Positioning Performance Improvement of GNSS Using Deep Learning for Autonomous Vehicles

Abstract

As autonomous vehicles are actively developed as a next-generation means of transportation, various core technologies required for autonomous driving are being developed rapidly. Among them, vehicle positioning technology is essential for convenient and safe driving. Currently, the most representative vehicle positioning technology is Global Navigation Satellite System (GNSS). For more accurate positioning, research related to sensor fusion that estimates the position by fusion of Inertial Measurement Unit (IMU) data with GNSS data is being actively conducted. However, sensor fusion has disadvantages in that complexity increases according to the number of sensors and dynamic analysis is difficult. In this thesis, a deep learning model using Recurrent neural network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), which can learn time series data, was implemented in order to estimate the accuracy of position of the vehicle with these sensor data, and the performance was compared and analyzed. In addition, the performance of prediction accuracy of position, noise filtering, and sensor fusion was compared and analyzed. In the model design stage, position information of three locations was obtained using Real Time Kinematic (RTK), and velocity and acceleration data were collected using Kalman filter. The collected data was pre-processed to train and test the deep learning model. In the data preprocessing process, data was normalized between 0 and 1, separated into data in the training set and test set, and four data sets to be used for comparative analysis were constructed. In the model training stage, hyperparameters were optimized for RNN, LSTM, LSTM dropout, GRU, and GRU dropout. An optimizer for back propagation learning is Adaptive Moment Estimation (Adam), and a learning process is stopped before overfitting using the Early stopping function of the Keras library. Batch size has a great influence on learning performance, so the learning results were compared by changing them to 8, 16, 32, 64, and 128. Finally, the activation function used tanh. The trained models created a total of 300 models by combining 3 locations, 5 models, 4 training sets, and 5 batch sizes. In the model testing stage, the performance of position prediction accuracy, noise filtering, and sensor fusion for each model and training data set were confirmed and compared. As a result, overall performance of prediction accuracy of position and noise filtering of LSTM and GRU models were better. In addition, as for the result of sensor fusion according to the training data set, the performance of the model trained with the training data set consisting of position, velocity, and acceleration data was the best. The data set consisting of position data added noise and normal velocity and acceleration data also showed high performance of prediction accuracy of position and noise filtering. Through this, it was confirmed that if the position, velocity, and acceleration sensors were fused using the RNN, LSTM, and GRU models, high performance positioning results could be obtained through the velocity and acceleration sensors despite the error of the position sensor.