GPS and Lidar Fusion Positioning Based on Unscented Kalman Filtering and Long Short-Term Memory Cosidering GPS Failure

Vehicle positioning accuracy has always been one of the important factors restricting the development of intelligent driving technology. After analyzing the positioning principles of GPS and Lidar, this paper proposes a GPS and Lidar fusion positioning method based on the Unscented Kalman Filter (UKF). Moreover, the Long Short-Term Memory(LSTM) neural network is used to solve the GPS failure in this paper. First, the vehicle motion is analyzed, and the vehicle kinematics model is established. Meanwhile, an UKF is designed. Then the overall scheme of this paper is designed according to the positioning characteristics of GPS and Lidar. When GPS is normal, the LSTM neural network algorithm is in training mode. Neural network model is trained through the data of IMU and GPS. Once the GPS is lost, LSTM enters the prediction mode. The IMU data is used as the input of the neural network system, and the GPS positioning is predicted to obtain a pseudo GPS positioning value based on the trained neural network model, which is fusion positioning with Lidar by UKF. Then the LSTM neural network algorithm and the lidar positioning algorithm based on Normal Distributions Transform(NDT) are designed. Finally, the simulation and test are carried out. The test results show that the positioning accuracy of this method is about 0.15 m and the time is less than 50 ms, which indicates the method of this paper has a better fusion effect.