Research On Indoor Positioning Method Based On Probability Model And Deep Learning

In today's highly developed artificial intelligence and 5G technology,indoor positioning technology has increasingly become a rigid demand in the fields of people's life,industrial production,smart city construction,and security,and plays a pivotal role.Currently,various indoor positioning technologies are flourishing,for example,indoor positioning technologies based on wireless communication,inertial navigation,lidar,vision,etc.However,indoor positioning has not been unified and standardized like outdoor positioning technology.Traditional purely inertial navigation positioning systems can achieve autonomous positioning without relying on deployment equipment,but the inherent drift of inertial devices causes cumulative errors in the positioning results.The map matching data fusion algorithm corrects the accumulated error on the basis of guaranteeing autonomous positioning through boundary constraints on the map.However,the map matching positioning technology of the traditional particle filter algorithm still has shortcomings in positioning accuracy and map utilization.Therefore,the paper studies the indoor positioning method based on map matching to solve the above problems,and constructs an indoor autonomous positioning framework based on map and inertial navigation data.The main research work is as follows:(1)Starting from the map representation method,the paper abstracts the map data as a collection of reachable points,and extracts its neighboring points in various directions to form a point sequence with map context information.(2)The paper discusses the intelligent extraction algorithm of map data.Inspired by the language model in the field of natural language processing,the paper explores the application of word embedding methods in the language model to map data,constructs an intelligent extraction algorithm for map data based on the dynamic word embedding model,and expresses the reachable points of the map as containing contextual distance and Real vector of direction information.(3)The paper combines particle filter algorithm,dynamic word embedding and variational autoencoder model to build an indoor positioning algorithm based on probability model and deep learning.The algorithm uses particle swarms to simulate the movement distribution of pedestrians;uses dynamic word embedding models to learn map features and embeds map information into particle states;uses variational autoencoders to learn the distribution characteristics of particle swarms,combined with the historical trajectory of particle swarms,Improve positioning effect.The paper uses the real scene map to train the dynamic word embedding model,which proves the effectiveness of the dynamic word embedding model for map data processing;in the two real scenes,the truth value label experimental site is set,and a total of 610 valid truth value trajectories are collected for training change Sub-self-encoders verify the feasibility of positioning algorithms based on particle filters,dynamic word embedding models and variational self-encoders for improving positioning accuracy.