Acoustic-assisted Indoor Pedestrian Dead Reckoning
Yang Wang, Heng Zhou, Takuya Maekawa · 2025
Pedestrian Dead Reckoning (PDR) using smartphone IMUs enables infrastructure-free indoor localization but suffers from accumulated drift due to sensor noise and unknown initial states. While prior work has explored sparse cues such as magnetic anomalies or BLE beacons to correct the drift, there remains a need for dense and low-cost landmarks. We propose a novel acoustic-assisted PDR method that detects ubiquitous doors and walls via active sensing with standard smartphone speakers and microphones. Our system generates Range-Doppler maps from audio echoes and applies a lightweight Long-term Recurrent Convolutional Network (LRCN) to classify acoustic events and estimate wall distance. These cues are fused with inertial priors in a segmented particle-filter framework: door events serve as absolute anchors, and detected walls are used to fine-tune trajectory estimates. Evaluated across nine diverse indoor environments and four smartphone models without environment-specific training, our method reduces the mean absolute positioning error by 86.83% compared to state-of-the-art neural PDR and achieves sub-meter mean absolute error over more than 10,000 meters of walking.