Research On Indoor Positioning Algorithm Based On Reinforcement Learning

In the fields of modern business,medicine,and life,people's demand for indoor locations is increasing.Traditional indoor positioning technologies have problems such as high cost and low accuracy.Map data is a type of data that can constrain pedestrian trajectories.An indoor positioning method incorporating map information can achieve low-cost and high-precision positioning.Based on inertial navigation technology,this paper combines deep learning and reinforcement learning to fuse inertial navigation data with map data,and proposes a map matching algorithm based on deep learning and a map matching algorithm based on deep reinforcement learning.Firstly,this paper studies the method of inertial navigation and map fusion.In order to solve the problem of manually extracting map features in traditional map matching methods,an automatic map matching algorithm based on deep learning is proposed.Deep learning is a representational learning method based on a large amount of data.Based on real-world data sets,this paper builds three sets of deep learning models to achieve map matching.They are CNN model,RNN model and CRNN model.During the research process,we collected 567 groups of pedestrian data through inertial navigation equipment,and after data enhancement,we compiled a labeled training data set.By comparing and analyzing the method based on inertial navigation alone,the positioning errors of the three models in this paper have been reduced by 59.9%,52.4%,and 64.9%,respectively.Experiments show that the map matching method based on deep learning model proposed in this paper realizes the fusion of map and inertial navigation data and meets the indoor positioning accuracy requirements.Then,this paper proposes a map-matching algorithm based on deep reinforcement learning based on the time-series characteristics of pedestrian inertial navigation data for local optimal problems in traditional map-matching algorithms.The algorithm takes deep reinforcement learning network as the basic framework,and designs the map matching process as a Markov decision process.For maps and inertial navigation data,the map cutting module,action space module,state space module,Designed for reward space modules.Then,the map data is converted to the positioning coordinate data by encoding and decoding,and the matching trajectory is completed.In the model optimization stage,this study also carried out a comparative experiment on the size of the coded map and the size of the pixel cluster,and analyzed the effect of the model on map matching under different conditions.Compared with the IMU model using inertial navigation alone,the positioning error of the deep reinforcement learning model is reduced by 71.4%,which avoids the phenomenon of the positioning trajectory passing through the wall.Experiments show that the map matching method based on deep reinforcement learning proposed in this paper can achieve the global optimal map matching effect.This paper applies deep learning and reinforcement learning to map matching,completes the training and testing of various models based on real data,and verifies the good performance of map matching algorithms based on deep learning and map matching algorithms based on deep reinforcement learning in indoor positioning.