Study On Zero-Velocity Detection And Location Of Foot-Mounted IMU Based On Deep Learning

Due to the development of the Internet,the popularity of mobile devices and various personal devices,location-based services have become increasingly important.At present,many satellite-based outdoor positioning technologies have achieved high positioning accuracy.However,indoor positioning technology still faces many challenges.Especially in emergencies or temporary application scenarios such as emergency rescue of firefighters and indoor drills of soldiers,it is very difficult to navigate and locate the human body due to weak GPS or Beidou signals.The typical solution is to use inertial navigation technology based on inertial sensors.This technology does not require pre-deployment of any basic equipment and does not rely on satellite signals.It is a very efficient indoor positioning solution,but accurate inertial navigation depends on accurate zero-velocity detection.Therefore,research on zerovelocity detection and positioning methods is of great significance to improve the accuracy of inertial navigation.This article focuses on the zero-velocity detection method based on IMU,and the main work done is as follows:Analyze the advantages and disadvantages of the existing zero-velocity detection methods,determine the introduction of deep learning methods and select IMU sensors to collect data.By analyzing the movement data of pedestrians in various gaits,it is determined that the IMU is worn in the middle of the shoelace of the right foot,and a data set including walking,running,mixed gait,and up and down movement is collected.For the collected data set,a method of data preprocessing is given,and the data is denoised,which is conducive to more accurate subsequent experimental results.Analyze and compare a variety of Kalman filter algorithms,determine two EKF algorithms and calculate multiple parameters to update the foot IMU at zero velocity.Aiming at the problem that the zero-velocity detection method based on a fixed threshold has poor adaptability in a complex gait environment,an improved zerovelocity detection algorithm based on adaptive threshold is studied,which improves the adaptability of multi-gait.The main content of the research is to introduce an improved LSTM network model to recognize and classify pedestrian gait.By improving the accuracy of gait recognition to adaptively adjust the zero-velocity judgment threshold,the positioning error of the foot IMU is reduced by 67.9% compared with the fixed threshold method.Aiming at the difficulty of making zero-velocity labels and the insufficient generalization ability of the model,a method of training the zero-velocity detection model by using an improved neural network is studied.The main content of the research is to optimize two neural network models,introduce the idea of eliminating pseudozero-velocity points to process the data set,compare three zero-velocity detection algorithms,and verify the superiority of the zero-velocity detection model based on deep learning Among them,the zero-velocity detection accuracy rate based on the improved RNN reaches 95.7%,and the zero-velocity detection algorithm based on the improved LSTM performs best,with a detection accuracy rate of 97.9%,and the positioning error of the foot IMU is significantly reduced.This thesis has 35 maps,8 tables and 80 references.