论文研究-基于Zynq-7000的V4L2双帧缓存驱动设计 .pdf

所需积分/C币:11 2019-08-15 534KB .PDF
评分

基于Zynq-7000的V4L2双帧缓存驱动设计,王宇,胡若澜,V4L2是Linux2.6版本的一大特点,它主要为视频驱动程序的开发提供了相关的协议标准。Zynq-7000处理器结合了ARM架构和FPGA的特点,其软硬件��
山到花论文在丝 http:/www.paper.edu.cn Lcgcndrc basis function expansion mcthod ethods 1.1 The classical tvar method The dctcrministic-rcgrcssion TVAR modcl is a spccial AR modcl whose associated modcl paramctcrs arc timc-varying. Thc TVaR modcl of ordcr p is given by ()=∑a1(t)y(t-)+e(t) where t is the sampling index or time instant of the output signal y(o), the term e(t)is assumed to be a sequence of independent and normal distributed random variables with zero mean and a 100 variance d2(i.d. N(0.02) and a, (are the time-varying parameters One solution to the time-varying estimation problem in Equation(1) is to expand the time-varying coefficients a; (() onto a set of basis functions m (t) for m=1, 2,, M,such that the following representations hold 105 where the Ci, m represent the time-invariant expansion parameters of the basis functions m((),M is the dimension of the basis functions. Substituting(2)into (1), we can obtain (3) By denoting z()=[x1(t),x2(1)…,zn(t 110 H(t)=y(t-i)x(), e.=②,⊙ ⊙=@1,⊙2,, Equation (3 )can then be expressed as 115 ()=H()@+e(t (4) whcrc thc uppcr script T is thc transpose of a vcctor or a matrix Equation (4)is a standard linear regression model and indicates that the TVaR model can now be converted into a time invariant model which can be resolved using a classical Ols algorithm, since Ci m are not functions of time. The task simplifies to solving coefficients Ci. m which 120 can be easy to estimate using an Ols algorithm The next step is to select appropriate basis functions so that the time-varying parameters can be effectively estimated. While a number of basis functions such as Fourier bases Walsh and haar functions, Legendre polynomials, wavelets and radial basis functions can be applied to general time-varying parametric modelling problems 23.24. 23.20, 6 there is not a specific guideline on how to 125 choose thc propcr basis function for timc-varying modelling problcms from thesc available basis functions. In fact, different basis functions possess its own unique tractable ability and 山到花论文在丝 http:/www.paper.edu.cn approximated propcrtics, for cxamplc, lcgcndrc polynomials pcrform wcll if thc paramctcrs arc smoothly and slowly varying with time e g, sinusoids whereas Walsh and Haar functions behave well for time-varying coefficients with fast transients and burst-like dynamics 6. 233 130 Most physiological signals inevitably involve both fast and slowly variations at different stages of thc signal In ordcr to alleviate the dilemma of determining an appropriate basis function that has to be highly relied on a priori information of the signals, and also to achicvc the time-varying modelling algorithm which is more flexible and can rapidly track both fast and Slowly variations, a novel TVAR modelling method by expanding time-varying coefficients onto 135 multiscale radial basis functions scheme is proposed in this study, where properties of different types multiscale basis functions with good generalization propcrty exploited might be morc appropriate for dealing with the fast and smoothly changing dynamics 1.2 Multiscale radial basis functions Radial basis functions are real-valued functions which depend on the distance between th 140 input pattern and a point c that is defined as the center of rBFI25.27. A function is a radial basis if its output depends on the distance of the input and the center. Typically, a radial basis function is usually realized by a gaussian distribution function. The gaussian rbf is an ideal choice since it is infinitely smooth and have good approximation properties [28] which can be presented as follows 271 145 p(x-cp (5) where x is an input vector, c is the location (or translation) parameters to determine the kernel center of Gaussian radial basis function, o is the scale (or dilation) parameters of the basis function to determine the kernel widths, and denotes the Euclidean norm One problem for time-varying model estimation by expanding time-varying coefficients onto 150 MRBF is how to dctcrminc unknown paramctcrs of MrBF including ccntcrs and scales. Thc centers of the rbf determine the location of rbF in the whole time-varying parameters. In order to make the rbf distribute throughout the time-varying parameters, and ensure that each part of the time-varying parameters is estimated accurately, the centers of the rBF are distributed in the time-varying parameters uniformly 2o The expression of kth rbF and the center of rBF are 155 dcfincd by 291 4(x-c1=e? I-c ,=1,2,…,M (6) 2 k×N M where M is the dimension of rbf, ci, is the center of the kth rbf, ol is the scale of the kth rBe, N is the length of sampling data 160 As to the issue of determining proper scales, a modified particle swarm optimization(Pso) can be performed to determine the optimal scales of rbf. the detailed procedure of the modified PSO algorithm will be presented in the next section 1.3 The modified Pso algorithm The PSo is a stochastic, population based optimization algorithm which can be exploited to 165 determine the optimal scales of MrBF automatically for time-varying parameter estimates. The 4 山到花论文在丝 http:/www.paper.edu.cn vclocity and the particlc's position updating Equations of the traditional PSo algorithm arc presented as below /301 (h+1) =×v1)+6×C1×(Ce-4)+6xc2×(gt (8) h+1) +vh+) 170h the iteration index. I <h<Ir. i is the maximum number of iterations the particle index, lsis, I is the particle size v( h) the velocity of ith particle at hth iteration The ith element of v( h) is in the range position of ith particle at hth iteration. The jth element u;(ht of u is in the range 175 rl 2 rior the inertia weight uniform random number between 0 and 1 CI, c2 the acceleration coefficients Chep the local best position of the hth iteration 180 ghest the global best position after the hth iteration Chen ct al. -discusscd that using timc-varying acceleration cocfficicnts can help particlcs to avoid local minima and converge quickly to a global minimum. As in the initial stages, a large cognitive component 0 x C x(Chest -u(h)and a small social component 0xc2 x(ghest -u;)) arc nccdcd to help particles to avoid local optima and whilc, in the latcr stagcs, a small 185 exci*(c(h\ a a large 6xC2 x(gbest ui )are needed to help particles to converge (h) best quickly to a global optima. Thus, we can adopt this mechanism which c, is changed from 2.5 to 0.5 and c, varies from 0.5 to 2.5, respectively, where the definition of c, and ca can be calculated as below 2×h CI 2.5 (10) IxH 2×h 190 C,=0.5+ (11) 1×H ve. If the velocity in Equation(8)approaches to zero, it is reinitialized randomly to the maximum locity as follows [29] ±6xy×Vn (12) where v(+ is the jth element of v /h+, y=0. 1 is a constant, and 0 is the uniform random 195 number between 0 and 1 If the inertia weight of the traditional Pso algorithm is a constant, the optimizing precision is poor for certain functions. To overcome the defect of optimizing precision and improve the convergence performance for the conventional Pso algorithm, the modified Pso algorithm which modified the inertia weight of traditional velocity updating Equation is proposed in the present 200 study In PSO algorithm, a proper inertia weight can provide a good balance between global and 山到花论文在丝 http:/www.paper.edu.cn local explorations and hclp to dctcrminc thc optimal results rapidly. In traditional PSO algorithm however, the inertia weight is usually a constant, which may be inflexible in searching the optimal particles and even leads to inaccurate results. Thus the modilied pso algorithm assigns the inertia 205 weights values of each particle in relation to its fitness function to find optimal particles more F-F 0)=1-6× (13) where 0 is uniform random number between 0 and 1, F is the fitness value of the ith particles, and best Is the global optimal fitness value. The modified inertia weight varies with the fitness value 210 of each particle and can thus make the PSo algorithm more flexibly and accurately in searching the optimal results. 1.4 The determination of the scales The traditional rbf based basis function expansion method adopts the single scale rbf. the advantage of the conventional single scale rbf method is that the model is easy to construct The 215 singlc scalc RBF modcl, howcvcr, may lack good generalization propcrtics. Whilc the rbf of different scales can possess different tracking features and flexible generalization performance Thus, multiscale RBF expansion method which is much more flexible as it exploits different kinds of scales for tracking properties can be employed to effectively capture multiples of time-varying cocfficicnts. A modified PSO algorithm can then bc cxploitcd to dctcrminc optimal rBF scales 220 which are appropriate for specific signal. The candidate scales can be given as follows,33/ T2v2-Sk (14) where N is the length of the sampling data, M is the dimension of the rbF. sr is an arbitrary integer which nccds to bc tuned, whcrc the maximum valuc of s, is usually undcr 10.So the particle position u, can be represented as a vector whose dimension is M and elements are Sk, that 225isl;-「S1,S2,…,SM1 ase the searching steps of optimal scales of MRBF using the modified PSO can be summarized follows. Pso initialization Gencrate thc particles u randomly in the rang [Umin, U mat ],1<isI. Sct the initial 230 velocities v(o)=0. 1sisI where 0 is the zero vector of dimension M. The initial fitness valuc of local best position F(Chev and fitness valuc of global best position F(ghey ) arc both roes Iteration loop: For(h=0; hsH; h++)i Calculate the fitness values ofeach particle 235 Fork<isl. calculate the fitness value F of u), Correlation index which can measure the model lilting performance is adopted to evaluate the fitness value of each particle. Correlation index is adopted to measure the model fitting performance. The greater the correlation index value the better the model fitting result. The correlation index is given by 34) 山到花论文在丝 http:/www.paper.edu.cn ∑(Y()-Y()2 R2=1-1= (15) ∑(()-) 240 where R is the correlation index, N is the length of the sampling data. Y is the predicted output of the original output y Update the local and global best positions 1)For(=0;i≤r;i+) If( F> F(Chest)) 245 (C (h) F 112 End if End for 2) If( F(Chel)> f(ghest )) 250 End if: Update velocities and positions o/particles 1)For(i-0;i≤;计++) 255 O=1-0× F-F(s =O×v+6×1×(Cn-1n2")+bxc2×(gmt-l1) For(=0:j≤Mx++) =0) If(<0.5 260 h+1 =b×y×V y End if: 265 If( [h > Else if( 山到花论文在丝 http:/www.paper.edu.cn End 270 End for End for. 2)For(i-0;i≤I;i++) For(=0:j≤M+) 275 If(u hiDL>Uma. (h+1) Else if(u(<U End if: 280 End for End for cnd of iteration loop This yields the global best position ghe) the mRBF expression with the optimal scales,and the time-invariant coefficients 0, respectively. Thus time-varying coefficients a, (0) In 285 Equation( 1)can be easy to recovered from@ and MRBF 1.5 Model identification and parameter estimation The time-invariant coefficients in the model (1) approximated by mrbF expansion can be easy to estimate using the ols algorithm and retaining only those coefficients that are considered to be significant. One approach that can be applied to determine the unknown parameters 290 including the optimal centres and widths of mrBF is adoptcd by a modified PSO algorithm 3oI Once time-invariant Cim are calculated, the resultant estimates will then be applied to recover the time-varying parameters a, (4) by Equation(2)in the TVAR model(1 In order to determine the proper number of Gaussian basis functions, a final prediction error (FPE)can be adopted 33. The FPE criteria for multiple scale basis functions for the AR process FPE N+p+o2+p2k, (16) where N is the length of the signal, p is the number of significant model terms, o4 is the prediction crror, and k, indicates the numbcr of Gaussian basis functions in the L multiple basis functions 1.6 Spectral estimation 300 Once the time-varying parameters in the TVaR model (1)are calculated, the time-varying power spectrum estimate can be easy to obtain in terms of the estimated TVAR coefficients Define a; (t)as the estimate of a: (0), and o2 as the estimate of a.The time-dependent spectral function derived from the TvAR model(1) is then calculated from the estimated coefficients and the variance of residual as follows 35.36 山到花论文在丝 http:/www.paper.edu.cn s(t,f (17 l-∑n4()e 2/s where is the sampling frequency, a: (Oare the TVaR parameter estimates at time and 0. is the variance of the estimated observation error process. It should be notable that the spectral function in Equation(17) is continuous function of frequency f and thus can be used to cvaluatc spcctral estimates at any desired frcqucncics from 0 up to the nyquist 310 frequency f/ 2, even though this does not improve the frequency resolution of the spectrum infinitely. However, the frequency resolution is naturally not infinite, but can be determined by the underlying proper model order and the associated parameters. The resolution of parametric approaches is higher than classical FFT-based sI extrapolation of the autocorrelation sequences e ectrum estimation approaches due to the implicit 315 2 Simulation example Prior to applying the proposed approach to real eeg data analysis, a benchmark on an artilicial time-varying signal was employed to test the reliability and illustrate the advantage of using MRBF over RLS and Legendre basis function expansion approaches. We consider an artificial EEG signal that has thc following form 2|Fsn(2z)+2|5si(2fm,tc∈(0.2) 4 sin(2zfpt)+z sin (2rft te2,4 320 (0=2 sin(2rfaD), ∈[45) (18) le sin(2rfat)+1t sin(2), tE[5, 6) 24 sin(2rS, ∈|6,8 wheres=0.5, U=0. 25, frequency components of fe=6, fa=12,.B=25,f, =40 Hz are meant to emulate the theta, alpha, beta and gamma bands of eeg data, respectively. The simulated EEG Signal that consists of four sinusoids with varying frequency and noise was generated. The frequencies of the sinusoids corresponded to typical EEG frequency components(see Figure 1) 325 The above signal was sampled with a sampling interval 0.0l, and thus a total of 800 observations were obtained. To obtain a more realistic simulation, an additive gaussian white noise sequence with mcan zero and variance of0. 16, was then addcd to the 800 data points The order of Tvar models was determined according to the fpe criteria in Equation(16).A ninth order TVAR model selected by FPE was adopted to describe the time-varying artificial EEG 330 Signal using the proposed method. The Gaussian MRBFs were employed to approximate the time-varying coefficients a; ()with i=1, 2, ,9 and t =1, 2, 800. A modified PSO approach and an ols algorithm were used to determine the unknown parameters in the model including the optimal centres and width of gaussian basis functions and model parameter estimation. We compared power spectral density(PSD)estimates obtained from the proposed PSd estimates 335 method with those generated by classical estimation techniques such as the rls in addition to single sets of Legendre basis functions, where the ten order TVAR model was determined by Equation(16) using the rls method and Legendre basis function approach for PSD estimation The simulation example aims at comparing the time- frequency resolution generated using three methodologies 山到花论文在丝 http:/www.paper.edu.cn 340 Figurc I shows a comparison of timc-frcqucncy spectra among thrcc mcthods for the simulated EEG signals given in Equation(18). We compare MRBF expansion method [ Fig. I(c) to rls method [Figure l(a)] and Legendre basis function expansion method [Figure 1(b)]. we can see that the rls method has the lowest frequency resolution for each frequency components Understandably, the Rls algorithm creates a tracking lag in thc parameter estimates of TVAR modcl. Thc frcqucncy rcsolution of Lcgcndrc polynomials expansion mcthod is highcr than that of RLS method because identification of nonstationary processes can be more effectively handled by using basis functions than by the rls approach. The MrBF expansion method dose give a better time-frequency spectra for the four different frequency components with higher resolution than the other two mcthods, bccausc the mrbf expansion mcthod can track rapidly simulated EEG signals 350 while traditional time-frequency methods like adaptive Rls algorithm and Legendre basis function expansion method often fail when the nonstationary signals change fast enough The time-frequency resolution was quantitatively evaluated by comparing the mean absolute error(MAE)and root mean squared error(Rmse)for four sinusoidal components at f6, 12, 25 and 40H7. Thc MAE and rmse of cach frcqucncy componcnt arc dcfincd as 355 M()=()51)∈(y (19) RMSE )1s(tf)-S(t,水儿,K∈(.a,B,y (20) N where N is length of the data S(i, )is the estimated power spectrum of the real power spectrum S(l,f)from the simulation signal The mae and rmse of the timc frcqucncy cstimates for four frcqucncy componcnts in 360 Equation(18), with respect to the corresponding true values, are estimated and shown in Table I From Table 1, both mae and rmse values of the proposed method are all smaller than the other two methods, which indicate that the time-frequency estimation accuracy of the proposed method is higher than the other two methods. From these results, it is clear that the proposed method can corrcctly cstimatc the timc-frcqucncy spectra of nonstationary signal and obtain much highcr 365 resolution in both time and frequency domains simultaneously £ 15 Time(s)

...展开详情
立即下载 最低0.43元/次 学生认证VIP会员7折
举报 举报 收藏 收藏
分享
img

关注 私信 TA的资源

上传资源赚积分,得勋章
相关资源标签
相关内容推荐