| As the global navigation satellite system(GNSS)technology matures,GNSS has become the best choice for outdoor positioning and is widely used in surveying and mapping,aviation,navigation,transportation,meteorology and many other environments.However,in the indoor environment,there is a problem that GNSS is unavailable due to signal occlusion.Therefore,how to effectively locate in the indoor environment is an urgent problem to be solved.Unlike previous foot-mounted micro-inertial navigation devices or handheld smart phones for positioning and navigation,this topic focuses on the waist position that is more convenient to wear but less studied,and uses a variety of deep learning models such as Long Short-Term Memory(LSTM)and Graph Convolutional Network(GCN)perform sensor feature data learning to achieve a more stable and accurate effect.The main contributions and work of this article are:(1)Reproduction of multiple sensor positioning.Due to the complex and interactive indoor positioning technology system,there is no complete frame application yet.Therefore,in order to study the positioning technology better,this topic reproduces geomagnetic positioning,pneumatic floor positioning,RSSI ranging positioning and inertial navigation positioning to learn the advantages and disadvantages of each sensor technology in indoor positioning.(2)Three deep learning methods are used to process the waist-mounted data to improve the accuracy of the heading angle estimation of the waist-mounted equipment.On the one hand,considering that the course estimation can be regarded as a time series prediction problem,thus introducing LSTM model;on the other hand,considering that the graph can represent the relationship between the data,it can well describe the direction and movement direction at different times,GCN model is introduced;Finally,the heading angle is calculated by predicting the pedestrian movement direction and the auxiliary direction sensor,so that even if the pedestrian movement direction and the face orientation are inconsistent,the true movement heading angle can be calculated.(3)When drawing waist-mounted user trajectories,we designed a two-stage behavior detection algorithm to improve the accuracy of step length and step counting estimation.We decompose the traditional gait detection mode into recognition in multi-motion state,design a double-layer discriminator,the first layer judges the movement state such as up and down stairs,escalators,walking,etc.,and the second layer judges the in-situ jump that occurs when walking,Body shaking and other actions to reduce the number of pseudo steps.We train the number of steps by inputting sensor data of equal interval duration,and estimate the step length within the duration,thereby improving the accuracy of step length and step counting estimation on the basis of reducing interference.(4)Designed the experimental scenarios of waist-mounted inertial navigation,including data fusion of waist-mounted inertial navigation equipment,UWB and mobile phone sensor equipment.Simulation experiments show that the deep learning model has achieved satisfactory performance on the server side,and due to the special requirements of indoor positioning on accuracy,the accuracy of deep learning on the mobile side needs to be further improved.At the same time,in order to reduce the mobile data upload server,we try to use transfer learning technology to fine-tune the model based on the pre-trained model on the server side with the data collected through the mobile terminal. |
PDF Full Text Request