PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning based Pedestrian Inertial Navigation: Methods, Dataset and On-Device Inference

Changhao Chen, Peijun Zhao, Chris Xiaoxuan Lu, Wei Wang, Andrew Markham, Niki Trigoni

arXiv:2001.04061v1cs.ROcs.LG

TL;DR

Low-cost IMU navigation faces drift and conventional methods with restrictive motion or sensor-placement assumptions. This paper releases OxIOD and proposes L-IONet, then reports competitive performance and efficient inference, including real-time operation on phones and a smartwatch.

  • Problem

    Low-cost IMU errors cause unbounded drift, while ZUPT and PDR rely on restrictive sensor-placement or pedestrian-motion assumptions.

    Source: §I · p. 2

  • Method

    The paper releases the OxIOD pedestrian dataset and proposes L-IONet, which uses causal dilated convolutions to learn inertial tracking from raw IMU data.

    Source: §I · p. 2, §IV · p. 4, C. Lightweight Inertial Odometry… · p. 6

  • Results

    L-IONet shows competitive performance with greater memory, inference, and training efficiency, and achieves real-time inference on smartphones and a smartwatch.

    Source: B. Model Performance at the edge · p. 8, §VI · p. 10

  • Takeaways & Limitations

    OxIOD is intended to support training and evaluation of inertial navigation models and future research in long-term ubiquitous ego-motion estimation.

    Source: §I · p. 2, §VI · p. 10

Abstract

from arXiv · show

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.

I. INTRODUCTION

IMUs offer a compact, infrastructure-light sensing option for navigation, but long-term drift, restrictive conventional methods, and gaps in labeled data and benchmarking constrain pedestrian inertial navigation. The paper introduces OxIOD and L-IONet to support data-driven navigation and evaluate it alongside conventional methods.

I. INTRODUCTION

  • IMUs are pervasive, low-cost sensors that can support location services with little infrastructure and process 6-dimensional time series in real time on resource-constrained devices.The paper also notes that off-cloud processing better protects user location and motion privacy.Source: §I · p. 1
  • Low-cost IMUs accumulate unbounded system error, while ZUPT requires a foot-fixed sensor and PDRs assume periodic pedestrian motion.These restrictions make the described conventional approaches difficult to apply in everyday usage.Source: §I · pp. 1–2
  • Data-driven navigation needs extensive, precisely labeled sensor data, but existing datasets poorly reflect everyday human tracking and common benchmarks are lacking.The paper also identifies limited study of DNN efficiency on low-end devices as a challenge.Source: §I · p. 2
  • OxIOD provides 158 pedestrian sequences totaling more than 42 km, with high-precision labels and multiple sensor attachments.The dataset is presented as a common benchmark for pedestrian inertial navigation.Source: §I · p. 2
  • L-IONet learns inertial navigation from raw IMU data without handcrafted engineering and is designed for more efficient training and inference than LSTM-based models.The authors evaluate classical PDRs and data-driven models on OxIOD as a common benchmark.Source: §I · p. 2

II. RELATED WORK

Existing inertial datasets differ in their suitability for pedestrian odometry: some target vehicles or lack trajectory ground truth, while human activity and gait datasets often lack substantial movement or reliable localization labels. OxIOD is described as larger and more diverse than most inertial navigation datasets.

A. Inertial Navigation Datasets

  • Vehicle datasets use chassis-mounted sensors, while TUM VI provides ground truth only at sequence endpoints and ADVIO is short with pseudo ground truth.The described vehicle datasets are suitable for vehicle movement studies, not directly for human movement.Source: A. Inertial Navigation Datasets · pp. 2–3
  • Human activity datasets often lack enough participant movement for odometry, while gait datasets such as MAREA lack solid ground truth.These limitations constrain their use for training and testing odometry models.Source: A. Inertial Navigation Datasets · p. 3
  • OxIOD contains 158 sequences spanning 42.587 km, exceeds most inertial navigation datasets in size, and offers greater diversity for everyday human motion.The paper argues that its scale and accurate labels suit data-hungry deep neural network methods.Source: A. Inertial Navigation Datasets · p. 3

B. Inertial Navigation Using Low-cost IMUs

Low-cost IMU navigation has moved from constrained, hand-engineered approaches toward data-driven motion models, but edge inference remains a concern. OxIOD provides varied pedestrian sequences and motion ground truth for training and evaluation.

B. Inertial Navigation Using Low-cost IMUs

  • Prior inertial methods use foot-mounted ZUPT or step-based PDR, while data-driven models learn more general motion from raw inertial data; few prior works consider low-end-device inference efficiency.PDR assumes periodic pedestrian motion; ZUPT requires the sensor to be firmly fixed to the foot.Source: B. Inertial Navigation Using Low-cost… · p. 3

III. OXFORD INERTIAL ODOMETRY DATASET

  • OxIOD contains 158 sequences spanning 42.5 km and 14.72 hours, with varied attachments, motion modes, four consumer-phone types, and five users.The collection includes handheld, pocket, handbag, and trolley/stroller attachments, and halting, slow walking, normal walking, and running.Source: §III · p. 3

A. Sensor Setup

  • The dataset records accelerations and angular rates from phone 6-axis IMUs and magnetic fields from 3-axis magnetometers, with Vicon motion capture providing orientation, velocity, and position ground truth.For large-scale collection across two office floors, Google Tango was used as pseudo ground truth.Source: A. Sensor Setup · pp. 3–4
  • The Vicon setup used 10 cameras at 250 fps and captured motion with precision down to 0.5 mm.Phone and magnetometer timestamps were shared; Vicon and phone clocks were synchronized to UTC before experiments.Source: A. Sensor Setup · p. 4

B. Data Collection

The collection varies phone placement, movement, devices, and users to reflect everyday pedestrian sensing. It also includes large-scale office trajectories with pseudo ground truth.

B. Data Collection

  • Four common attachments were studied: phone in hand, pocket, handbag, or on a trolley.The attachment experiments used one participant walking naturally inside the Vicon room.Source: B. Data Collection · p. 4
  • The collected motion modes were halting, slow walking, normal walking, and running, with approximate speeds of 0.5, 1, and 1.5 m/s for the three walking modes.Experiments varied motion mode with one user and an iPhone 7 Plus to reduce user-habit and sensor-property influences.Source: B. Data Collection · p. 4
  • Data came from four off-the-shelf phones and five participants, with participants carrying phones in hand, pocket, and handbag.The mixed device and user data can also be applied to device and user identification.Source: B. Data Collection · p. 4
  • Large-scale tracking covered two office floors of about 1,650 m2 and 2,475 m2, using Google Tango attached to the participant’s chest for pseudo ground truth.Phones were placed in hand, pocket, or handbag, and participants walked freely in the building.Source: B. Data Collection · p. 4

IV. INERTIAL NAVIGATION MODELS

The section presents PDR and IONet as baselines and introduces L-IONet as a lightweight alternative. PDR builds trajectories from detected steps, estimated stride length, and heading, but its hand-built assumptions can produce long-term errors.

IV. INERTIAL NAVIGATION MODELS

  • The models include step-based PDR, raw-data IONet, and the proposed L-IONet, intended to make inertial tracking more accurate and efficient.PDR mitigates SINS drift from exponential to linear growth; IONet learns self-motion without hand-engineered motion features.Source: §IV · p. 4, Figure 3 · p. 5

A. Pedestrian Dead Reckoning

  • PDR detects strides from acceleration thresholds, estimates stride length with Weinberg’s empirical equation [31], derives heading from integrated gyroscope yaw, and updates location per step.Its output is a step-length and step-heading pair for constructing a 2D trajectory.Source: A. Pedestrian Dead Reckoning · pp. 4–5
  • PDR step classification can be difficult in practice, and its empirical stride-length model varies with users’ walking habits and body properties, causing accumulating errors over long operation.The limitation persists even when step detection and classification are accurate.Source: A. Pedestrian Dead Reckoning · p. 5

B. Inertial Odometry Neural Networks

IONet learns motion transformations from inertial sequences rather than relying on detected walking steps, allowing it to model more general motion. Its recurrent network compresses each sequence before predicting displacement and heading.

B. Inertial Odometry Neural Networks

  • IONet learns self-motion from raw inertial data and can handle motions such as trolley movement that lack the steps or periodicity patterns required by PDR.It predicts a location transformation over a fixed time window without explicit step detection or step-length estimation.Source: B. Inertial Odometry Neural Networks · pp. 5–6
  • IONet normalizes six-axis acceleration and angular-rate sequences, then uses an LSTM hidden state to retain input history and compress the sequence into a motion representation.The final hidden feature summarizes the full sequence for motion-transformation prediction.Source: B. Inertial Odometry Neural Networks · p. 5
  • A fully connected layer maps the final LSTM features to the polar transformation vector of displacement and heading, which is used to update location.The figure compares LSTM-based IONet with WaveNet-based L-IONet.Source: B. Inertial Odometry Neural Networks · p. 5, Figure 4 · p. 6

C. Lightweight Inertial Odometry Neural Networks

L-IONet replaces IONet’s recurrent model with a WaveNet-inspired feed-forward architecture for inertial tracking. Its causal dilated convolutions process long sequences without recurrent states, supporting parallel training and faster inference.

C. Lightweight Inertial Odometry Neural Networks

  • Compared with IONet, L-IONet is designed to use fewer resources and computations; the paper attributes its speed advantage to parallel training and faster convolutional operations.The authors also describe WaveNet’s dilated causal convolutions as easier to optimize on long continuous sensor signals than recurrent models.Source: C. Lightweight Inertial Odometry… · pp. 6–7
  • L-IONet replaces IONet’s recurrent model with an autoregressive feed-forward model that predicts polar vectors and connects them with previous states to reconstruct trajectories.The model has no recurrent module and fewer complex nonlinear operations, making it easier to train in parallel and faster at state inference.Source: C. Lightweight Inertial Odometry… · p. 6
  • Causal convolutions use current and past inputs, while stacked dilated layers expand the receptive area to capture long sequences without an oversized model.The causal design prevents the use of future states.Source: C. Lightweight Inertial Odometry… · p. 6
  • The gated activation combines filtered features and gates through element-wise multiplication; residual and skip connections support deeper regression models and improve non-linearity.The convolutional filters are indexed by layer, and the model uses residual and skip connections to enable a deeper structure.Source: C. Lightweight Inertial Odometry… · p. 6
  • Eight dilated-convolution layers, with dilation increasing from 1 to 128, are sufficient to process 200-frame inertial sequences.The dilation doubles at each layer.Source: C. Lightweight Inertial Odometry… · p. 6
  • L-IONet replaces WaveNet’s softmax output with pooling and a fully connected layer that maps extracted features to two-dimensional polar vectors.These vectors are connected with previous system states to produce current locations.Source: C. Lightweight Inertial Odometry… · pp. 6–7

D. Sliding Window

A fixed-size sliding window segments inertial readings into independent sequences; overlapping predictions are merged to reconstruct locations at a higher output rate.

D. Sliding Window

  • 200-frame windows with stride 10 feed the network, and overlapping predictions increase the output rate to 10 Hz.The current location is updated using the location 200 frames earlier, rather than the state 10 frames earlier.Source: D. Sliding Window · p. 7
  • A merging module connects each window’s predicted polar vector to generate locations, while low-pass filters smooth vectors and locations for trajectory reconstruction.Source: D. Sliding Window · p. 7
  • The 200-frame window corresponds to 2 seconds of inertial readings.Source: D. Sliding Window · p. 7

V. EXPERIMENTS

The experiments evaluate IONet and L-IONet on OxIOD for device inference, velocity estimation, and localization. Training and testing are split by sensor-attachment scenario, and models are also deployed on consumer phones and a smartwatch.

V. EXPERIMENTS

  • The experiments assess model performance on low-end devices, velocity estimation, and localization using IONet and L-IONet trained on OxIOD.Source: §V · p. 7

A. Setup

  • OxIOD is split into training and testing sets separately for handheld, pocket, handbag, and trolley attachment scenarios, using 200-frame windows and stride 10.Source: A. Setup · p. 7
  • Models are trained by minimizing mean squared error with ADAM at learning rate 1e−5, batch size 256, and one NVIDIA TESLA K80.Source: A. Setup · p. 7
  • The deployed models are converted to TensorFlow Lite and tested on three consumer smartphones and a Sony Smartwatch 2.The listed devices span phone configurations from 2 to 4 GB RAM and a smartwatch with 256 MB RAM.Source: A. Setup · p. 7

B. Model Performance at the edge

WaveNet-based L-IONet matches or exceeds the tested IONet configurations on accuracy while reducing model size and training time. Its lightweight models also support real-time inference on a smartwatch and retain accuracy across tested motion modes.

B. Model Performance at the edge

  • The edge study compares LSTM, GRU, and Basic RNN variants with WaveNet across model size, training speed, and accuracy, varying layer count and model-specific widths.Source: B. Model Performance at the edge · p. 7
  • 0.0069 prediction error for WaveNet (32) is the highest accuracy reported, slightly below the error of 1-layer Bi-LSTM (128), with one-quarter the parameters and roughly 10× faster training.Source: B. Model Performance at the edge · p. 7, Figure 5 · p. 8
  • On the smartwatch, L-IONet WaveNet (32) and WaveNet (16) achieved real-time inference at 56.78 ms and 27 ms per step, respectively, and ran faster than LSTM-based IONet models.GRUs had larger prediction errors than WaveNet (32): 0.0186, 0.0391, and 0.0626, around 3, 6, and 9 times higher. Basic RNNs had fewer parameters and faster inference but test errors of 0.268, 0.288, and 0.296.Source: B. Model Performance at the edge · p. 8
  • WaveNet-based L-IONet trains faster than Basic RNNs, LSTMs, and GRUs; the authors attribute this to feed-forward models being easier to train and optimize.Source: B. Model Performance at the edge · p. 8

C. Velocity and Heading Estimation

D. Deep Learning based Pedestrian Inertial Navigation

The experiments compare IONet and L-IONet with PDR across device attachments and on office floors outside the training room. The learned models outperform PDR across most attachment settings and generalize to the office floors.

D. Deep Learning based Pedestrian Inertial Navigation

  • Evaluation used Vicon ground truth for normal indoor walking on test-fold trajectories absent from training; training covered four device-attachment categories.IONet used a one-layer bidirectional LSTM with 128-dimensional hidden states, while L-IONet used WaveNet with 32 filters.Source: D. Deep Learning based Pedestrian… · p. 9
  • Both models localize well on two office floors despite training only in the Vicon room, indicating generalization to new environments.The office-floor trajectories are compared with pseudo ground truth from Google Tango.Source: D. Deep Learning based Pedestrian… · pp. 9–10, Figure 7 · p. 10
  • Across hand, pocket, and handbag settings, IONet and L-IONet outperform PDR; on trolley, PDR fails while the learned models reconstruct meaningful trajectories.L-IONet is closer to ground truth in handheld and trolley settings, while IONet performs better in-pocket, where device motion is more constrained.Source: D. Deep Learning based Pedestrian… · p. 9, Figure 8 · p. 10

VI. CONCLUSIONS AND FUTURE WORK

The paper presents L-IONet for inertial tracking and releases OxIOD for training and evaluating inertial navigation models. It reports competitive performance and efficient inference, including real-time use on smartphones and a smartwatch, while identifying broader data and model coverage as future work.

VI. CONCLUSIONS AND FUTURE WORK

  • L-IONet offers competitive performance over previous deep inertial odometry models with improved memory, inference, and training efficiency, including real-time inference on smartphones and a smartwatch.The paper also describes a systematic evaluation of deep inertial odometry on low-end devices.Source: §VI · p. 10
  • OxIOD is released as a large-scale, diverse dataset for training and evaluating inertial navigation models.The authors hope it will support future research in long-term ubiquitous ego-motion estimation.Source: §VI · p. 10
  • Future work targets 3D tracking, broader IoT-device coverage, shared benchmarks, network distillation, and WaveNet-style models for other sensor streams.The paper also identifies other machine-learning models, including DBN and BLS, and potential applications such as health monitoring and intelligent transportation.Source: §VI · pp. 10–11
Loading 2001.04061v1…