Particle Filter-Based Indoor Localization with Learning Based PDR and Monocular Depth-Aided BIM Matching

Abstract

Modern smartphones offer several sensors that enable indoor pedestrian localization without the need for additional hardware. Pedestrian dead reckoning (PDR) provides a low-cost and efficient solution. However, it suffers from error accumulation and drift over time. Image-based localization methods can mitigate these limitations but are very computationally intensive to run at a higher frequencies. To address this, we propose a hybrid localization framework based on a particle filter, where a frequent, low-cost deep learning based inertial PDR is fused with a less frequent image-based updates to achieve accurate and robust localization. Our method does not require an offline mapping process and instead utilizes Building Information Models (BIM) generated maps. Furthermore, we incorporate recent monocular depth estimation models to generate depth directly from single images, thereby eliminating the need for continuous image streams to generate point clouds. Experimental results show that our proposed method can effectively track pedestrian poses using primarily smartphone sensors and BIM data.