GNSS-based Environment Retrieval: A Deep Learning Approach for Fine-grained Environment Perception

GNSS positioning accuracy for vehicles confronts different challenges from a variety of environments in urban navigation. To provide a suitable positioning solution for each environment, perceiving and categorizing the environment is necessary. First, this paper proposed a fine-grained environment categorization to symbolize the environment by considering building structures and signal propagation characteristics. Then, a GNSS-based environment perception method is introduced in this work. Supervised by satellite visibility labels, a deep learning model is trained to extract environmental features from GNSS measurements. A database is constructed to associate these features and environment categories. Finally, the user environment category can be retrieved by matching its extracted features with the database. The effectiveness of our method is validated on vehicle-collected GNSS data in Hong Kong with different environments. Through experiment results, the proposed method achieves 7% higher in F1-score and 5% higher in Top-5 accuracy compared to the conventional environment classifier.