused filtering based on single frames and multiframe, to moving small target in video image easily by the situation that other objects in complex background or noise block or flood, propose the elimination of opening and closing conversion or weaken the algorithm of background and noise; To the small and weak characteristic of Small object, propose the adaptive neural network competitive model of on-line study, the active unit utilizing it to compete extracts the multidimensional characteristic parameter of Weak target; For the kinetic characteristic of video Small Target, utilize the sudden change of gray scale, give Small object motion state model and forecast model; To the real-time detection and tracking of moving small target, have employed fuzzy pushdown automata chain and carry out track identification and tracking, with the fuzzy pushdown automata chain degree of depth for threshold value carries out track judgement, thus propose a kind of based on the moving small target track algorithm under complex environment and recognition methods, the present invention contribute to target identification and image procossing personnel understand detect target the characteristics of motion, active degree and the impact on other targets thereof, thus provide corresponding decision-making, seek to suppress or eliminate the impact of undesirable element on itself or other important goal and be all very important; Important reference and reference role is played by the development of all target recognition and trackings based on video system such as military, civil, public security system, road traffic.The average correct recognition rata of feature extraction method of identification of the present invention is the highest, and in experiment, along with increasing of sample number, average correct recognition rata constantly increases, and increases sample curve again and tend to be steady gradually when reaching certain sample size.

Accompanying drawing explanation

Fig. 1 is the realization flow figure of the recognition and tracking method of moving small target under the low signal-to-noise ratio that provides of the embodiment of the present invention;

Fig. 2 is the schematic diagram of the filter result of the opening and closing transfer pair moving small target image that the embodiment of the present invention provides;

In figure: the 5th frame of (a) image sequence; (b) the 5th frame opening and closing mapping algorithm filter effect;

Fig. 3 be the embodiment of the present invention provide set up the tracking of tracker state fusion to moving small target in image or video;

Fig. 4 is that the fuzzy pushdown automata chain that provides of the embodiment of the present invention is to the pursuit path image of moving small target;

Fig. 5 be the proposition that provides of the embodiment of the present invention compare schematic diagram with the correct recognition rata of existing method of identification.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with embodiment, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

Below in conjunction with drawings and the specific embodiments, application principle of the present invention is further described.

As shown in Figure 1, the method for the recognition and tracking of the low signal-to-noise ratio moving small target of the embodiment of the present invention comprises the following steps:

Step S101, provides and extract order calibration method from the single-frame images of video sequence, weakens or eliminate the impact of background and noise;

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 target;

Step S104, based on the information fusion of multiple image superposition, utilizes fuzzy pushdown automata chain bullet stack recursive operation to propose track algorithm and the recognition methods of large-scale image and video image motion Small object under network environment.

In embodiments of the present invention, in step S101, provide and extract order calibration method from the single-frame images of video sequence, when weakening or eliminate the affecting of background and noise, concrete methods of realizing is:

Based on the basis of mathematical morphology, propose the image object detection algorithm of opening and closing conversion.The main thought of this algorithm utilizes mathematical morphology combinatorial operation to ask for local maximum and minimal value, alleviates the calculated amount of subsequent treatment, reduces false alarm rate as far as possible and count.Region growing is carried out to each Local modulus maxima, minimum point enforcement weakens or eliminates, possible target is selected.

Utilize this algorithm to make to reach and convert filtered image for each frame through opening and closing it shows as large-area zero background and comprises pinpoint target, and random spotted noise is in interior high fdrequency component.Then according to the correlativity of moving target between consecutive frame, carry out the superposition of difference multiframe, Small object is due to its motility, superposition frame shows as the very strong tracing point of correlativity, and high frequency spotted noise, then because its randomness is cancelled out each other a part, unmatched part shows as the very poor stochastic distribution noise spot of correlativity in geometric area.

Such as, use simple a kind of opening operation conversion g=f-f ο B, filtering is carried out to single-frame images, the place changed greatly in image can be detected, namely be equivalent to high fdrequency component, the place that in the image of energy elimination simultaneously, gray-value variation is comparatively mild, is equivalent to low frequency component.Utilize this conversion just can carry out filtering to a width single-frame images, filter low-frequency component, be equivalent to filter extended background, leave the HFS comprising Small object.In formula, f is gradation of image frame, and B is structure, and ο represents opening operation.Use opening and closing converts, to the wherein frame result of the video image of moving small target as shown in Figure 2.

In embodiments of the present invention, in step s 102, when providing Weak target extraction of motion information and status predication modeling, concrete methods of realizing is:

Detection and tracking is carried out to target, first sets up a reference template as standard form.

Given known moving target, namely under certain video condition containing in a two field picture of this target, detect its movement velocity, current location, direction of motion o, height h, the gray average μ of imaging and variances sigma.Using the image of this frame as reference template.

Set up reference template as follows: with target location central point for the center of circle, to clap in front half-circle area that the product of frame time and the projection speed of target in imaging plane is radius, search for.With the change frequency of gray scale for thresholding, threshold value is set.Search the change frequency of gray scale along image width and image height direction respectively, if the change frequency of both direction is all less than or equal to 2, extends and clap frame time, until the grey scale change number of times at least one direction is more than 3 or 3 times.Then, be starting point with the center of circle, calculate the distance d between continuous three adjacent grey scale change ₁and d ₂, the motion amplitude of target and the target imaging width at a frame can be obtained, with this amplitude and width for Target Motion Character parameter s during current location ₁and s ₂.Can the position of localizing objects by the saltus step of gray scale and the height etc. of direction of motion, speed and imaging is calculated with it.The gray average of half-circle area internal object and variance are another two characteristic parameters, obtain a principal eigenvector V used for successive image process as standard feature vector.

By the movable information of Small object, will provide motion state forecast model is

${\stackrel{\mathrm{\&Lambda;}}{x}}_{k+1|k}^{\left(i\right)}={\stackrel{\mathrm{\&Lambda;}}{x}}^{\left(i\right)}(k+1|k)=F_i\left(k\right){\stackrel{\mathrm{\&Lambda;}}{x}}^{\left(i\right)}\left(k\right|k)$

Here, F _ik () is the motion state transition matrix from previous frame to present frame, must select suitable metastasis model to it, 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 implementation method setting up the incidence matrix of two inter frame image moving small target is:

Set up m × n object matching matrix M.Here, m is the number of the moving small target of present frame, and n is the number of the moving small target of previous frame.The value of element M (i, j) is given by following formula.

$M(i,j)=\left\{\begin{array}{ccc}|r_i-r_j|,& \mathrm{if}& r_i+r_j>\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}\\ \&infin;,& & \mathrm{else}\end{array}\right.$

R in formula _iit is the radius of the i-th target; r _jfor the radius of jth target; (x _i, y _i) be the center-of-mass coordinate of the i-th target; (x _j, y _j) be the center-of-mass coordinate of jth target; ∞ represents a very large numerical value.

Coupling matrix is utilized to realize mating of moving small target in present image and moving small target in previous frame image.First, in matrix M, selective value is minimum and be not the element of ∞, and the row and column corresponding to this element is the numbering of current kinetic Small object and previous frame moving small target respectively, the moving small target that row is corresponding like this with arrange corresponding moving small target and match.Then all elements value of the row and row that complete coupling is become ∞.Continue to find minimum value in matrix M, complete the coupling of moving small target, until all values in matrix all becomes ∞.After search terminates, do not find the row representative of coupling target in present image, have the appearance of new moving small target, do not find row representative certain moving small target in present image of coupling target to disappear.

In embodiments of the present invention, in step S104, based on the information fusion of multiple image superposition, fuzzy pushdown automata chain bullet stack recursive operation is utilized to propose the track algorithm of large-scale image and video image motion Small object under network environment and recognition methods is:

Based on multiple features fusion, propose the joint probability based on the characteristic sum motion state of present frame and previous frame moving small target and coarseness data association algorithm, hypothesis testing method.Utilize joint probability and coarseness data association algorithm pre-estimation multimode small target tracking algorithm.Information updating hypothesis state after adopting multihypothesis test method to utilize in frame.While these algorithms of proposition and method, provide suitable " thresholding ", only calculate and be positioned at the observed reading of " thresholding " and the association probability of Small object and roughness, this greatly reduces calculated amount, solve Small object collision problem by the distance of observation position and predicted position.

Realize the moving small target recognition and tracking of image, a kind of suitable recognizer just must be proposed, the moving small target proper vector of the known class can thought to be identified and train is all fuzzy number vector, its proper vector is compared, defined feature subordinate function, degree of membership and matched rule, provide fuzzy object recognizer.Be embodied as:

For implementing to follow the tracks of to the track of moving small target in image or video, the present invention proposes the motion state blending algorithm of multiframe superposition: according to the motion state of moving small target in image or video, a kind of blending algorithm of detection and tracking system will be designed, to realize the tracking to moving small target.This algorithm is made up of three parts: (a) maps tracker and enter in the neural network with N number of feedback neural unit; B () allows N number of feedback neural unit be the state fusion completing tracker on multiframe merges in time; C () makes M non-feedback output neuron carry out spatially the state fusion of the tracker in i.e. single frames fusion again, the Space Time accumulation finally calculating tracker state is merged, and has carried out the fusion tracking algorithm of tracker to moving small target.

Designing tracker state fusion algorithm is: here, by the state estimation of the 1st frame to i-th Small object of the current kth frame of motion information prediction of kth frame, the state estimation of all Small object of current kth frame, it is the fuzzy membership of model.The key of following the tracks of is carved into by measuring the fuzzy membership that the k-1 moment predicts current k moment i-th tracker model from the outset ${\mathrm{\&mu;}}_{k|k-1}^{\left(i\right)}\stackrel{\mathrm{\&Delta;}}{=}\mathrm{Pr}\left\{m_k^{\left(i\right)}\right|z^{k-1}\}=\underset{j}{\mathrm{\&Sigma;}}{\mathrm{\&pi;}}_{\mathrm{ji}}{\mathrm{\&mu;}}_{k-1}^{\left(j\right)},$ Here ${\mathrm{\&mu;}}_{k-1}^{\left(j\right)}=\mathrm{Pr}\left\{m_{k-1}^{\left(j\right)}\right|z^{k-1}\}$ The fuzzy membership of k-1 moment tracker model, known, π _ji=Pr{m _k=m ⁽ⁱ⁾| m _k-1=m ^(j)from model m _k-1to model m _kstate transfer fuzzy membership.The tracker state fusion set up to the tracing process of moving small target as shown in Figure 3.

On last superposition frame, Small object point shows as the very strong track of a correlativity.But identify the track still non-easy thing of Small object point, mainly some discrete noise spots still likely exist.Also may occur in addition tracing point and noise spot interlaced.Therefore can only proceed from the situation as a whole to adjudicate according to overall relevancy under certain hypothesis.For this reason, propose based on fuzzy pushdown automata chain, utilize play stack with the method for popping with the persistence length of track for thresholding carries out track judgement.Tracking results as shown in Figure 4.As can be seen from Figure 4, illustrate that the fuzzy pushdown automata D-chain trace algorithm that the present invention proposes achieves larger success in pattern recognition system.

The fuzzy membership μ of each component of proper vector of moving small target to be identified or track is calculated according to blending algorithm _ij, the proper vector that can obtain Unknown Motion Small object or track is U _i=[μ _i1, μ _i2..., μ _ik] ^t.It with trained known i-th ₀the multi-Dimensional parameters proper vector U of classification _i0compare, and if only if time, adjudicate moving small target to be identified or track belongs to i-th ₀class.Even make then judge that moving small target to be identified or track belong to i-th ₀class.Here, δ is threshold value, and B is the index set of target or track class.

Concrete steps of the present invention are:

Step one, provides and extract order calibration method from the single-frame images of video sequence, weakens or eliminate the impact of background and noise;

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 target;

Step 4, based on the information fusion of multiple image superposition, utilizes fuzzy pushdown automata chain bullet stack recursive operation to propose track algorithm and the recognition methods of large-scale image and video image motion Small object under network environment.

Further, in step one, provide and extract order calibration method from the single-frame images of video sequence, when weakening or eliminate the affecting of background and noise, concrete methods of realizing is:

Because the complexity of scene, degree of stability affect the effect of target following.Such as, target image is by the impact etc. of the variation of object in uneven, the background of illumination.Effectively to split from image, extract target, a kind of algorithm weakening or eliminate background or noise effect must be proposed.The present invention, by based on the basis of mathematical morphology, proposes the image object detection algorithm of opening and closing conversion.The main thought of this algorithm utilizes mathematical morphology combinatorial operation to ask for local maximum and minimal value, alleviates the calculated amount of subsequent treatment, reduces false alarm rate as far as possible and count.Region growing is carried out to each Local modulus maxima, minimum point enforcement weakens or eliminates, possible target is selected.

Utilize this algorithm to make to reach and convert filtered image for each frame through opening and closing it shows as large-area zero background and comprises pinpoint target, and random spotted noise is in interior high fdrequency component.Then according to the correlativity of moving target between consecutive frame, carry out the superposition of difference multiframe, Small object is due to its motility, superposition frame shows as the very strong tracing point of correlativity, and high frequency spotted noise, then because its randomness is cancelled out each other a part, unmatched part shows as the very poor stochastic distribution noise spot of correlativity in geometric area.

Such as, use simple a kind of opening operation conversion g=f-f ο B, filtering is carried out to single-frame images, the place changed greatly in image can be detected, namely be equivalent to high fdrequency component, the place that in the image of energy elimination simultaneously, gray-value variation is comparatively mild, is equivalent to low frequency component.Utilize this conversion just can carry out filtering to a width single-frame images, filter low-frequency component, be equivalent to filter extended background, leave the HFS comprising Small object.In formula, f is gradation of image frame, and B is structure, and ο represents opening operation.Use opening and closing converts, to the wherein frame result of the video image of moving small target 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 less, and filtering low frequency background is more thorough, and the target size that can retain is less.

Further, in step 2, implement Weak target extraction of motion information and status predication modeling, concrete methods of realizing is:

Detection and tracking is carried out to target, first sets up a reference template as standard form.

Given known moving target, namely under certain video condition containing in a two field picture of this target, detect its movement velocity, current location, direction of motion o, height h, the gray average μ of imaging and variances sigma.Using the image of this frame as reference template.

Set up reference template as follows: with target location central point for the center of circle, to clap in front half-circle area that the product of frame time and the projection speed of target in imaging plane is radius, search for.With the change frequency of gray scale for thresholding, threshold value is set.Search the change frequency of gray scale along image width and image height direction respectively, if the change frequency of both direction is all less than or equal to 2, extends and clap frame time, until the grey scale change number of times at least one direction is more than 3 or 3 times.Then, be starting point with the center of circle, calculate the distance d between continuous three adjacent grey scale change ₁and d ₂, the motion amplitude of target and the target imaging width at a frame can be obtained, with this amplitude and width for Target Motion Character parameter s during current location ₁and s ₂.Can the position of localizing objects by the saltus step of gray scale and the height etc. of direction of motion, speed and imaging is calculated with it.The gray average of half-circle area internal object and variance are another two characteristic parameters, obtain a principal eigenvector V used for successive image process as standard feature vector.

By the movable information of Small object, will provide motion state forecast model is

${\stackrel{\mathrm{\&Lambda;}}{x}}_{k+1|k}^{\left(i\right)}={\stackrel{\mathrm{\&Lambda;}}{x}}^{\left(i\right)}(k+1|k)=F_i\left(k\right){\stackrel{\mathrm{\&Lambda;}}{x}}^{\left(i\right)}\left(k\right|k)$

Here, F _ik () is the motion state transition matrix from previous frame to present frame, must select suitable metastasis model to it, for the state estimation of previous frame, for the state estimation of present frame.

Further, in step 3, the implementation method setting up the incidence matrix of two inter frame image moving small target is:

Set up m × n object matching matrix M.Here, m is the number of the moving small target of present frame, and n is the number of the moving small target of previous frame.The value of element M (i, j) is given by following formula.

$M(i,j)=\left\{\begin{array}{ccc}|r_i-r_j|,& \mathrm{if}& r_i+r_j>\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}\\ \&infin;,& & \mathrm{else}\end{array}\right.$

R in formula _iit is the radius of the i-th target; r _jfor the radius of jth target; (x _i, y _i) be the center-of-mass coordinate of the i-th target; (x _j, y _j) be the center-of-mass coordinate of jth target; ∞ represents a very large numerical value.

Coupling matrix is utilized to realize mating of moving small target in present image and moving small target in previous frame image.First, in matrix M, selective value is minimum and be not the element of ∞, and the row and column corresponding to this element is the numbering of current kinetic Small object and previous frame moving small target respectively, the moving small target that row is corresponding like this with arrange corresponding moving small target and match.Then all elements value of the row and row that complete coupling is become ∞.Continue to find minimum value in matrix M, complete the coupling of moving small target, until all values in matrix all becomes ∞.After search terminates, do not find the row representative of coupling target in present image, have the appearance of new moving small target, do not find row representative certain moving small target in present image of coupling target to disappear.

Further, in step 4, based on the information fusion of multiple image superposition, fuzzy pushdown automata chain bullet stack recursive operation is utilized to propose the track algorithm of large-scale image and video image motion Small object under network environment and recognition methods is:

Based on multiple features fusion, the present invention proposes Describing Motion Small object and the data association algorithm following the tracks of its movement locus.Joint probability based on the characteristic sum motion state of present frame and previous frame moving small target and coarseness data association algorithm, hypothesis testing method are proposed simultaneously.Utilize joint probability and coarseness data association algorithm pre-estimation multimode small target tracking algorithm.Information updating hypothesis state after adopting multihypothesis test method to utilize in frame.While these algorithms of proposition and method, provide suitable " thresholding ", only calculate and be positioned at the observed reading of " thresholding " and the association probability of Small object and roughness, this greatly reduces calculated amount, solve Small object collision problem by the distance of observation position and predicted position.

Realize the moving small target recognition and tracking of image, a kind of suitable recognizer just must be proposed, the moving small target proper vector of the known class can thought to be identified and train is all fuzzy number vector, its proper vector is compared, defined feature subordinate function, degree of membership and matched rule, provide fuzzy object recognizer.Be embodied as:

For implementing to follow the tracks of to the track of moving small target in image or video, the present invention proposes the motion state blending algorithm of multiframe superposition: according to the motion state of moving small target in image or video, a kind of blending algorithm of detection and tracking system will be designed, to realize the tracking to moving small target.This algorithm is made up of three parts: (a) maps tracker and enter in the neural network with N number of feedback neural unit; B () allows N number of feedback neural unit be the state fusion completing tracker on multiframe merges in time; C () makes M non-feedback output neuron carry out spatially the state fusion of the tracker in i.e. single frames fusion again, the Space Time accumulation finally calculating tracker state is merged, and has carried out the fusion tracking algorithm of tracker to moving small target.

Designing tracker state fusion algorithm is: here, by the state estimation of the 1st frame to i-th Small object of the current kth frame of motion information prediction of kth frame, the state estimation of all Small object of current kth frame, it is the fuzzy membership of model.The key of following the tracks of is carved into by measuring the fuzzy membership that the k-1 moment predicts current k moment i-th tracker model from the outset ${\mathrm{\&mu;}}_{k|k-1}^{\left(i\right)}\stackrel{\mathrm{\&Delta;}}{=}\mathrm{Pr}\left\{m_k^{\left(i\right)}\right|z^{k-1}\}=\underset{j}{\mathrm{\&Sigma;}}{\mathrm{\&pi;}}_{\mathrm{ji}}{\mathrm{\&mu;}}_{k-1}^{\left(j\right)},$ Here ${\mathrm{\&mu;}}_{k-1}^{\left(j\right)}=\mathrm{Pr}\left\{m_{k-1}^{\left(j\right)}\right|z^{k-1}\}$ The fuzzy membership of k-1 moment tracker model, known, π _ji=Pr{m _k=m ⁽ⁱ⁾| m _k-1=m ^(j)from model m _k-1to model m _kstate transfer fuzzy membership.The tracker state fusion set up to the tracing process of moving small target as shown in Figure 3.

On last superposition frame, Small object point shows as the very strong track of a correlativity.But identify the track still non-easy thing of Small object point, mainly some discrete noise spots still likely exist.Also may occur in addition tracing point and noise spot interlaced.Therefore can only proceed from the situation as a whole to adjudicate according to overall relevancy under certain hypothesis.For this reason, propose based on fuzzy pushdown automata chain, utilize play stack with the method for popping with the persistence length of track for thresholding carries out track judgement.Tracking results as shown in Figure 4.As can be seen from Figure 4, illustrate that the fuzzy pushdown automata D-chain trace algorithm that the present invention proposes achieves larger success in pattern recognition system.

The fuzzy membership μ of each component of proper vector of moving small target to be identified or track is calculated according to blending algorithm _ij, the proper vector that can obtain Unknown Motion Small object or track is U _i=[μ _i1, μ _i2..., μ _ik] ^t.It with trained known i-th ₀the multi-Dimensional parameters proper vector U of classification _i0compare, and if only if time, adjudicate moving small target to be identified or track belongs to i-th ₀class.Even make then judge that moving small target to be identified or track belong to i-th ₀class.Here, δ is threshold value, and B is the index set of target or track class.

The recognition and tracking problem of the moving small target under low signal-to-noise ratio provided by the invention, existing document also has similar research, but all exist in these literature methods follow the tracks of real-time speed comparatively slow, follow the tracks of or the weakness such as recognition effect is poor.But the present invention provides different implementation methods stage by stage, according to target respectively with a kind of, propose a kind of small target tracking algorithm and recognition methods of the time-space domain fused filtering based on single frames and multiframe.To moving small target in video image easily by the situation that other objects in complex background or noise block or flood, propose the elimination of opening and closing conversion or weaken the algorithm of background and noise; To the small and weak characteristic of Small object, propose the adaptive neural network competitive model of on-line study, the active unit utilizing it to compete extracts the multidimensional characteristic parameter of Weak target; For the kinetic characteristic of video Small Target, utilize the sudden change of gray scale, give Small object motion state model and forecast model; To the real-time detection and tracking of moving small target, have employed fuzzy pushdown automata chain and carry out track identification and tracking, with the fuzzy pushdown automata chain degree of depth for threshold value carries out track judgement, thus propose a kind of based on the moving small target track algorithm under complex environment and recognition methods.These researchs of the present invention contribute to target identification and image procossing personnel understand detect target the characteristics of motion, active degree and the impact on other targets thereof, thus provide corresponding decision-making, seek to suppress or eliminate the impact of undesirable element on itself or other important goal and be all very important.This plays important reference and reference role by the development of all target recognition and trackings based on video system such as military, civil, public security system, road traffic.

This moving small target recognition and tracking method that the present invention proposes is compared with existing recognition methods, and simulation result is as shown in Fig. 5 and table 1.

To width moving small target image every in video, selecting to dimensioning in emulation is 2 ⁸wavelet basis function carry out repeating for 10 times experiment according to the method that the present invention proposes, the sample of different number is got in each experiment respectively.The more two kinds of moving small target of the utilization that the inventive method and current document provide are followed the tracks of and method of identification compares, and the correct average recognition rate when emulation 500 times is respectively 95.14%, 92.45%, 88.17%.And also comparatively the recognition methods of order first two is very fast for the recognition speed of the inventive method, and result is as shown in Fig. 5 and table 1.

Know from Fig. 5, the average correct recognition rata based on feature extraction method of identification of the present invention is the highest.In experiment, along with increasing of sample number, average correct recognition rata constantly increases, and increases sample curve again and tend to be steady gradually when reaching certain sample size.

In order to evaluate the combination property of each algorithm, we are according to several aspects such as computing velocity, memory space, the traffic, correct recognition ratas, adopt the method for fixed guantity combining with fixed quality, Integrated comparative.Have rated the quality of different personal recognition method.Table 1 gives comprehensive comparison.

The Integrated comparative of table 1 different motion Small object method of identification

Algorithm	Correct recognition rata average	Computing velocity	Memory space	The traffic
					Existing method 1	0.8817	0.82s	In	In
Existing method 2	0.9245	0.49s	Lower	Lower
					Put forward the methods of the present invention	0.9514	0.426s	Low	Low

Computing velocity in table 1 is that algorithm often walks all to calculate and repeats for 10 times to test average calculation times used under simulated environment, the just computing time of algorithm itself.The computing machine that emulation uses is the internal memory to 4,2G, and programming language used is MATLAB.Storage in table 1 and traffic demands are just according to the computation process of various algorithm, and complexity is roughly estimated.As can be seen from the result of table, memory space and the traffic are closely linked to be.In table correct average recognition rate be each algorithm under given emulation experiment ambient conditions, after 500 emulation experiments are averaged, then get the mean value of 10 time steps.In fact they are average on room and time of correct recognition rata, because of but the population mean of correct recognition rata.

Know from simulation result, the method for identification based on feature extraction of the present invention not only has processing speed, lower memory space and the traffic faster, and also has good recognition effect.

Implementation of the present invention is as follows:

(1) provide extract order calibration method from the single-frame images of video sequence, weaken or eliminate background and noise impact based on the basis of mathematical morphology, propose the image object detection algorithm of opening and closing conversion.The main thought of this algorithm utilizes mathematical morphology combinatorial operation to ask for local maximum and minimal value, alleviates the calculated amount of subsequent treatment, reduces false alarm rate as far as possible and count.Region growing is carried out to each Local modulus maxima, minimum point enforcement weakens or eliminates, possible target is selected.

Utilize this algorithm to make to reach and convert filtered image for each frame through opening and closing it shows as large-area zero background and comprises pinpoint target, and random spotted noise is in interior high fdrequency component.Then according to the correlativity of moving target between consecutive frame, carry out the superposition of difference multiframe, Small object is due to its motility, superposition frame shows as the very strong tracing point of correlativity, and high frequency spotted noise, then because its randomness is cancelled out each other a part, unmatched part shows as the very poor stochastic distribution noise spot of correlativity in geometric area.

Such as, use simple a kind of opening operation conversion g=f-f ο B, filtering is carried out to single-frame images, the place changed greatly in image can be detected, namely be equivalent to high fdrequency component, the place that in the image of energy elimination simultaneously, gray-value variation is comparatively mild, is equivalent to low frequency component.Utilize this conversion just can carry out filtering to a width single-frame images, filter low-frequency component, be equivalent to filter extended background, leave the HFS comprising Small object.In formula, f is gradation of image frame, and B is structure, and ο represents opening operation.Use opening and closing converts, to the wherein frame result of the video image of moving small target as shown in Figure 2.

(2) Weak target extraction of motion information and status predication modeling is provided

Detection and tracking is carried out to target, first sets up a reference template as standard form.

Given known moving target, namely under certain video condition containing in a two field picture of this target, detect its movement velocity, current location, direction of motion o, height h, the gray average μ of imaging and variances sigma.Using the image of this frame as reference template.

Set up reference template as follows: with target location central point for the center of circle, to clap in front half-circle area that the product of frame time and the projection speed of target in imaging plane is radius, search for.With the change frequency of gray scale for thresholding, threshold value is set.Search the change frequency of gray scale along image width and image height direction respectively, if the change frequency of both direction is all less than or equal to 2, extends and clap frame time, until the grey scale change number of times at least one direction is more than 3 or 3 times.Then, be starting point with the center of circle, calculate the distance d between continuous three adjacent grey scale change ₁and d ₂, the motion amplitude of target and the target imaging width at a frame can be obtained, with this amplitude and width for Target Motion Character parameter s during current location ₁and s ₂.Can the position of localizing objects by the saltus step of gray scale and the height etc. of direction of motion, speed and imaging is calculated with it.The gray average of half-circle area internal object and variance are another two characteristic parameters, obtain a principal eigenvector V used for successive image process as standard feature vector.

By the movable information of Small object, will provide motion state forecast model is

${\stackrel{\mathrm{\&Lambda;}}{x}}_{k+1|k}^{\left(i\right)}={\stackrel{\mathrm{\&Lambda;}}{x}}^{\left(i\right)}(k+1|k)=F_i\left(k\right){\stackrel{\mathrm{\&Lambda;}}{x}}^{\left(i\right)}\left(k\right|k)$

Here, F _ik () is the motion state transition matrix from previous frame to present frame, must select suitable metastasis model to it, for the state estimation of previous frame, for the state estimation of present frame.

(3) incidence matrix of two inter frame image moving small target is set up

Set up m × n object matching matrix M.Here, m is the number of the moving small target of present frame, and n is the number of the moving small target of previous frame.The value of element M (i, j) is given by following formula.

$M(i,j)=\left\{\begin{array}{ccc}|r_i-r_j|,& \mathrm{if}& r_i+r_j>\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}\\ \&infin;,& & \mathrm{else}\end{array}\right.$

R in formula _iit is the radius of the i-th target; r _jfor the radius of jth target; (x _i, y _i) be the center-of-mass coordinate of the i-th target; (x _j, y _j) be the center-of-mass coordinate of jth target; ∞ represents a very large numerical value.

Coupling matrix is utilized to realize mating of moving small target in present image and moving small target in previous frame image.First, in matrix M, selective value is minimum and be not the element of ∞, and the row and column corresponding to this element is the numbering of current kinetic Small object and previous frame moving small target respectively, the moving small target that row is corresponding like this with arrange corresponding moving small target and match.Then all elements value of the row and row that complete coupling is become ∞.Continue to find minimum value in matrix M, complete the coupling of moving small target, until all values in matrix all becomes ∞.After search terminates, do not find the row representative of coupling target in present image, have the appearance of new moving small target, do not find row representative certain moving small target in present image of coupling target to disappear.

(4) based on the information fusion of multiple image superposition, fuzzy pushdown automata chain bullet stack recursive operation is utilized to propose track algorithm and the recognition methods of large-scale image and video image motion Small object under network environment

Based on multiple features fusion, the present invention proposes Describing Motion Small object and the data association algorithm following the tracks of its movement locus.Propose the joint probability based on the characteristic sum motion state of present frame and previous frame moving small target and coarseness data association algorithm, hypothesis testing method simultaneously.Utilize joint probability and coarseness data association algorithm pre-estimation multimode small target tracking algorithm.Information updating hypothesis state after adopting multihypothesis test method to utilize in frame.While these algorithms of proposition and method, provide suitable " thresholding ", only calculate and be positioned at the observed reading of " thresholding " and the association probability of Small object and roughness, this greatly reduces calculated amount, solve Small object collision problem by the distance of observation position and predicted position.

Realize the moving small target recognition and tracking of image, a kind of suitable recognizer just must be proposed, the moving small target proper vector of the known class can thought to be identified and train is all fuzzy number vector, its proper vector is compared, defined feature subordinate function, degree of membership and matched rule, provide fuzzy object recognizer.Be embodied as:

For implementing to follow the tracks of to the track of moving small target in image or video, the present invention proposes the motion state blending algorithm of multiframe superposition: according to the motion state of moving small target in image or video, a kind of blending algorithm of detection and tracking system will be designed, to realize the tracking to moving small target.This algorithm is made up of three parts: (a) maps tracker and enter in the neural network with N number of feedback neural unit; B () allows N number of feedback neural unit be the state fusion completing tracker on multiframe merges in time; C () makes M non-feedback output neuron carry out spatially the state fusion of the tracker in i.e. single frames fusion again, the Space Time accumulation finally calculating tracker state is merged, and has carried out the fusion tracking algorithm of tracker to moving small target.

Designing tracker state fusion algorithm is: here, by the state estimation of the 1st frame to i-th Small object of the current kth frame of motion information prediction of kth frame, the state estimation of all Small object of current kth frame, it is the fuzzy membership of model.The key of following the tracks of is carved into by measuring the fuzzy membership that the k-1 moment predicts current k moment i-th tracker model from the outset ${\mathrm{\&mu;}}_{k|k-1}^{\left(i\right)}\stackrel{\mathrm{\&Delta;}}{=}\mathrm{Pr}\left\{m_k^{\left(i\right)}\right|z^{k-1}\}=\underset{j}{\mathrm{\&Sigma;}}{\mathrm{\&pi;}}_{\mathrm{ji}}{\mathrm{\&mu;}}_{k-1}^{\left(j\right)},$ Here ${\mathrm{\&mu;}}_{k-1}^{\left(j\right)}=\mathrm{Pr}\left\{m_{k-1}^{\left(j\right)}\right|z^{k-1}\}$ The fuzzy membership of k-1 moment tracker model, known, π _ji=Pr{m _k=m ⁽ⁱ⁾| m _k-1=m ^(j)from model m _k-1to model m _kstate transfer fuzzy membership.The tracker state fusion set up to the tracing process of moving small target as shown in Figure 3.

On last superposition frame, Small object point shows as the very strong track of a correlativity.But identify the track still non-easy thing of Small object point, mainly some discrete noise spots still likely exist.Also may occur in addition tracing point and noise spot interlaced.Therefore can only proceed from the situation as a whole to adjudicate according to overall relevancy under certain hypothesis.For this reason, propose based on fuzzy pushdown automata chain, utilize play stack with the method for popping with the persistence length of track for thresholding carries out track judgement.Tracking results as shown in Figure 4.As can be seen from Figure 4, illustrate that the fuzzy pushdown automata D-chain trace algorithm that the present invention proposes achieves larger success in pattern recognition system.

The fuzzy membership μ of each component of proper vector of moving small target to be identified or track is calculated according to blending algorithm _ij, the proper vector that can obtain Unknown Motion Small object or track is U _i=[μ _i1, μ _i2..., μ _ik] ^t.It with trained known i-th ₀the multi-Dimensional parameters proper vector U of classification _i0compare, and if only if time, adjudicate moving small target to be identified or track belongs to i-th ₀class.Even make then judge that moving small target to be identified or track belong to i-th ₀class.Here, δ is threshold value, and B is the index set of target or track class.

Innovative point of the present invention is as follows:

(1) how the Major Difficulties that data correlation is moving small target track algorithm is set up to two inter frame image Small object in video sequence.The complicacy of background under network environment, the ratio that Small object occupies in the picture is little, color of object and the similarity degree of background color, the degree of stability of background, and multiobject generation that is mutual and various special circumstances all can be followed the tracks of to moving small target and bring difficulty.Small object external appearance characteristic as the information such as target shape and texture, because the generation of the process of blocking almost is submerged in the picture, and Small object motion uncertainty, cause the loss of Small object information, be easy to occur follow the tracks of unsuccessfully.How effectively process is blocked, and particularly seriously blocks, and is a difficult point during moving small target is followed the tracks of always.In monitor video, the Small object outward appearance in each frame is often closely similar, and how choosing suitable feature and realize data correlation accurately to distinguish different Small object motion state preferably, is the key issue that this project will be studied.And the domestic and international research to this problem is at present almost a blank.

To the discussion of this problem, the present invention analyzes Small object imaging characteristics in the picture and movable information, first to the small and weak of target and kinetic characteristic, adopts opening and closing conversion to remove background, elimination or the process of noise decrease scheduling algorithm.Then according to the motion relevance of moving small target between consecutive frame, the superposition of difference multiframe is carried out.

(2) utilizing the characteristic of moving small target itself, set up movable information vector, extracting that the multi-Dimensional parameters of Small object multi-characteristic points moving small target extracts is present many departments difficult problem thorny and urgently to be resolved hurrily.Present invention utilizes the characteristic of moving small target itself, excavate the unique point of each characteristic parameter of moving small target and characteristic parameter, set up kinematics character vector, the multi-Dimensional parameters giving moving small target extracts.

(3) adopt the information fusion superposed based on multiple image and fuzzy pushdown automata chain recursive operation to carry out track judgement, Modling model algorithm and feature based are chosen parallel mechanism and are carried out image recognition.

Along with the lifting that the development of science and technology and the security protection of the mankind are realized, under network environment, video monitoring system obtains applying more and more widely in every field.And be a very useful job based on the identification of the moving target of video monitoring, space flight can be applied in, military affairs, guided missile track identification follow the tracks of, the various fields such as break in traffic rules and regulations detection.But in some occasion as in infrared guidance, needs can be intercepted and captured and locking tracking target as soon as possible.So to the accurate detection and tracking of moving small target, the application in the every field such as military, civil is seemed more and more important, also more and more urgent.Under the strong clutter background condition of network environment, the test problems of low signal-to-noise ratio moving small target directly determines operating distance and the detection perform of detection system, and it solves has very important practical significance for raising detection system performance.

For the research of these problems, the present invention analyzes the various features of polymorphic target, utilize tracker state fusion, fuzzy pushdown automata chain bullet stack recursive operation carries out track following to the moving small target of video image, provide Algorithm of Moving Point Targets Detection.

Specific embodiments of the invention:

The recognition and tracking problem of the moving small target under low signal-to-noise ratio provided by the invention, existing document also has similar research, but all exist in these methods follow the tracks of real-time speed comparatively slow, follow the tracks of or the weakness such as recognition effect is poor.But the present invention provides different implementation methods stage by stage, according to target respectively with a kind of, propose a kind of small target tracking algorithm of the time-space domain fused filtering based on single frames and multiframe and recognition methods (based on such as Fig. 3).1. to moving small target in video image easily by the situation that other objects in complex background or noise block or flood, propose the elimination of opening and closing conversion or weaken the algorithm of background and noise; 2. to the small and weak characteristic of Small object, propose the adaptive neural network competitive model of on-line study, the active unit utilizing it to compete extracts the multidimensional characteristic parameter of Weak target; 3. for the kinetic characteristic of video Small Target, utilize the sudden change of gray scale, give Small object motion state model and forecast model; 4. to the real-time detection and tracking of moving small target, have employed fuzzy pushdown automata chain and carry out track identification and tracking, with the fuzzy pushdown automata chain degree of depth for threshold value carries out track judgement, thus propose a kind of based on the moving small target track algorithm under complex environment and recognition methods.

To technical scheme enforcement be 1.:

Use opening operation conversion g=f-f ο B or closed operation conversion g=f B-f, filtering is carried out to single-frame images, the place changed greatly in image detected, namely be equivalent to high fdrequency component, the place that in the image of energy elimination simultaneously, gray-value variation is comparatively mild, is equivalent to low frequency component, this conversion is utilized just to carry out filtering to a width single-frame images, filter low-frequency component, be equivalent to filter extended background, leave the HFS comprising Small object; In formula, f is gradation of image frame, and B is structure, and ο represents opening operation, and represents closed operation.

Claims (6)

1. the method for the recognition and tracking of a low signal-to-noise ratio moving small target, it is characterized in that, the method of the recognition and tracking of this low signal-to-noise ratio moving small target provides different implementation methods stage by stage, according to target respectively with a kind of, to moving small target in video image easily by the situation that other objects in complex background or noise block or flood, propose the elimination of opening and closing conversion or weaken the algorithm of background and noise; To the small and weak characteristic of Small object, propose the adaptive neural network competitive model of on-line study, utilize the active unit of competition to extract the multidimensional characteristic parameter of Weak target; For the kinetic characteristic of video Small Target, utilize the sudden change of gray scale, give Small object motion state model and forecast model; To the real-time detection and tracking of moving small target, have employed fuzzy pushdown automata chain and carry out track identification and tracking, with the fuzzy pushdown automata chain degree of depth for threshold value carries out track judgement;

Specifically comprise the following steps:

Step one, provides and extract order calibration method from the single-frame images of video sequence, weakens or eliminate the impact of background and noise;

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 target;

Step 4, based on the information fusion of multiple image superposition, utilizes fuzzy pushdown automata chain bullet stack recursive operation to propose track algorithm and the recognition methods of large-scale image and video image motion Small object under network environment.

2. the method for the recognition and tracking of low signal-to-noise ratio moving small target as claimed in claim 1, it is characterized in that, in step one, provide and extract order calibration method from the single-frame images of video sequence, when weakening or eliminate the affecting of background and noise, concrete methods of realizing is:

Mathematical morphology combinatorial operation is utilized to ask for local maximum and minimal value, alleviate the calculated amount of subsequent treatment, reduce false alarm rate to count as far as possible, region growing is carried out to each Local modulus maxima, minimum point enforcement weakens or eliminates, possible target is selected;

Use opening operation conversion g=f-f o B or closed operation conversion g=f B-f, filtering is carried out to single-frame images, the place changed greatly in image detected, namely be equivalent to high fdrequency component, the place that in the image of energy elimination simultaneously, gray-value variation is comparatively mild, is equivalent to low frequency component, this conversion is utilized just to carry out filtering to a width single-frame images, filter low-frequency component, be equivalent to filter extended background, leave the HFS comprising Small object; In formula, f is gradation of image frame, and B is structure, and o represents opening operation, and represents closed operation;

For accurately identifying target or track, carry out the suppression of Small object enhancing and interference, because Small object point moves between each frame, multiframe superposition can be carried out to video image, on the last frame of superposition, Small object point shows as the very strong tracing point of correlativity, but noise still likely floods Small object track, propose multi-frame difference superposition algorithm, choose the image sequence comprising moving small target point, the superposition value of odd-numbered frame and even frame each n frame difference, that is:

In formula, f _ifor the i-th frame in image sequence, f _zfor last superposition frame;

Add frame through the stack and take thresholding process, method is as follows:

δ is threshold value, gets m, N are the size of superposition two field picture.

3. the method for the recognition and tracking of low signal-to-noise ratio moving small target as claimed in claim 1, it is characterized in that, in step 2, when providing Weak target extraction of motion information and status predication modeling, concrete methods of realizing is:

First construct the adaptive neural network competitive model of on-line study, the active unit utilizing it to compete extracts the multidimensional characteristic parameter of Weak target:

The first step, initialization network: the dimension of fixing output nerve network grid is N × M, and input layer is quadravalence network, and the weight that random initializtion input neuron is connected with output neuron, make t represent algorithm iteration number of times, put t=0;

Second step, selects victor: gray scale, colourity, the motion parameter X={x of each frame Small object image ₁, x ₂..., x _dbe input to input neuron in network, to each input neuron value x _j, the output of the node i in competition layer for g is an activation function, as got α > 0 is constant, the slope of controlling curve; μ _it () is that p ties up input vector x _jwith p dimensional weight vector ω _jieuclidean distance between (t) || x _j-ω _ji(t) || with, namely ω _jit () is in t from input layer node j to the connection weight of competition layer node i vector, j ∈ J, J={1 ..., d}, i ∈ I, I={1 ..., N _iit is certain regional area of competition layer;

Select the output neuron i won ^*, in competition layer, corresponding minimum node will be won, if namely the node of so winning in competition layer is i ^*, then with i ^*the weight of association and and i ^*the weight of the some association that point is contiguous all can be adjusted;

3rd step, upgrades weight: N (i ^*) be triumph output neuron i ^*neighbour, specifically specified by the distance between output neuron, to each output neuron i ∈ { N (i ^*), i ^*, adjust renewal according to the following formula:

η (t)=η determines in advance; This rule only upgrades the neighbour of triumph output neuron;

4th step, standardized weight: to standardization after renewal weight, thus be consistent with input measurement standard;

5th step, continues circulation: repeat the first step to the 4th step, the number of times of iteration is set to t=t+1, shuts down criterion until meet, and shutting down criterion is || x _j-ω _ji(t) || < ε, get ε=0.5, or till having exceeded maximum cycle index.

4. the method for the recognition and tracking of low signal-to-noise ratio moving small target as claimed in claim 1, it is characterized in that, in step 3, the implementation method setting up the incidence matrix of two inter frame image moving small target is:

Set up m × n object matching matrix M, here, m is the number of the moving small target of present frame, and n is the number of the moving small target of previous frame, and the value of element M (i, j) is given by following formula:

R in formula _iit is the radius of the i-th target; r _jfor the radius of jth target; (x _i, y _i) be the center-of-mass coordinate of the i-th target; (x _j, y _j) be the center-of-mass coordinate of jth target; ∞ represents a very large numerical value;

First, in matrix M, selective value is minimum and be not the element of ∞, row and column corresponding to this element is the numbering of current kinetic Small object and previous frame moving small target respectively, the moving small target that the moving small target of row correspondence is corresponding with row matches, and then all elements value of the row and row that complete coupling is become ∞; Continue to find minimum value in matrix M, complete the coupling of moving small target, until all values in matrix all becomes ∞; After search terminates, do not find the row representative of coupling target in present image, have the appearance of new moving small target, do not find row representative certain moving small target in present image of coupling target to disappear.

5. the method for the recognition and tracking of low signal-to-noise ratio moving small target as claimed in claim 1, it is characterized in that, in step 4, based on the information fusion of multiple image superposition, fuzzy pushdown automata chain bullet stack recursive operation is utilized to propose the track algorithm of large-scale image and video image motion Small object under network environment and recognition methods is:

The first step, each fuzzy pushdown automata is identified in temporal fusion to moving small target:

F _ip(t) and Ξ _iwhat t () represented that t records by fuzzy pushdown automata i respectively is identified fuzzy membership and the Fuzzy Distribution that moving small target belongs to p class, represent to the l moment, accumulate that fusion obtains by i-th fuzzy pushdown automata be identified the fuzzy membership that target belongs to p class, represent the Fuzzy Distribution being identified target obtained by i-th fuzzy pushdown automata accumulation fusion to the l moment, here, l=1,2 ..., t, namely

With

O _p(p ∈ U) is moving small target, the measurement Fuzzy Distribution that Fuzzy Distribution and t are merged in the accumulation in t-1 moment is merged, and obtains the target identification accumulation of i-th fuzzy pushdown automata to t and merges Fuzzy Distribution for:

Wherein, s ₂fuzzy Integration Function, usual S ₂take off formula:

Now, with Fuzzy Distribution the motion state of corresponding Small object is the state estimation of the Small object of current t:

F _it (), for being carved into the motion state transition matrix of current time from upper a period of time, selects metastasis model, for the state estimation in a upper moment, for the state estimation of current time;

Second step, fuzzy pushdown automata is to the Space integration of Small object identification:

Obtaining the accumulation Fuzzy Distribution of t each fuzzy pushdown automata target identification afterwards, i=1 here ..., N, utilizes Fuzzy Integration Function to merge this N number of Fuzzy Distribution, merges Fuzzy Distribution to be till just obtaining t to the Space Time accumulation of target identification:

Utilize Fuzzy Integration Function theoretical, can obtain

S _nalso Fuzzy Integration Function is represented; If p=1 ..., M;

Now, with Fuzzy Distribution Ξ ^tthe motion state of corresponding Small object is be the state estimation of all Small object of current kth frame, motion state fusion results is: by the state estimation of the 1st frame to i-th Small object of the current kth frame of motion information prediction of kth frame, the state estimation of all Small object of current kth frame, it is the fuzzy membership of model; The key of following the tracks of is carved into by measuring the fuzzy membership that the k-1 moment predicts current k moment i-th tracker model from the outset here the fuzzy membership of k-1 moment tracker model, known, π _ji=Pr{m _k=m ⁽ⁱ⁾| m _k-1=m ^(j)from model m _k-1to model m _kstate transfer fuzzy membership.

6. the method for the recognition and tracking of low signal-to-noise ratio moving small target as claimed in claim 5, it is characterized in that, after multiframe merges, Small object point is further enhanced, major part noise spot is cut, and filtering random noise disturbance is adjudicated by fuzzy pushdown automata chain length, in order to reduce fuzzy pushdown automata chain length as far as possible while reservation Small object point, take thresholding process to fusion frame, method is as follows:

Threshold value T is the length of fuzzy pushdown automata chain, but this chain must ensure containing Small object point, f _afor total fusion frame, on last fusion frame, Small object point shows as the very strong track of a correlativity;

The each component f of proper vector of moving small target to be identified or track is calculated according to blending algorithm _afuzzy membership μ _ij, the proper vector namely obtaining Unknown Motion Small object or track is U _i=[μ _i1, μ _i2..., μ _ik] ^t; It with trained known i-th ₀the multi-Dimensional parameters proper vector of classification compare, and if only if time, adjudicate moving small target to be identified or track belongs to i-th ₀class; Even make then judge that moving small target to be identified or track belong to i-th ₀class; Here, δ is threshold value, and B is the index set of target or track class.