| Autonomous underwater vehicle (AUV) has tremendous potential value and commercialprospects in the military, scientific and engineering fields, and it has become a hot issue in thefield of robotic research. Navigation is one of the key technologies which determine thedevelopment of AUV, and with the current of AUV toward the long-range and deep-sea thetraditional navigation methods are unable to satisfy the requirements of fast developing.Simultaneous localization and mapping (SLAM) is of great practical significance forimproving the autonomy of AUV Navigation because of its simple structure of externalsensor and no need for a priori environment map. In this paper, some key algorithms andtechnologies of SLAM for AUV are studied.The main research contents and achievements of this paper are as follows:Firstly, the motion control model, environment model and sensor observe model of AUVare discussed and simplified reasonably. And then, the description of AUV’s localizationproblem and mapping problem are proposed. On this basis, the probabilistic model of SLAMproblem and criteria of algorithm performance are established.EKF-SLAM algorithm has simple structure and rigorous mathematical theory for AUV’sSLAM problem in small-scale environment. Using UKF instead of EKF can eliminate thelinearization error introduced by EKF-SLAM in processing non-linear problem. But UKF islack of adaptive ability in practical application, because its estimate accuracy depends onprecise prior noise model. To solve this problem, a robust UKF-SLAM algorithm is proposed,which tracks actual noise statistical characteristics by using a multi-dimensional observenoise scale factor. As a result, it keeps high estimate accuracy and stability when the priornoise is unknown or the noise statistical characteristics are time-varying. The comparativesimulations between EKF-SLAM, UKF-SLAM and robust UKF-SLAM prove theeffectiveness of the new algorithm.For large-scale or dense-feature environment, the FastSLAM algorithm based on particlefilter (PF) is suitable for AUV SLAM problem. An improved particle filter based onmaximum likelihood estimation is proposed to solve the problem that traditional PFalgorithm can not obtain the estimation of a nonlinear system with the assumption ofnon-Gaussian non-stationary noise. In the improved PF algorithm, the real noise distributionis approximated by a series of independent weighted Gauusian noise sequences, and theimportance weights of PF are transformed to probability density function of each Gauusian noise sequence. Then the distribution parameters and weights of these noise sequences can becaculated by maximum likelihood estimation, and thus the weights of PF algorithm can bedetermined. By using maximum likelihood based PF algorithm in FastSLAM framework, theSLAM of AUV with non-Gaussian non-stationary noise is implemented. At last, theeffectiveness of proposed algorithm is verified by simulations.Data association is one of the core problems of SLAM, which directly affects theaccuracy of AUV’s localization and environmental map estimation. A data associationalgorithm based on fuzzy logic is proposed on the basis of analyzing and comparing principleand characteristic of common data association methods. In this algorithm, the relatedinformation between feature measurement and feature estimation is mined from their ellipses,and projected into the fuzzy set on the domain. A variety of information is fusioned andreasoned by establishing certain fuzzy rules, and the data association result is just the fuzzyoutput variable. The simulations indicate that the fuzzy data association has higherassociation accuracy and stronger anti-disturbance ability compared with traditionalalgorithms.In traditional feature based SLAM model, the space information is described by vectorsequence. As a result, this kind of model can not effectively express multiple observeinformation such as loss detecting, false alarm and observe uncertainty etc. for AUVnavigation problem in the cluttered environment, and the algorithm performance is severelyaffected. To solve this problem, a SLAM model based on random finite set (RFS) theory isproposed. In the novel model, the system state, observation and environment map in SLAMare all represented in the form of RFS. The estimation of joint target state variable is carriedout through probability hypothesis density (PHD) filter in the Bayes estimate framework, andthe PHD filter is realized by particle filter. Additionally, a target state extracting method basedon particle set time-delay outputting is putted forward to overcome the defaults of traditionalextracting algorithm. Simulations show that in the cluttered environment the RFS-SLAM canobtain higher estimate accuracy and stability compared with traditional SLAM model. |
PDF Full Text Request