[Paper Review] Deep Learning based Pedestrian Inertial Navigation: Methods, Dataset and On-Device Inference

Changhao Chen, Peijun Zhao+4 more|arXiv (Cornell University)|Jan 13, 2020

Indoor and Outdoor Localization Technologies34 references4 citations

TL;DR

This paper proposes L-IONet, a lightweight deep learning framework for real-time pedestrian inertial navigation using raw IMU data, and introduces the Oxford Inertial Odometry Dataset (OxIOD), a large-scale, high-precision dataset with motion capture ground truth. The method achieves accurate on-device inference on smartphones and smartwatches, outperforming traditional PDR and prior DNN models in diverse motion scenarios including handheld, pocketed, and trolley-borne phone configurations.

ABSTRACT

Modern inertial measurements units (IMUs) are small, cheap, energy efficient, and widely employed in smart devices and mobile robots. Exploiting inertial data for accurate and reliable pedestrian navigation supports is a key component for emerging Internet-of-Things applications and services. Recently, there has been a growing interest in applying deep neural networks (DNNs) to motion sensing and location estimation. However, the lack of sufficient labelled data for training and evaluating architecture benchmarks has limited the adoption of DNNs in IMU-based tasks. In this paper, we present and release the Oxford Inertial Odometry Dataset (OxIOD), a first-of-its-kind public dataset for deep learning based inertial navigation research, with fine-grained ground-truth on all sequences. Furthermore, to enable more efficient inference at the edge, we propose a novel lightweight framework to learn and reconstruct pedestrian trajectories from raw IMU data. Extensive experiments show the effectiveness of our dataset and methods in achieving accurate data-driven pedestrian inertial navigation on resource-constrained devices.

Motivation & Objective

• Address the lack of large-scale, high-precision, publicly available datasets for training and evaluating deep learning models in inertial navigation.
• Develop a lightweight deep neural network framework capable of real-time, on-device inference for pedestrian inertial odometry.
• Enable accurate and robust trajectory estimation from raw IMU data across diverse user motion patterns and device placements.
• Demonstrate generalization of trained models to unseen environments and motion types beyond the training data.

Proposed method

• Propose L-IONet, a lightweight deep neural network using dilated causal convolutions (WaveNet-style) to model long-range temporal dependencies in raw IMU sequences.
• Design a two-branch architecture: one branch processes raw IMU data (accelerometer and gyroscope) to predict 3D position, and another learns velocity predictions to constrain error drift.
• Integrate a residual connection and normalization layers to improve training stability and model performance on resource-constrained devices.
• Train the model end-to-end using supervised learning with motion capture ground truth from the OxIOD dataset.
• Optimize the model for edge deployment by reducing parameter count and inference latency, enabling real-time inference on smartphones and smartwatches.
• Utilize knowledge distillation as a future extension to further compress the model for even lower-resource devices.

Experimental results

Research questions

• RQ1 Can a lightweight deep learning model achieve accurate pedestrian inertial odometry from raw IMU data on resource-constrained edge devices?
• RQ2 How does the proposed L-IONet framework compare in performance and efficiency to existing DNN-based and model-based (e.g., PDR) inertial odometry methods?
• RQ3 To what extent can a model trained on a single indoor environment generalize to new environments and motion types not seen during training?
• RQ4 How effective is the use of dilated causal convolutions (WaveNet) in modeling complex human motion dynamics from raw IMU sequences?
• RQ5 Can a large-scale, publicly available dataset with fine-grained ground truth significantly accelerate research in deep inertial navigation?

Key findings

• The proposed L-IONet model achieves superior trajectory estimation accuracy compared to the baseline PDR algorithm, especially in non-periodic motion such as trolley movement where PDR fails.
• L-IONet outperforms the IONet baseline in handheld and handbag scenarios due to its enhanced ability to model complex motion dynamics via WaveNet-based temporal modeling.
• The model achieves real-time inference on low-end smartphones and smartwatches, demonstrating feasibility for on-device deployment.
• Despite being trained exclusively on data from a Vicon motion capture room, L-IONet generalizes effectively to large-scale indoor navigation across two floors of an office building.
• The OxIOD dataset, with 158 sequences totaling 42.587 km and 0.5 mm ground-truth accuracy, provides a robust benchmark for future research in inertial navigation.
• The WaveNet-based architecture in L-IONet is more effective than the LSTM-based IONet in capturing free-form motion, particularly when the device is not fixed in a stable position.