| In practical applications of signal,image or point cloud processing,the working environment or channel may be subject to interference from intelligent devices,human activities,natural phenomena,and other factors.As a result,signals or samples may be corrupted by a wide range of noise.How to estimate information of interest from noisepolluted data is a common concern in modern signal processing,machine learning and computer vision,which is called robust estimation.At present,most of the research on robust estimation is aimed at a certain noise,that is,it is assumed that noise follows some a priori distribution.However,in engineering practice,it is usually difficult for practitioners to know the statistical information of noise in advance,which makes the realistic model deviate from the hypothetical model.Relevant studies show that even slight deviation may lead to significant degradation of the performance of the estimation algorithm.Aiming at the diversity and uncertainty of noise,main contents and contributions of this thesis are as follows:1.Introduces the classical least mean square(LMS)algorithm and its mean square error optimization criteria,the robust algorithm based on M-estimation,the robust algorithm based on information theoretic learning and their optimization criteria,analyzes the existing problems of algorithms,and illustrates the correlation among algorithms.This thesis introduces the concept of multi-kernel correntropy recently proposed in the field of information theoretic learning,analyzes its differences and advantages with other algorithms,and points out its limitations,which lays a theoretical foundation for two improved multi-kernel correntropy algorithms proposed in subsequent practical applications.2.In order to deal with the problem of performance degradation of traditional distributed estimation algorithms due to diverse and unknown noise interference in distributed network deployment environment or communication,a robust diffusion semi-parameter adaptive estimation algorithm based on multi-kernel correntropy is proposed.In order to overcome the limitation of offline updating of the existing multikernel correntropy,an online strategy is designed to update the variable parameters of the adaptive cost function,which expands the applicability of multi-kernel correntropy in adaptive filtering,unsupervised learning,online learning and other scenarios.The convergence of the algorithm is verified by theoretical derivation and the range of step size parameter is obtained.In addition,the proposed algorithm is compared with other advanced algorithms in different noise scenes in the simulation experiment,and the effectiveness of the algorithm is verified.3.Two key challenges in the application of robust estimation in 3D point cloud registration are discussed,namely,the large number of anomalies in the observed data and the non-smooth tail of the error distribution.To solve the above problems,this thesis constructs a multi-kernel correntropy optimization problem with fiducial points,and introduces a natural geometric property in point cloud registration,namely strong second-order spatial compatibility,to resist the interference of a large number of outliers on the estimation and alleviate the locality problem.Through the multi-kernel parameter design of error sensing,the algorithm can adapt to different error distributions.The rigid transformation is re-parameterized to Lie algebra,and the optimization is performed on the manifold to maintain the geometric constraints of the rigid transformation.Based on the above improvements,a compatibility-guided and error-aware robust geometric estimation algorithm is proposed.Compared with other advanced algorithms on open dataset,the superiority and effectiveness of the algorithm are verified.Based on the framework of robust estimation,this thesis discusses robustness in general estimation,distributed estimation and geometric estimation.The common problems faced by existing methods in distributed network and 3D point cloud registration are expounded,and the connection,difference and advantage between information theoretic learning methods and existing methods are analyzed.Another important contribution of this thesis is to propose two effective and stable algorithms based on information theoretic learning through problem modeling,algorithm design,theoretical analysis and experimental simulation. |
