**5. Conclusions**

In this study, we propose a new smartphone-based indoor localization system using deep learning. The system used the following ideas to increase the practicality for pedestrians: 1. The movement speed calculated using outdoor GPS coordinates was set as the label of the training dataset. 2. A consistent training dataset regardless of the smartphone placement was constructed by converting accelerometer data to earth-oriented coordinates the 3-axis is fixed. 3. An unsupervised learning model was implemented to identify the minimum size of the training dataset in real time to minimize the battery consumption. In addition, the following concepts were proposed to increase the accuracy of estimating the moving speed: 1. Noise was removed to increase the accuracy of GPS and accelerometer data. 2. Data augmentation methods were applied to obtain a uniform distribution of moving speeds of pedestrians. 3. By implementing seven different deep learning models, the optimal deep learning model with the best moving-speed estimation performance was selected. After learning the parameters for the indoor localization system with the data collected outdoors, the proposed system showed a localization estimation error of approximately 3 to 5 m as a result of direct experimentation inside a horseshoe-shaped building compared to the test data of approximately 240 m in length. Compared to existing indoor location recognition studies, the proposed system shows a high location estimation accuracy, and the practicality for pedestrians is also very high. In the future, we plan to develop an in-building floor-recognition technology to provide 3D location information for indoor pedestrians. The three-dimensional location information also allows pedestrians to

know which floor of the building they are on, so that they can receive much more diverse location-based services and identify safer evacuation routes in the event of a disaster.

**Author Contributions:** Conceptualization, S.K.; formal analysis, S.K.; funding acquisition, S.K.; methodology, J.Y. and S.K.; project administration, S.K.; software, J.Y.; supervision, S.K.; validation, J.Y.; visualization, J.Y.; writing—original draft, J.Y.; writing—review & editing, S.K. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was supported by the National Research Foundation of Korea (NRF) gran<sup>t</sup> funded by the Korea governmen<sup>t</sup> (MSIT) (No. 2022R1A5A8026986) and Korea Institute for Advancement of Technology (KIAT) gran<sup>t</sup> funded by the Korea Government (MOTIE) (P0020536, HRD Program for Industrial Innovation).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.
