Research On Vehicle Position Tracking Based On Temporal Convolutional Networks

This project comes from a school-enterprise cooperation project between the supervisor and the team of Didi Chuxing company.It is used to solve the problem that mobile phones cannot locate vehicles in GNSS(Global Navigation Satellite System)blocked environments.Since the birth of GNSS,such as the Global Positioning System(GPS)of the United States and the Bei Dou Navigation Satellite System(BSD)of China,it has been the main choice for vehicle positioning.However,in scenes such as urban canyons and tunnels,satellite signals are reflected causing huge pseudo-range errors or unreceivable signals which resulting in positioning drift or wrong location.In order to solve the problems of GNSS positioning failure,researches mainly use inertial navigation whose fundamental theory is dead reckoning(DR).Inertial navigation technology uses an Inertial Measurement Unit(IMU)which usually includes a three-axis accelerometer and a three-axis gyroscope to collect inertial data,which is used to calculate posture and velocity.And it combines posture and velocity to calculate relative position changes.Although the traditional algorithm is theoretically strong,it has many shortcomings.First of all,it requires the mobile phone and the vehicle to be relatively fixedly connected and around three minutes of inertial data input in advance to ensure that the calculated posture is accurate enough.Secondly,traditional algorithms have large positioning errors and serious accumulation of errors.Thirdly,as the largest travel platform in China,Didi Chuxing company can produce massive amounts of travel data every day,but traditional algorithms cannot take advantage of big data.This paper proposes a vehicle position tracking model based on temporal convolutional networks(TCN).The model takes the features value of the mobile phone IMU and the initial speed and bearing of the vehicle as input,and takes the speed and bearing changes as the output.The contribution of this paper is threefold.Firstly,a deep learning tracking model based on mobile phone inertial data is proposed to replace the traditional algorithm and delay the accumulation of inertial errors.The second is to design a loss function based on Smooth L1 Loss,which further improve the positioning accuracy from the perspectives of multi-task learning and time series prediction.The third is to verify the feasibility of training a personalized model with a single user’s short-term data offline,and use the inertial data collected in tunnels to transfer the model to fit the tunnel scene.This paper trains and tests the model through crowd-testing data,uses Mean Absolute Error(MAE)and error quantiles to evaluate the effect of the model.And the paper uses real stagnant data to complete the model’s business index evaluation offline.The cumulative sampling time of the whole dataset exceeds 2,200 hours,and the cumulative mileage exceeds 1.25 million kilometers,covering 15 major cities in China.By the end of 2020,the number of models deployed on the Didi Chuxing platform exceeded 7.15 million units,and the number of model inferences in a single day exceeded 36 billion.