Learning-based Localization of AUV with Outlier Sensor Data

This paper describes localization of autonomous underwater vehicles(AUV) in case of some navigation sensor data are an outlier. In that situation, using existing navigation algorithms causes problems in long-range localization including integration. All the integration-based existing navigation algorithms can estimate exact position, only if the integration is performed under the assumption that the position of previous step is correct. Therefore, even if an outlier sensor data occurs in a short period of time, problems in localization will continue. Also, outlier sensor data related to heading (direction of AUV) causes bigger problems. In this work, we propose a localization method through learning that is used in a situation with outlier sensor data. To do so, a learning model is designed by fully connected model and is trained through partly contaminated real sea data. For Training, displacement between subsequent GPS data is taken as reference. As a result, average Euclidean error in incremental displacements of the existing navigation algorithm result and the reference is 0.45m. On the other hand, the proposed learning-based localization shows the average Euclidean error of 0.24m. This result doesn't mean that the localization method through learning is accurate than the existing navigation algorithm. However, if there are outlier sensor data related to heading, we can conclude that using our proposed method shows more accurate localization.