Source-linked AI summary
IDOL: Inertial Deep Orientation-Estimation and Localization
Scott Sun, Dennis Melamed, Kris Kitani
TL;DR
Inertial odometry can be undermined by inaccurate device orientation estimates, which are commonly used to rotate IMU measurements. The paper combines learned orientation estimation with an EKF and position estimation, and reports better orientation and localization performance than prior methods.
Problem
Data-driven inertial localization methods typically rely on conventional device orientation estimates, which can be inaccurate and corrupt network outputs.
Source: §1 · p. 1Method
A supervised two-stage pipeline estimates device orientation from IMU data with a recurrent network and EKF, then estimates position with another recurrent network.
Source: §1 · pp. 1–2, §3.3 · p. 5Results
The system localizes better than previous methods across multiple buildings and users; its orientation estimate is 0.28 radians (16.04°) more accurate than CoreMotion’s estimate.
Source: §5.3 · p. 7, §6 · p. 8Takeaways & Limitations
The orientation module is a swap-in component for existing systems, with orientation performance comparable to visual-inertial systems in known environments.
Source: §6 · p. 8Cross-building generalization is limited: magnetic-field differences can degrade orientation, and all data-driven methods studied regress position inaccurately on held-out buildings.
Source: §5.5 · p. 8
Abstract
from arXiv · showhide
Many smartphone applications use inertial measurement units (IMUs) to sense movement, but the use of these sensors for pedestrian localization can be challenging due to their noise characteristics. Recent data-driven inertial odometry approaches have demonstrated the increasing feasibility of inertial navigation. However, they still rely upon conventional smartphone orientation estimates that they assume to be accurate, while in fact these orientation estimates can be a significant source of error. To address the problem of inaccurate orientation estimates, we present a two-stage, data-driven pipeline using a commodity smartphone that first estimates device orientations and then estimates device position. The orientation module relies on a recurrent neural network and Extended Kalman Filter to obtain orientation estimates that are used to then rotate raw IMU measurements into the appropriate reference frame. The position module then passes those measurements through another recurrent network architecture to perform localization. Our proposed method outperforms state-of-the-art methods in both orientation and position error on a large dataset we constructed that contains 20 hours of pedestrian motion across 3 buildings and 15 subjects.
1 Introduction
Inertial localization depends on accurate orientation estimates to convert IMU measurements into a global frame, but conventional estimates can be inaccurate. The paper proposes a supervised two-stage approach that estimates orientation and position.
1 Introduction
- IMUs offer infrastructure-free, energy-efficient sensing, but accurate localization has often relied on WiFi, Bluetooth, LiDAR, or cameras with practical trade-offs.The passage describes beacon-based solutions as requiring heavy instrumentation, LiDAR as expensive and power-hungry, and image-based localization as power-intensive and a privacy concern.Source: §1 · p. 1
- Orientation errors can corrupt inertial localization: conventional estimates may be inaccurate and can contribute to substantial position error.Existing deep-learning approaches still typically use the phone’s conventional orientation estimate to rotate local IMU measurements before applying a network.Source: §1 · p. 1
- The proposed supervised deep network pairs an explicit orientation-estimation module with a position-estimation module.The model uses accelerometer and magnetometer measurements as well as gyroscope data, drawing on information about gravity and Earth’s magnetic field.Source: §1 · pp. 1–2
- The work contributes an orientation architecture, an end-to-end position estimator reported as more accurate than prior classical and learning-based techniques, and an annotated 6D-pose dataset.The model estimates 3D orientation and 2D position, with 3D position also possible.Source: §1 · p. 2
2 Related Work
The related work spans traditional and data-driven approaches to localization and orientation estimation. The paper identifies dependence on conventional orientation estimates as a gap in data-driven inertial localization and describes its learned orientation estimate as an alternative EKF update.
2.1 Traditional Localization Methods:
- Traditional inertial localization integrates gyroscope readings for orientation and double-integrates corrected acceleration for position, so errors grow; movement constraints and EKF sensor fusion can reduce drift.The passage gives foot-mounted ZUPT systems as an example of using movement constraints.Source: §2.1 · p. 2
2.2 Data-driven Localization Methods:
- RIDI corrects acceleration before double integration, while Cortes, Solin, and Kannala use CNN-regressed speed as an EKF pseudo-measurement.RIDI first classifies holding modality, then applies a modality-specific regressor to estimate the correction.Source: §2.2 · p. 2
- IONet and RoNIN regress motion with deep networks, and TLIO fuses learned displacement with raw IMU data; these approaches still depend on device orientation estimates that can be unstable or drift.IONet uses phone-API frame conversion, while TLIO uses a stochastic-cloning EKF to estimate position, orientation, and sensor biases.Source: §2.2 · p. 2
2.3 Traditional Orientation Estimation Methods:
- Madgwick’s filter assumes non-gravitational acceleration is small, which is impractical for pedestrian motion; MUSE uses acceleration updates only when static and relies mainly on magnetic north.MUSE’s magnetic basis is challenged by local magnetic-field variation, while a static device is rare during pedestrian motion.Source: §2.3 · pp. 2–3
- The paper uses an EKF that replaces gravity- or magnetic-vector updates with a learned model output, aiming for a less noisy orientation estimate and simpler Kalman updates.Source: §2.3 · p. 3
2.4 Data-driven Orientation Estimation Methods:
- Earlier learned orientation methods primarily filter gyroscope data or estimate corrections; this method directly estimates orientation from all IMU channels and fuses short-term gyro data through an EKF.The passage notes that earlier data-driven approaches had not included magnetic observations.Source: §2.4 · p. 3
3 Method
The method uses supervised learning to estimate device orientation and position in stages. It uses the orientation estimate to transform inertial measurements into the world frame before localization.
3 Method
- The supervised method uses accelerometer and magnetometer information to estimate absolute device orientation, then uses the transformed IMU measurements for position estimation.The design exploits the sensors’ measurements of gravity and Earth’s magnetic field.Source: §3 · p. 3, §3.1 · p. 3
3.1 Estimating 3D Orientation:
- An LSTM-based Orientation Network predicts orientation and covariance from IMU readings; an Extended Kalman Filter fuses those estimates with gyroscope updates to stabilize orientation.The filter uses quaternion operations to fuse angular data and rotates accelerometer and gyroscope readings into the world frame.Source: §3.1 · pp. 3–4
3.2 Orientation Module Training:
- Orientation training uses ellipsoid-fit magnetometer calibration and a negative log likelihood loss to train the covariance estimator.The covariance head outputs a six-dimensional parameterization of the covariance matrix.Source: §3.2 · p. 4
3.3 Estimating 2D Position:
- The position network maps world-frame gyroscope and accelerometer inputs to Cartesian displacement, which is accumulated into positions across windows.At test time, LSTM hidden states reset between sequence windows to help limit drift.Source: Figure 2 · p. 4, §3.3 · p. 5
3.4 Position Module Training:
- Position training progressively lengthens batch sequences from 100 to 2000 timesteps to reduce drift accumulation, with the option to shorten sequences afterward to reduce latency.Training uses mean squared error over each LSTM window’s displacement.Source: §3.4 · p. 5
4 Dataset
The dataset uses a phone mounted to a SLAM rig to record pedestrian motion and ground-truth pose in multiple buildings. Its collection setup yields varied walking and device orientations, though carrying the rig makes motion somewhat unnatural.
4 Dataset
- 20 hours of motion from 15 users were recorded across three buildings with an iPhone 8 sampling raw accelerometer, gyroscope, and magnetometer readings at 100 Hz.The Kaarta Stencil rig estimated pose at 200 Hz; testing measured less than 1.5° RMS orientation error and less than 10 cm RMS position error.Source: §4 · p. 5, Figure 3 · p. 5
- Participants walked at variable speeds, paused, changed direction arbitrarily, and could adjust their grip, producing variation in rig orientation relative to the user.Most trajectories were about 10 minutes long.Source: §4 · p. 5
- Carrying the mapping rig makes participants’ motions somewhat unnatural, although the authors argue key localization challenges remain, including no zero-velocity points or axis-constrained movement.The authors primarily rely on motion induced during walking because acquiring ground-truth phone orientation for modalities such as pocket carrying is harder.Source: §4 · p. 5
- Available datasets were unsuitable for this model for differing reasons, including missing raw IMU, phone orientations, magnetometer readings, or common-frame alignment.The authors also note that many such datasets have shorter trajectories, making drift less evident.Source: §4 · p. 5
5 Experiments
The experiments evaluate orientation and position accuracy against established baselines, then test generalization and the effect of replacing phone orientation estimates. The full orientation module and position pipeline outperform competing methods in the reported evaluations, while cross-building generalization remains challenging.
5 Experiments
- The evaluation targets better orientation estimates, improved position localization, and evidence that other odometry methods benefit from the proposed orientation module.Orientation is assessed by angular RMSE; position is assessed using ATE, one-minute T-RTE, and one-meter D-RTE.Source: §5 · p. 6
5.1 Training/Testing:
- At test time, the initial orientation may be supplied or estimated relative to a global frame with a calibrated magnetometer; inference takes approximately 65 ms per second of IMU data on a CPU.The orientation network is trained first, then initialized weights are used to train the position network.Source: §5.1 · p. 6
5.2 Baselines:
- Orientation comparisons include CoreMotion, Brossard et. al. (2020), and MUSE; position comparisons include PDR, IONet, RoNIN variants, and TLIO.The evaluation uses the correct RMSE calculation for RoNIN metrics, correcting an omission in the released evaluation code while retaining relative comparisons.Source: §5.2 · p. 6
5.3 Orientation Analysis:
- OrientNet+EKF significantly outperforms the orientation baselines across metrics, with 0.08 radians (4.6°) error in Building 1, nearly matching the ground-truth rig.The EKF eliminates high-error outliers; the full module's error grows less steeply over time, and its predicted uncertainty approximately matches observed error frequencies.Source: Table 1 · p. 6, §5.3 · p. 7, Figure 4 · p. 7
5.4 Position Analysis:
- A single model trained on Building 1 generalizes to known and unknown subjects and outperforms other methods on both; separately trained models also outperform other methods across buildings, especially on RTE.Example trajectories show rotational drift in RoNIN and TLIO estimates can compromise position enough to leave the floorplan.Source: §5.4 · pp. 7–8, Table 2 · p. 7, Table 3 · p. 7
- Replacing phone orientation estimates with the proposed module improves TLIO and RoNIN position performance and nearly reaches the theoretical maximum obtained using ground-truth orientations.The comparison uses Building 1.Source: §5.4 · p. 8, Table 4 · p. 8
5.5 Generalization
- Cross-building generalization has two failure modes: magnetic-field differences can degrade orientation, while changes in building shape and size affect position estimates across held-out locations.Training on a building with magnetic distortions can yield accurate orientation estimates there, but at the cost of degraded generalization to novel environments.Source: §5.5 · p. 8
6 Conclusion
The paper integrates deep orientation estimation with deep inertial odometry, reporting better localization across buildings and users and identifying cross-building generalization as future work.
6 Conclusion
- The combined system localizes better than previous methods across multiple buildings and users, while its orientation module outperforms traditional and data-driven techniques.The orientation module uses an LSTM and EKF; the paper describes it as a swap-in component for existing systems.Source: §6 · p. 8
- The paper integrates deep orientation estimation with deep inertial odometry, combining an orientation module with a position module.Source: §6 · p. 8
- The authors report that the orientation module can empower existing systems with orientation performance comparable to visual-inertial systems in known environments.Source: §6 · p. 8
- The paper contributes a dataset spanning 20 hours of pedestrian motion across 3 buildings and 15 people.Source: §6 · p. 8
- The authors identify generalization across buildings as a future-work issue, suggesting further data collection, data augmentation, or architectural modifications.Source: §6 · p. 8