Transfer Learning Approach for Indoor Localization with Small Datasets

Abstract

Indoor pedestrian localization has been the subject of a great deal of recent research. Various studies have employed pedestrian dead reckoning, which determines pedestrian positions by transforming data collected through sensors into pedestrian gait information. Although several studies have recently applied deep learning to moving object distance estimations using naturally collected everyday life data, this data collection approach requires a long time, resulting in a lack of data for specific labels or a significant data imbalance problem for specific labels. In this study, to compensate for the problems of the existing PDR, a method based on transfer learning and data augmentation is proposed for estimating moving object distances for pedestrians. Consistent high-performance moving object distance estimation is achieved using only a small training dataset, and the problem of the concentration of training data only on labels within a certain range is solved using window warping and scaling methods. The training dataset consists of the three-axes values of the accelerometer sensor and the pedestrian’s movement speed calculated based on GPS coordinates. All data and GPS coordinates are collected through the smartphone. A performance evaluation of the proposed moving pedestrian distance estimation system shows a high distance error performance of 3.59 m with only approximately 17% training data compared to other moving object distance estimation techniques.