Efficient Data Collection and Training for Deep-Learning-Based Indoor Vehicle Navigation

While data quality and quantity directly affect the estimation accuracy of deep learning-based localization techniques, obtaining empirical position data is often costly due to time-consuming experiments, especially for large-scale indoor environments. This paper proposes a novel data collection method to build deep learning-based localization models for indoor vehicle navigation services using multi-directional beacons. By defining the directed graph of the model applicability, we show that high-resolution position labeling data of a limited number of beacons is sufficient to train the localization model for the whole indoor scenario. However, obtaining the ideal directed graph of the model applicability still requires costly high-resolution empirical data from all beacons. We thus propose an autoencoder-based predicted graph method using only cost-effective low-resolution proximity data. We derive the minimum set of beacons needed to collect high-resolution empirical data to train the localization model. We evaluate the proposed beacon selection method in terms of the estimation accuracy and the data collection cost to the beacon-specific model and the ideal directed graph-based method through the extensive set of experimental measurements. The evaluation results show that the proposed localization model using the autoencoder-based predicted graph provides good estimation accuracy while considerably reducing the required training beacons.