Fusing Vehicle Trajectories and GNSS Measurements to Improve GNSS Positioning Correction Based on Actor-Critic Learning

Obtaining high-accuracy vehicle localization is essential for safe navigation in automated vehicles, however, complex urban environments result in inaccurate positioning because of multipath and non-line-of-sight errors, making precise positioning in complex environments an important and unsolved problem. Different from conventional model-based approaches which require rigid assumptions on sensors and noise models, data-driven artificial intelligent approaches can provide new solutions for Global Navigation Satellite System (GNSS) positioning problems in multipath environments. Recent learning-based works employ the generalization ability of the deep neural network to model these complex errors in urban environments, and provide effective positioning corrections for rough GNSS localizations by learning from a vehicle trajectory-based Reinforcement Learning (RL) environment. However, existing RL-based approaches employ insufficient features as input, and cannot take advantage of the vehicle trajectory information. To tackle these issues, we propose a novel Deep Reinforcement Learning-based (DRL) positioning correction framework with fused features considering the vehicle trajectory information and GNSS features in the multi-input observations. To help the agent learn the correction policy efficiently with comprehensive inputs, we develop a new positioning correction RL environment with multi-input observations, which augments GNSS device readings of different satellites as complementary to the vehicle trajectory, and moreover employ a correction advantage reward to help the agent learn an efficient policy in positioning correction. To fully utilize comprehensive multi-input observations in the proposed environment, we employ the LSTM module to separately extract features to fuse brief states from input time series data. Finally, to address the long-term positioning accuracy, we construct the learning model based on actor-critic DRL structure with the cumulative reward setting and the continuous-action actor by using joint features from vehicle trajectory and GNSS measurements, for learning the optimal positioning correction strategy. We test our proposed approach on the real-world dataset, i.e., Google Smartphone Decimeter Challenge (GSDC), and results show that our algorithm can obtain improved localization performances over both model-based methods with 23% improvement from Kalman filter, and 15% from existing learning-based DRL method for positioning correction.