Deep Learning-Based Multi-Floor Indoor Localization Using Smartphone IMU Sensors With 3D Location Initialization

Pedestrian Dead Reckoning (PDR) is a widely used indoor localization technique that does not require infrastructure and can exploit smartphone inertial sensors. However, it suffers from accumulated drift, and accurate initial position and floor information are especially important in multi-floor environments. This study proposes a deep learning-based multi-floor indoor localization method that estimates a three-dimensional position using only smartphone sensors without relying on external infrastructure. The proposed method utilizes the smartphone magnetometer to perform accurate 3D position initialization and drift correction, thereby addressing major limitations of conventional PDR methods such as sensitivity to initialization errors and performance degradation caused by drift. In this paper, we explain conventional IMU sensor-based PDR methods and describe the problems of PDR as well as the limitations of barometer-based floor transition detection techniques. The proposed initial 3D position estimation method detects the initial floor using a deep learning model trained on magnetic field sequences from the user's first 10 steps, and estimates the 2D position by comparing the current magnetometer data with a pre-built magnetic field map using the k-nearest neighbors algorithm. After detecting the initial 3D position, the method combines barometer-based floor transition detection and PDR-based position estimation to enable multi-floor indoor localization, and addresses the drift problem using correction nodes based on magnetic field data. As a result of experiments conducted in multi-floor buildings, the proposed method demonstrated significantly superior performance in terms of path similarity and localization accuracy compared to conventional methods based on PDR, complementary filter, and Kalman filter. In particular, it achieved a localization error of 64 cm and up to 99.8% accuracy within a 2-meter threshold, proving the effectiveness and practicality of the proposed approach.