GPS-blocked Vehicle Tacking Algorithm Based On Mobile Crowd-sensing

Recently,GNSS(Global Navigation Satellite System)have been widely developed to provide accurate and convenient outdoor positioning services for vehicles.However,there are still a series of areas in the city that cannot be covered by satellites,such as large tunnels,underground parking lots and multi-level overpasses.The number of direct lineof-sight satellites are extremely reduced,thus causing delay of signal phase,reduction of load-noise ratio,multi-path propagation and other serious interference,which seriously interfere with the vehicle positioning results.To solve this kind of missing positioning problem,recent research usually leverages the vehicle inertia tracking method based on smartphone sensor,which basic principle is Dead Reckoning.However,the performance of this method is seriously affected by the internal fluctuation of the mobile phone sensor and the external disturbance of the car body.This study makes a large-scale evaluation of the car trajectory of Didi ride-hailing platform and summarizes the relevant technical challenge.Then,this paper proposes an inertia learning framework to solve the above problems.The main research work of this paper is as follows:Build a vehicle tracking prototype system based on smart phone inertial sensors,this paper point out three important technical challenges it faces: 1)various ways of placing smartphones inside the vehicle body and unknown posture during driving;2)unstable noise on low-quality inertial sensors in commercial smartphones;3)diversity of perception of smartphones and vehicles through crowds.Aiming at the above three threats,this paper proposes an inertia learning framework based on smart phones to infer the location of vehicles in real time.It includes a phone posture estimation algorithm based on Principal Component Analysis,which can generate the motion of the vehicle from the inertia reading of the smartphone.Secondly,this paper also designs a vehicle tracking model based on Temporal Convolutional Networks to replace the traditional double integration strategy to infer vehicle trajectories.Finally,this paper proposes a personalized model training and migration method for mobile devices to improve the tracking accuracy of individual devices.In this paper,the vehicle tracking model is trained and tested through the large-scale crowd-sensing data set of Di Di platform.Compared with the traditional method,the three-axis angle error of this method is less than 5 degrees,and the GPS-free positioning accuracy for 30 seconds is reduced from 74.77 m to 23.64 m.This method has been deployed on 7.59 million devices of Didi ride-hailing platform,and location reasoning is carried out 4.26 billion times a day.