Research On Particle Filtering Localization Algorithm Based On Deep Learning

In an unknown environment,Simultaneous Localization And Mapping(SLAM)is one of the most important research topics in the field of robot navigation.With the continuous development of deep learning technology and its excellent performance in image processing in recent years,visual SLAM based on deep learning technology has gradually attracted more and more attention.As the most important component of SLAM,localization technology research is one of the important directions of robotics research.Based on the present stage deep learning algorithm and the technique of robot localization,on the basis of in-depth study,the traditional robot localization based on the Bayesian filtering algorithm are discussed,analyzing the imperfection of the traditional localization algorithm.In this paper,a new algorithm model is proposed by combining the deep learning algorithm with the traditional probability model.Combining the traditional probabilistic model algorithm with neural network,the new model algorithm has strong learning ability and good location effect.The specific research contents of this paper are as follows:(1)Firstly,the traditional Bayesian filtering localization algorithm is described systematically,and the traditional Bayesian filtering localization algorithm is experimented and analyzed,and the limitations and shortcomings of the traditional algorithm are found.(2)In this paper,the Gradient Propagation Particle Filter Network(GPPFN)is proposed,which embedding Particle filtering into the neural Network to carry out end-to-end training.The model is mainly composed of observation model and motion model,each model is composed of several modules with network structure,which can be optimized by neural network to achieve the overall optimal result.In this paper,the working procedure of GPPFN network is described in detail,and it is applied to three simulation environments of Deep Mind Lab.By comparing the model with the training data of other localization methods,it is found that the GPF network is better than other models in error,error rate and localization effect.(3)The plurality of information input can improve the accuracy of robot positioning.The input of GPF network model is only observation information and motion information.In order to make full use of the information contained in the map,we introduce a Semantic Map Particle Filter Network(SMPF-net)based on Semantic Map.The experiment is based on SUNCG data set and conducted in the House3 D simulator in an environment close to the real environment.In this model,semantic map,observation information and odometer information are used as input sources,in which semantic map is marked differently.The particle filter network uses different sensors as the input source to compare the localization effect,and further compares the method with other localization methods.By changing the number of initial particles and setting the value of the initial state,the accuracy of the final positioning of the robot is obtained.Through the comparison of many aspects,it is concluded that the positioning accuracy of the particle filter network is higher than other models.However,the particle filter network also has some shortcomings.The model lacks a decision-making mechanism,and it cannot judge whether resampling process is needed during training.