NeuMAP: End-to-End Neural Map Matching-Based Pedestrian Dead Reckoning

Abstract

Pedestrian Dead Reckoning (PDR) using inertial sensors embedded in wearable devices and smartphones has gained attention for indoor localization, but still suffers from drift due to cumulative errors. Traditional map matching addresses this issue but often requires manually constructed geometric floorplan models or computationally expensive algorithms. To address these limitations, we propose NeuMAP, a neural map matching-based indoor PDR framework for end-to-end inference that takes a binary 2D floorplan image, along with raw IMU data and an initial position and heading, as input, and outputs a 2D pedestrian trajectory. NeuMAP represents the user's location as a belief over floorplan pixels and updates it using PDR motion estimates and learned map priors from the floorplan image. Trained on diverse simulated and real environments, NeuMAP achieves the lowest MAE and RMSE and a zero block rate among the compared methods, while reducing map-matching inference time by 89-92% compared with the two particle-filter-based map-matching baselines.