Research On Vehicle Positioning Algorithm Based On IMU

Intelligent vehicles have become one of the hot topics in Unmanned driving research field,but how to achieve the location information of intelligent vehicle is one of the key and difficult points.Now,vehicle positioning technology,especially Satellite-based navigation system could provide a high-precision position result.However,there are still a lot of challenge in vehicle positioning technology.Although with a significant development in the area of RTK-based,RFID-based and camera-based positioning technology,it is hard to achieve a sufficient accuracy and robust result,and the cost is high,so it can’t be applied to general intelligent vehicles.In the other hand,IMU-based positioning has been popular among researchers for its low cost.This thesis focuses on vehicle positioning technology based on IMU,and mainly performs the following work:On the basis of combining the kinematic constraints of the vehicle with the Kalman filter algorithm,the parameters of process noise and measurement noise in Kalman filter algorithm are set manually and fixed.In order to solve this problem,this paper presents a vehicle location algorithm based on noise parameter adaptive model.The main contribution is using the deep convolutional neural networks to design two different noise adaptive models for the positioning system.In order to reduce the influence of noise such as singular values in the IMU raw data and enhance the capture ability of the deep network model,we also use Z-score standardization method to process it,and the gradient clipping and weight attenuation techniques are used to effectively alleviate the problem of gradient explosion and overfitting in the deep network model.By calculating the relative translation error between the predicted trajectory and the real trajectory,the network model is optimized,and the parameters of process noise and measurement noise in the positioning system are adjusted dynamically to improve the positioning accuracy of the vehicle.The experiment proves that this method can obtain high positioning accuracy only by using a low-cost IMU inertial sensor,and the whole process does not involve other sensors,which greatly reduces the cost of vehicle positioning.In view of the different characteristics of two mainstream deep neural networks in processing tasks,In addition to using the features of CNN’s weight sharing to design two different adaptive models for the system’s process noise and measurement noise,we also take advantage of the ability of LSTM to capture the timing dependence between time series to design an adaptive model for the two noises of the system.At the same time,we also combine the advantages of deep network models to propose an improved CNN-LSTM network model to adaptively adjust noise parameters,we also use the Mish activation function to optimize the Relu activation function which has the problem of hard saturation.By comparing with CNN and LSTM network models,it is verified that this model can better adjust the noise parameters self-adaptively,so as to achieve higher positioning accuracy.This thesis has 35 maps,11 tables and 80 references.