Research On The Back-end Of SLAM Based On Maximum Correntropy Criterion Filtering

Thanks to modern technology and artificial intelligence,using robotics to replace traditional productivity becomes a major research trend.The Simultaneous Localization and Mapping(SLAM)algorithm based on artificial intelligence is the product conforming trends and demands of the era.The SLAM algorithm has been widely used in human life in modern society and brought great convenience,e.g.,underwater detection,unmanned driving,and automated guided vehicle.The structure of the SLAM algorithm is mainly divided into front-end and back-end.The front-end primarily utilizes RGBD or stereo camera to collect data and helps the SLAM algorithm interact with the external environment.The back-end is responsible for further processing the information obtained from the front-end.And the back-end algorithms are mainly divided into two categories,i.e.,the Kalman filtering algorithms and nonlinear optimization methods.This thesis mainly focuses on the Kalman filtering algorithms.In the early research,the SLAM algorithm is mainly applied in laboratories with prominent characteristics,small environment scope,and simple structures.During this period,as an extension of adaptive filtering,the Kalman filtering algorithms make a significant contribution to the back-end research.However,as the application scenarios of SLAM gradually turn much richer and more complex,the shortcomings of the traditional Kalman filtering algorithms as its back-end are steadily increased.Traditional Kalman filtering algorithms are only suitable for Gaussian noises.However,the SLAM experiment environment is dominated by non-Gaussian noises,which can lead to a non-ideal filtering effect.This thesis mainly focuses on the non-ideal filtering performance of the traditional Kalman filtering algorithms under a small scene of relatively simple environment structure with non-Gaussian noises,and introduces the basic principles of the typical SLAM back-end algorithms.Combining the advantages of maximum correntropy in information theory under non-Gaussian environments,this thesis proposes the Kalman filtering algorithms based on the maximum correntropy criterion as the back-end of SLAM.The main contents of this thesis are as follows:(1)In this thesis,the multi-kernel generalized maximum correntropy algorithm was proposed to solve the stability problem of the traditional kernel adaptive filtering algorithms.Meanwhile,the quantized multi-kernel generalized maximum correntropy algorithm was proposed with the vector quantization strategy to reduce the computational complexity.Simulations on Mackey-Glass chaotic time series prediction are used to verify the accuracy and robustness of the proposed algorithms,and thus provide the foundations for developing the Kalman filtering algorithms combined with the maximum correntropy criterion in the SLAM back-end for non-Gaussian noises.(2)In indoor scenes with simple environment structures and distinguishing features,the Kalman filtering algorithms as the back-end of SLAM,often suffer from the issues of insufficient precision and failure in non-Gaussian noises.Therefore,this thesis uses the Kalman filtering algorithms based on the maximum correntropy criterion as the backend of SLAM,and performs the simulations under Gaussian noises and non-Gaussian noises.The simulation results indicate that the proposed algorithms are robust against non-Gaussian noises and improve the performance of the traditional Kalman filtering algorithms as the back-end of SLAM.