Overbounding Multipath Error in Urban Canyon With LSTM Using Multi-Sensor Features

Multipath error is a major challenge for positioning integrity monitoring in autonomous driving, which requires conservative and effective overbounding methods. Traditional methods based on Gaussian overbound need the error distribution information, which is hard to obtain when using multiple features related to multipath as prior knowledge. A recent study used quantile overbound based on multi-layer perceptron (MLP) network with promising results but ignored the temporal correlation of multipath and the dynamic surrounding information. This paper proposes a multipath overbounding method based on a designed long short-term memory (LSTM) network using LiDAR, cameras, and global navigation satellite systems data. The method aims to improve the model's effectiveness while ensuring the model's conservatism. The effectiveness of the model is assessed using a generalized coefficient of determination, which shows how close the predicted quantile is to the actual value. Results show that the LSTM model outperforms the previous MLP-based study in predicting the quantile of multipath error in Hong Kong urban data. By using multi-sensor data as input, the effectiveness of LSTM improves by over 10% when using time windows of 15 seconds for different urban scenarios. The longer the time window, the better the performance of effectiveness. The predicted quantile is then used to compute an overbounded standard deviation based on a zero-mean Gaussian distribution, whose conservatism is verified by the cumulative distribution function. Overall, this study indicates that the use of multi-sensor data and a longer time window can better facilitate the effective bounding of the multipath error while ensuring conservatism.