Deep-Learning-Enhanced Outlier Detection for Precise GNSS Positioning With Smartphones

A robust outlier detection method is proposed for effective detection of smartphone global navigation satellite system (GNSS) measurement outliers frequently caused by multipath and non-line-of-sight (NLOS) effects. The new method has been developed through a novel combination of a binary-tree-based solution separation test (SS-test) and deep-learning techniques, which leverages the statistical testing’s outlier detection efficiency and the deep learning’s proficiency in modeling complex relationships. First, the binary-tree-based SS-test identifies specific subsets of outlier-free measurements. Second, the network predicts 3-D positioning errors through those subsets. Third, we obtain an accurate positioning solution from the measurement subset with the smallest predicted errors. To validate the proposed approach, a vehicle-based field test was conducted with a smartphone. The testing results indicate that the proposed method has reduced the rate of large positioning errors (LPERs) by 79% and improved the positioning accuracy by 41% compared to the conventional methods without robust statistical testing.