Research On Robot Localization Integrating Deep Learning And Particle Filter

With the continuous development of artificial intelligence,mobile robots have been widely used in various fields in recent years.Estimating the current position is the key task of these robots,which is called robot localization.Localization is an important process in robot technology,and it is also a research hotspot in the field of mobile robot.Probability based localization algorithms mainly include Kalman filter localization,Markov localization and Particle Filter localization.Among them,particle filter algorithm has higher accuracy and robustness in dealing with nonlinear non-Gaussian systems,and has become the mainstream algorithm at present.The traditional localization algorithm based on particle filter has difficulties and challenges in the construction and learning of probabilistic system model.With the rapid development of deep learning,particle filter localization algorithm based on deep learning has attracted extensive attention of domestic and foreign scholar.In this thesis,deep learning algorithm and traditional probability model are integrated to improve,and applied to the research of mobile robot localization.The specific research contents of this thesis are as follows:(1)Aiming at the robot localization problem,the traditional Monte Carlo Localization algorithm is studied.Since Monte Carlo Localization algorithm itself has particle degradation problem and cannot solve the problem of recovery in global localization failure,two improved algorithms,Augmented Monte Carlo Localization algorithm and Adaptive Monte Carlo Localization algorithm,are systematically expounded.And the Adaptive Monte Carlo Localization algorithm is experimented and analyzed in ROS environment to verify its effectiveness in global localization.(2)An Adaptive Soft-Resampling Particle Filter Network(ASRPF-Net)is proposed.Based on the framework of recurrent neural network,it realizes the overall differentiability of the network,allows end-to-end training and avoids the difficulty of traditional model learning.The network is mainly composed of motion model,observation model and adaptive softresampling model.They are fused into an RNN unit,and a particle filter algorithm is encoded in a recurrent neural network to learn the model of end-to-end timing state estimation.This thesis introduces the working steps of Adaptive Soft-Resampling Particle Filter Network in detail,and applies it to House3 D and Deep Mind Lab experimental environment.Comparing the proposed model with other methods,it is found that the Adaptive Soft-Resampling Particle Filter Network is superior to other methods in error,localization success rate and localization effect,and performs well in the global localization task.(3)In order to alleviate the need for annotated data and the high cost of obtaining the real state of the model,an Adaptive Soft-Resampling Semi-supervised Particle Filter Network model is proposed,which can improve state estimation when most of the true states is unknown.Because the resampling process of particle filter is non-differentiable,zero gradient will be generated during training and the gradient calculation will be stopped,which cannot provide gradient information for subsequent training.The model adopts a differentiable resampling method,so that the gradient information generated by particles in the training process can be propagated and the localization accuracy of the robot can be improved.The simulation experiment is carried out using the simulation environment in Deep Mind Lab.when a large part of the ground truth data is unknown,the proposed algorithm is better than other models in the state estimation task of global localization and tracking.Moreover,the proposed algorithm further improves the particle degradation problem,so as to improve the accuracy and robustness of localization.