CN103903242B - Adaptive targets compressed sensing fusion tracking method based on video sensor network - Google Patents

Abstract

The present invention relates to a kind of adaptive targets compressed sensing fusion tracking method based on video sensor network, the present invention builds the video sensor network test system of the video node including some different visual angles；At the image sparse degree of initial time, the measurement initial value calculating each video node respectively, its initial pictures sequence is compressed perception CS sampling according to video node each in some video node；Using the expectation of variance of image as threshold value, by the measured value of t two field picture, iteration obtains the measured value of each video node t+1 two field picture and its t frame image sequence is carried out CS sampling.Measured value is regulated by image sparse degree, thus obtain optional sampling amount and with higher compression ratio, image can be compressed, reduce and calculate and the data volume of communication, multi-view image in conjunction with tracking creditability and testing result merges, more preferable tracking accuracy, and longer Web vector graphic time can be obtained.

Description

Adaptive targets compressed sensing fusion tracking method based on video sensor network

Technical field

The present invention relates to a kind of targeted compression perception fusion tracking method, particularly relate to a kind of based on video sensing The adaptive targets compressed sensing fusion tracking method of device network.

Background technology

Owing to Computer Vision and transmitted data amount are big, limited by sensor node resource and energy consumption, In order to extend network life, it is necessary to be compressed view data reducing amount of calculation and the traffic.

Compressive sensing theory is pointed out: as long as signal is compressible or is sparse at certain transform domain, then Just with conversion base incoherent observing matrix, conversion gained height dimensional signal can be projected to a low-dimensional with one Spatially, then by solve an optimization problem just can from the projection that these are a small amount of with high probability reconstruct Go out original signal, may certify that such projection contains the enough information of reconstruction signal.The required measured value of decoding Number much smaller than the sample number under traditional theory.

As shown in equation (1), < < n(m is for measuring square with m for compressed sensing CS (Compressive Sensing) Battle array dimension and, n is signal dimension) calculation matrix signal is measured.

y=φx=φΨα=Θα (1)

Wherein α is the rarefaction representation of signal, and Ψ is rarefaction representation matrix,For calculation matrix.

If x signal has openness at certain domain of variation, as shown in equation (2), and proper build up φ.

α=Ψ^Tx (2)

So can pass through equation (3) come this signal of perfect recovery:

$\underset{\mathrm{\&alpha;}}{\min }{\left|\right|\mathrm{\&alpha;}\left|\right|}_{l_0}s.\mathrm{ty}=\mathrm{\&Phi;\&Psi;\&alpha;}---\left(3\right)$

In recent years, much study and compressed sensing algorithm is incorporated into video sensor network, but traditional pressure Contracting perception algorithm uses fixing measured value M to sample image sequence, when image sparse degree changes greatly Time, the sampled data output of minimum can not be obtained.For the Target Tracking System of single-view, it is easily subject to hide , there is the problems such as tracking accuracy difference in gear, therefore, a kind of solution to the problems described above of invention is the most necessary 's.

Summary of the invention

It is an object of the invention to provide a kind of adaptive targets compressed sensing based on video sensor network to melt Close tracking, solve when image sparse degree changes greatly, the sampled data output of optimum can not be obtained, And for the Target Tracking System of single-view, the problem of tracking accuracy difference.

In order to solve the problems referred to above, the present invention relates to a kind of adaptive targets based on video sensor network Compressed sensing fusion tracking method, comprises the following steps:

S1: build the video sensor network test system of the video node including some different visual angles, by mixing Gauss model GMM builds t frame background, the t frame video sequence that test obtains^x _tIt it is t frame foreground picture Picture and t frame background sum, wherein t=0,1,2 ...；

S2: according to video node each in some video node at the image sparse degree K of initial time₀, calculate respectively The measurement initial value M of each video node₀Its initial pictures sequence is compressed perception CS sampling, obtains every Initial pictures y after the sampling of individual video node_oAfter carry out S4, wherein M₀≈K₀log(n/K₀), n is signal dimension； S3: use the expectation E (σ of the variance of image_t) as threshold value, by the measured value M of t two field picture_t, iteration Obtain the measured value M of each video node t+1 two field picture_t+1, It is target image x_d Variance；

WhenThen M_t+1=β₀M_t；

WhenThen M_t+1=β₁M_t；

Wherein,For the target image recovered, x_dFor target image,Represent and use norm 2 to recover Error amount；β₀≤1,β₁≥1,β₀,β₁Respectively reducing and amplification factor, α is regulation parameter；

And by the measured value M of t+1 two field picture_t+1T+1 two field picture is compressed perception CS sampling, obtains every T two field picture y after the sampling of individual video node_tAfter carry out S4;

S4: by t two field picture y after sampling_tRecover to obtain t frame video sequence and recovery obtains t frame target figure Picture.

It is preferred that further comprising the steps of in S2 and S3:

S51: after each video node is recovered to obtain t frame target image, uses Unscented kalman filtering UKF to carry out The detection of target following information;

S52: the target following information many video node detected, is sent to Centroid and carries out data fusion, the T frame ground level coordinate [u_t,v_t] and t frame image plane coordinate [x_t,y_tMapping relations between] are:

$\mathrm{\&lambda;}\left[\begin{array}{c}{u}_t\\ v_t\\ 1\end{array}\right]=H_j\left[\begin{array}{c}{x}_t\\ y_t\\ 1\end{array}\right],$ Wherein H_jMapping matrix for jth pixel: $H_j=\left(\begin{array}{ccc}{H}_{11}& H_{12}& H_{13}\\ H_{21}& H_{22}& H_{23}\\ H_{31}& H_{32}& H_{33}\end{array}\right),$ Draw each Ground level coordinate [the u of pixel_t,v_t]: $u_t=\frac{H_{11}{x}_t+H_{12}{y}_t+H_{13}}{H_{31}{x}_t+H_{32}{y}_t+H_{33}},v_t=\frac{H_{21}{x}_t+H_{22}{y}_t+H_{23}}{H_{31}{x}_t+H_{32}{y}_t+H_{33}},$ λ is yardstick Parameter；

S53: for each video node, by G (u, v)=∑_jw_jG_j(u, v) andTarget after being merged The weighted value of each pixel, wherein G in tracking information_j(u, v) is the target location of jth pixel, and (u v) is G The position of target, η after fusion_jFor jth pixel confidence level, w_jFor based on confidence level η_jFusion after The weighted value of jth pixel.

It is preferred that sampling rate m/n is at least 25%, wherein m is calculation matrix dimension, and n is signal dimension Degree.

It is preferred that described image sparse degree is: K_t=(λ₀logS_t+λ₁)S_t, wherein (λ₀,λ₁)∈R²,λ₀,λ₁For Function model parameter, R is Spatial Dimension size, S_tThe pixel number of target.

It is preferred that S4 also includes pass throughT two field picture y after sampling_tRecover to obtain the T frame target imageWhereinFor t frame background after sampling, Δ represents the recovery process of CS.

It is preferred that described compressed sensing CS is sampled as by t frame video sequence x_tBy random measurement matrix φ_t? Spatial domain obtains sampled images.

Due to the fact that the above technical scheme of employing, compared with prior art, have the following advantages that and actively Effect:

1) present invention regulates measured value according to image sparse degree, it is thus achieved that optional sampling amount, with higher compression ratio Image is compressed, thus obtains the longer Web vector graphic time；

2) present invention uses multi-view image fusion to be obtained under different visual angles by video sensor network node The image information of Same Scene is in addition comprehensive, obtains target image the most accurate, comprehensive, reliable description, Thus improve tracking performance.

Accompanying drawing explanation

Fig. 1 is a kind of adaptive targets compressed sensing fusion tracking side based on video sensor network of the present invention The flow chart of method；

Fig. 2 is to be the FB(flow block) of the embodiment of the present invention；

Fig. 3 is to use tradition fixation measuring value compressed sensing algorithm to compare comparison diagram with the measurement of the present invention；

Fig. 4 is the root-mean-square error comparison diagram using tradition fixation measuring value compressed sensing algorithm with the present invention；

Fig. 5 is target following track result figure in the embodiment of the present invention.

Detailed description of the invention

The present invention is further illustrated with reference to the accompanying drawings with specific embodiment.

As shown in accompanying drawing 1-5, a kind of based on video sensor network the adaptive targets compressed sensing that the present invention provides Fusion tracking method, comprises the following steps:

S1: build the video sensor network test system of the video node including some different visual angles, pass through Mixed Gauss model GMM builds t frame backgroundWherein t=0,1,2 ...；

For t frame video sequence x_t, image can be expressed as t frame foreground imageWith t frame backgroundIt With, such as equation (4).

$x_t=x_t^f+x_t^b---\left(4\right)$

S2: according to video node each in some video node at the image sparse degree K of initial time₀, respectively Calculate the measurement initial value M of each video node₀Come its initial pictures sequence x_oIt is compressed perception CS to adopt Sample, initial pictures y after being sampled_oAfter carry out S4.

Image sequence x at the t frame of time-varying_tIn, can guarantee that again image to obtain higher compression ratio simultaneously The quality recovered.The degree of rarefication that measured value M allows for according to image is automatically adjusted size.

Image sparse degree function K_t,

K_t=(λ₀logS_t+λ₁)S_t (6)

Wherein, (λ₀,λ₁)∈R², λ₀,λ₁For function model parameter, R is Spatial Dimension size, S_tTarget Pixel number.

At initial time by equation (7) computation and measurement initial value M₀Carry out image sampling M₀≈K₀log(n/K₀)

(7)

Wherein n is signal dimension.

T frame video sequence x_tBy random measurement matrix φ_tObtain sampled images in spatial domain, be i.e. compressed Perception CS samples.Calculate the measurement initial value M of each video node respectively₀Come its initial pictures sequence x_oEnter Row CS samples, initial pictures y after being sampled_o。

Knowable to compressive sensing theory, the least M value can cause recovering poor image quality, and M value is excessive, adopts Sampled images data volume is big, and compression ratio is low, and calculating and transmission energy consumption can be caused bigger.Can from CS sampling experimental Show that the recovery image obtained with sampling rate m/n=10% can clearly comprise profile and the position letter of image Breath, for the segmentation of foreground target, the most at least needs the sampling rate of m/n=25%, and wherein m is for measuring square Battle array dimension, n is signal dimension.

S3: use the expectation E (σ of the variance of image_t) as threshold value, by the measured value M of t two field picture_t, Iteration obtains the measured value M of each video node t+1 two field picture_t+1, and by the measurement of t+1 two field picture Value M_t+1T+1 two field picture is compressed perception algorithm CS sampling, obtains t after the sampling of each video node Two field picture y_tAfter carry out S4.

This method uses the expectation of the variance of equation (8) employing image as threshold value.

$E\left({\mathrm{\&sigma;}}_t\right)=\sqrt{(N-M_t){\mathrm{\&sigma;}}_d^2}---\left(8\right)$

WhereinIt is target image x_dVariance, measured value M can be passed through_tControl the size of threshold value.

Represent and use norm 2 restoration errors value

IfThen M_t+1=β₀M_t (9)

IfThen M_t+1=β₁M_t (10)

Wherein,For the target image recovered, x_dFor target image,Represent and use norm 2 Restoration errors value；β₀≤1,β₁≥1,β₀,β₁Respectively reducing and amplification factor, α is regulation parameter；

S4: by t two field picture y after sampling_tRecover to obtain t frame video sequence x_t。

T two field picture y after equation (5) sampling can be passed through_tRecover to obtain t frame target imageWhereinFor The background of t frame after sampling.By the y after sampling_tThe background of t frame after the sampling of figure image subtractionIt is easy to The target image of t frame after sampling.

y_tFor sampling after image,The recovery process of CS is represented for the target image recovered, Δ.

Various visual angles target following based on video sensor network, the most further comprising the steps of in S2 and S3:

S51: after each video node is recovered to obtain t frame target image, use Unscented kalman filtering UKF Carry out the detection of target following information.

Concrete model is as follows

$x_t=\left[\begin{array}{cc}{I}_{2\&times;2}& {\mathrm{\&Delta;tI}}_{2\&times;2}\\ 0& I_{2\&times;2}\end{array}\right]x_{t-1}+\left[\begin{array}{c}\frac{{\mathrm{\&Delta;t}}^2}{2}{I}_{2\&times;2}\\ {\mathrm{\&Delta;tI}}_{2\&times;2}\end{array}\right]v_{t-1}---\left(11\right)$

z_t=Hx_t+w_t (12)

Wherein x_tTarget is at the state vector of t, I_2×2It is 2 dimension unit matrixs, v_t-1There is Gauss distribution Process noise, Δ t is interval time, z_tGround level coordinate figure z_t=[u_t,v_t].Assume that photographic head is demarcated, as Plane is H to the homography matrix of ground level.

Ground level coordinate figure [u to target_t,v_t] be updated, renewal process is as follows:

${\hat{x}}_t^{-}=F{\hat{x}}_{t-1}---\left(13\right)$

$P_t^{-}={\mathrm{FP}}_{t-1}{F}^T+Q---\left(14\right)$

$K_t=P_t^{-}{H}^T{({\mathrm{HP}}_t^{-}{H}^T+R)}^{-1}---\left(15\right)$

${\hat{x}}_t={\hat{x}}_t^{-}+K_t(z_t-H{\hat{x}}_t^{-})---\left(16\right)$

$P_t=(I-K_{t}H)P_t^{-}---\left(17\right)$

WhereinBeing respectively t two field picture dbjective state and the covariance matrix of prediction, F is dbjective state Transfer matrix.

This method uses equation (18) weighted value as confidence level η_jThe performance of target following weighed by measured value. Comprise two factors: the number of 1, target image size, i.e. pixel；2, degree of accuracy P of target following_t.Mesh Dimensioning is the biggest, illustrates that Detection results is the best.P_tRepresent the uncertainty followed the tracks of.

${\mathrm{\&eta;}}_j=\frac{{\mathrm{\&rho;S}}_t}{\mathrm{trace}\left(P_t\right)}---\left(18\right)$

Wherein trace (P_t) it is the order seeking covariance matrix, ρ is normalized parameter, S_tObject pixel for detection Point number.

S52: the information of each video node detecting and tracking, is sent to Centroid and carries out data fusion.T frame ground Plane coordinates [u_t,v_t] and t frame image plane coordinate [x_t,y_tMapping relations between] are by H_jDetermine:

$\mathrm{\&lambda;}\left[\begin{array}{c}{u}_t\\ v_t\\ 1\end{array}\right]=H_j\left[\begin{array}{c}{x}_t\\ y_t\\ 1\end{array}\right]---\left(19\right)$

Wherein, λ scale parameter, H_jMapping matrix for jth pixel:

$H_j=\left(\begin{array}{ccc}{H}_{11}& H_{12}& H_{13}\\ H_{21}& H_{22}& H_{23}\\ H_{31}& H_{32}& H_{33}\end{array}\right)---\left(20\right)$

For each video node, the ground level coordinate position [ut, vt] of target can be calculated:

$u_t=\frac{H_{11}{x}_t+H_{12}{y}_t+H_{13}}{H_{31}{x}_t+H_{32}{y}_t+H_{33}}---\left(21\right)$

$v_t=\frac{H_{21}{x}_t+H_{22}{y}_t+H_{23}}{H_{31}{x}_t+H_{32}{y}_t+H_{33}}---\left(22\right)$

S53: for each video node, the weighted value of each pixel after being merged by (23) and (24):

G(u,v)=∑_jw_jG_j(u,v) (23)

$w_j=\frac{{\mathrm{\&eta;}}_j}{{\mathrm{\&Sigma;}}_j{\mathrm{\&eta;}}_j}---\left(24\right)$

Wherein G_j(u, v) is the target location of jth pixel, and (u, v) for the position of target, η after merging for G_jFor jth Individual pixel confidence level, w_jFor based on confidence level η_jFusion after the weighted value of jth pixel.

In order to verify the effect of the method, the present embodiment builds indoor test systems based on 8 video node. Video node resolution be 640 × 480pixels frame per second be 25fps, 15 meters of visuals field, 60 degree of visual angles, with 13 seconds track persistent period.As it is shown on figure 3, conventional compression perception algorithm is measured is fixed on 15% than m/n, no Can be adjusted according to image sparse degree, and self-adapting compressing perception algorithm, according to image sparse degree, especially It is when image sequence 150-220 frame, automatically reduces measured value, thus obtain higher compression ratio.But from Adaptation compressed sensing is owing to reducing the hits of image, therefore the quality of image restoring passes through root-mean-square error (RMES) weigh, as shown in Figure 4, the most traditional compressed sensing algorithm.

And the image of self adaptation CS compression algorithm can well obtain the status information of target, each video is saved Point is followed the tracks of the target information of various visual angles and is carried out data fusion at Centroid, can obtain and follow the tracks of knot the most accurately Really.

As it is shown in figure 5, X, Y-axis is respectively abscissa and vertical coordinate, from various visual angles the track of fusion tracking relative to The measured value of single-view, is closer to real track.This illustrates that method in this paper is reducing communication While amount, can effectively target be tracked.

In sum, the self-adapting compressing perception algorithm that the present invention is carried out based on testing result is dilute according to image Dredge degree regulation measured value, thus obtain optional sampling amount and with higher compression ratio, image can be compressed, Reducing and calculate and the data volume of communication, the multi-view image in conjunction with tracking creditability and testing result merges, More preferable tracking accuracy, and longer Web vector graphic time can be obtained.

The disclosed above specific embodiment being only the present invention, this embodiment is only the clearer explanation present invention Used, and not limitation of the invention, the changes that any person skilled in the art can think of, all should fall In protection domain.

Claims (6)

1. an adaptive targets compressed sensing fusion tracking method based on video sensor network, it is special Levy and be, comprise the following steps:

S1: build the video sensor network test system of the video node including some different visual angles is logical Cross mixed Gauss model GMM and build t frame background, the t frame video sequence x that test obtains_tIt is t Frame foreground image and t frame background sum, wherein t=0,1,2 ...；

S2: according to video node each in some video node at the image sparse degree K of initial time₀, point Do not calculate the measurement initial value M of each video node₀Its initial pictures sequence is compressed perception CS adopt Sample, obtains initial pictures y after the sampling of each video node_oAfter carry out S4, wherein M₀≈K₀log(n/K₀), N is signal dimension；

S3: use the expectation E (σ of the variance of image_t) as threshold value, by the measured value M of t two field picture_t, Iteration obtains the measured value M of each video node t+1 two field picture_t+1, It it is target Image x_dVariance；

WhenThen M_t+1=β₀M_t；

WhenThen M_t+1=β₁M_t；

Wherein,For the target image recovered, x_dFor target image,Represent and use norm 2 restoration errors values；β₀≤1,β₁≥1,β₀,β₁Respectively reducing and amplification factor, α is regulation parameter, and N is The picture size size that video node obtains, i.e. pixel number；

And by the measured value M of t+1 two field picture_t+1T+1 two field picture is compressed perception CS sampling, Obtain t two field picture y after the sampling of each video node_tAfter carry out S4；

S4: by t two field picture y after sampling_tRecover to obtain t frame video sequence and recovery obtains t Frame target image.

A kind of adaptive targets compressed sensing based on video sensor network Fusion tracking method, it is characterised in that further comprising the steps of in S2 and S3:

S51: after each video node is recovered to obtain t frame target image, use Unscented kalman filtering UKF Carry out the detection of target following information；

S52: the target following information many video node detected is sent to Centroid and carries out data fusion, T frame ground level coordinate [u_t,v_t] and t frame image plane coordinate [x_t,y_tMapping relations between] are:Wherein H_jMapping matrix for jth pixel:Draw Ground level coordinate [the u of each pixel_t,v_t]:λ is Scale parameter；

S53: for each video node, by G (u_,V)=∑_jw_jG_j(u_,V) andAfter being merged Target following information in the weighted value of each pixel, wherein G_j(u v) is the target position of jth pixel Putting, (u, v) for the position of target, η after merging for G_jFor jth pixel confidence level, w_jFor based on confidence level η_jFusion after the weighted value of jth pixel.

A kind of adaptive targets compressed sensing based on video sensor network Fusion tracking method, it is characterised in that sampling rate m/n is at least 25%, wherein m is calculation matrix Dimension, n is signal dimension.

A kind of adaptive targets compressed sensing based on video sensor network Fusion tracking method, it is characterised in that described image sparse degree is: K_t=(λ₀logS_t+λ₁)S_t, wherein (λ₀,λ₁)∈R²,λ₀,λ₁For function model parameter, R is Spatial Dimension size, S_tThe pixel number of target.

The compression sense of a kind of adaptive targets based on video sensor network Know fusion tracking method, it is characterised in that S4 also includes pass throughT after sampling Two field picture y_tRecover to obtain t frame target imageWhereinFor t frame background after sampling, Δ represents CS Recovery process, φ_tThe random measurement matrix built for t.

A kind of adaptive targets compressed sensing based on video sensor network Fusion tracking method, it is characterised in that described compressed sensing CS is sampled as by t frame video sequence x_tBy Random measurement matrix φ_tSampled images is obtained in spatial domain.