Research On Several Key Techniques About Signal Estimation And Filtering

Signal estimation and signal filtering theories and techniques are widely used in information fusion systems. In some sense, signal estimation and signal filtering are both dealing with noise. Here the noise is in general sense, including the disturbance signal and some uncertainties of the signal state. The signal has noise, and the measurements also have noise, thus we need to estimate the true signal. Target tracking is actually target's state signal estimation which is stochastic signal processing in essence. Signal filtering is to remove the noise from the signals in some sense.The focus of this dissertation is estimator design for ground target's state signal estimation and measurement signal estimation and the linear phase two dimensional FIR filter design which are in the area of signal estimation and filtering. Five main results are achieved given as follows:(a) Because maneuvering vehicle state estimation with road constraint is highly nonlinear problem, traditional approaches just consider the roads without width. To overcome this problem, a distributed estimator with feedback is proposed. This estimator is based on the proposed mileage-lateral distance road model. With the image sensor, we get the estimation of the vehicle's vertical distance to the central line of the road (called lateral distance) and with the radar, we can get the lateral distance estimation. With the fused lateral distance estimation, the parameter accuracy of local estimators are improved. The system outputs both of the lateral distance estimation and the mileage estimation that the vehicle has traveled. The computer simulation results show that the estimator presented reached much better estimation accuracy comparing with the typical traditional method.(b) To overcome the defects of traditional system's state model that can only have good fitting for certain motion, an adaptive motion model is presented in this paper. The relationship between the estimated target's acceleration and the variance of the noise in the motion model is given based on the expectation of the absolute value of the estimated target's acceleration. Then the adaptive motion model based filter is presented based on the proposed adaptive motion model. The simulation results show that the adaptive motion model based filter achieves the nearly estimation accuracy as that two models based interacting multiple-model (IMM) can achieve. Such estimation accuracy is much better than the performances that traditional one motion model based filters can achieve.(c) Traditional measurement estimation algorithms rely on state model of the system very much and it doesn't work well for varieties of ground truth. The neural network based centralized multi-sensor signal estimation fusion and the neural network based batch filter (NNBF) are proposed. For the first algorithm, the coefficients of the filter is adjusted adaptively measurement signal training based on neural network at the next time is estimated using the proposed filter. For the second algorithm, the coefficients of the filter are adjusted adaptively based on the neural network training using the measurements data from the initial time to the current time of each sensor, then the distributed signal estimation fusion algorithm is presented based on the NNBF. In computer simulation, the neural network based centralized multi-sensor signal estimation fusion algorithm is compared with Extended Kalman filter, and NNBF based distributed fusion algorithm is compared with Kalman filter. The simulation results show that both proposed algorithms have better estimation accuracy.(d) Based on the transfer function of the two dimensional band-pass FIR filter of type-â…¢given in frequency domain, the structure and algorithm of neural network with double sine basis function is proposed. The convergence theorem of the neural network with double sine basis function is presented, and the proof is given. The simulation results show that this approach achieves much better magnitude response than the window function based approach and the frequency sampling based approach. The magnitude response achieved by this approach is nearly as the ideal filter. And, the computational complexity of this approach is low.(e) Usually the image noise is in high frequency domain. The double cosine basis function based neural network image noise filter is proposed by removing the high frequency signals. Based on the transfer function of two dimensional FIR filter given in frequency domain, the structure of double cosine basis function based neural network is given. And the convergence theorem of this neural network is presented and proved. The simulation results show that comparing the proposed filter with the median filter, the proposed filter has better performance in smoothing and removing the image noise than median filter.