Research On Bavesian Filters Based Simultaneous Localization And Mapping Algorithms For Mobile Robots

With the ongoing development of artificial intelligence technology,mobile robots are widely employed in many fields such as manufacturing industry,military denfense,aeronautics and astronautics,medical health,and domestic service.The simultaneous localization and mapping(SLAM),which is one of the most popular research topics in modern mobile robotics,is considered as a critical prerequisite for truly autonomous and intelligent robots.As various uncertainties exist in the actual work environments,the robot system oftern suffers from the measuremnt noises with non-Gaussian and heavy-tailed distributions or unknown prior parameters.In these complicated and uncertain environments,the performance of the conventional SLAM algorithms based on Bayesian filtering technology is severely degraded,whose estimation accuracy and computational efficiency are unable to meet the requirements of practical applications.To improve the performance under the complicated environment,the SLAM algorithms based on Gaussian filter,particle filter and probability hypothesis density filter are studied in this research work.The details are described as follows:(1)From the perspective of Bayesian filtering,the SLAM algorithms based on the Gaussian filter,particle filter and probability hypothesis density filter are summarized respectively.This provides the theoretical basis for the research work of improved algorithms in the subsequent three chapters.(2)The problem of a measurement system with a non-Gaussian and heavy tailed noise distribution is invesited,and a robust Gaussian filter SLAM algorithm based on statistical linear regression is proposed accordingly.Firstly,the squre-root cubature Kalman filter is used to predict the joint state vector.Then,the measurement update equation is recast as a linear regression formlation,and the generalized maximum likelihood estimator is used to calculate the associated weight matrix of the measurement residual.Finally,the posterior mean and covariance squared root factor of the joint state are estimated by the iterative re-weighted least square method.(3)The problem of sampling particles with low quality and computational efficiency in the FastSLAM algorithm is investigated,and a new UFastSLAM algorithm based on an improved proposal distribution and particle resampling method is proposed accordingly.Firstly,the robot state is augmented with both the control and measurement noise,and the optimal proposal distribution is estimated by utilizing the square-root transformed unscented Kalman filter.Then,the robot states of the particles are sampled according to the proposal distribution,and the feature maps of the particles are updated subsequently.Finally,the required number of particles is determined adaptively based on the KLD-resampling method in the particle resampling process.(4)The problem of complex cases with clutter and unknown measument noise variance is investigated,and a probability hypothesis density based SLAM algorithm with adaptive estimation of the measurement noise variance is proposed accordingly.Firstly,the joint posterior intensity of the feature map and measurement noise variance is represented by a weighted sum of a product of inverse Gamma and Gaussian distributions,and the parameters of the model are updated iteratively through the variational Bayesian approximation method.Then,the mixtures of the updated intensity are pruned and merged to obtain both feature state and number estimates.Finally,the weights of the particles are calculated based on the single feature strategy and the robot pose is estimated subsequently.