Indoor Trajectory Reconstruction of Walking, Jogging, and Running Activities Based on a Foot-Mounted Inertial Pedestrian Dead-Reckoning System

Jesus D. Ceron, Christine F. Martindale, Diego M. López, Felix Kluge, Björn M. Eskofier

TL;DR

This paper presents a robust pipeline for reconstructing indoor human trajectories using a single foot-mounted IMU, extending beyond simple walking to include jogging and running. By improving stride segmentation and gait event detection (Toe-Off and Mid-Stance), the system achieves significantly better accuracy in complex movements and diverse environments, paving the way for more sophisticated indoor navigation and activity monitoring.

What Problem Does This Paper Address?

Indoor positioning systems (IPS) are crucial for applications like rehabilitation, navigation for visually impaired individuals, and emergency response. While infrastructure-based IPS exist, infrastructure-free methods like Inertial Pedestrian Dead-Reckoning (IPDR) offer advantages in cost and privacy. However, traditional IPDR systems struggle with accuracy due to sensor drift, especially when users perform activities beyond simple walking, such as jogging, running, or natural movements like stopping and turning. Existing research often focuses on controlled environments and limited activities, leaving a gap in reconstructing trajectories in more realistic, dynamic indoor scenarios. This paper aims to bridge that gap by developing and evaluating an IPDR system capable of handling a wider range of human gaits and movements.

Methods

The proposed pipeline builds upon existing IPDR techniques but introduces several key improvements:

  • Stride Segmentation: A multi-subsequence Dynamic Time Warping (msDTW) approach is used for segmenting individual strides. This method is adapted to generate templates for walking, jogging, and running, trained on a large number of strides.
  • Gait Event Detection (Toe-Off and Mid-Stance): Novel algorithms are introduced for detecting Toe-Off (TO) and Mid-Stance (MS) events. The TO detection algorithm is enhanced to find the maximum peak in the gyroscope signal and then the nearest zero-crossing before it, improving accuracy in dynamic gaits. The MS detection algorithm is adapted to focus on a specific portion of the stride and identify the point of lowest energy, making it more robust for jogging and running.
  • Stride Length and Orientation Estimation: A Zero Velocity Update (ZUPT) algorithm is employed, empowered by a complementary Kalman filter. This filter helps to model and correct errors in velocity and position estimates, particularly crucial for the accuracy of stride length and orientation calculations. The ZUPT detection is adapted to use the detected MS events, making it more reliable for non-walking gaits.

Datasets

The study utilizes two datasets:

  • Unicauca Dataset: This dataset was collected specifically for this study at the University of Cauca, Colombia. It features 10 participants performing walking, jogging, and running on a "P"-shaped path (approximately 150m). This dataset was used for evaluating the system's performance in a controlled, yet more dynamic, setting than typical benchmarks.
  • FAU Dataset: This dataset is based on a previous study and is publicly available. It includes gait data from 80 participants performing various task-driven activities in an indoor environment. For this study, data from 56 participants (Germany) were used for evaluation, and data from 24 participants (Slovenia) were used for training. This dataset provides a richer variety of activities, including non-circuit tasks and circuit activities, offering a more comprehensive evaluation.

Key Insights and Ideas

  • Beyond Walking: The most significant insight is that improving gait event detection (TO and MS) is critical for extending IPDR accuracy beyond simple walking to jogging and running. The proposed algorithms demonstrate this by achieving high F-scores for these more dynamic activities.
  • The Importance of Accurate Gait Events: The study highlights that even with a well-tuned complementary filter, accurate trajectory reconstruction hinges on precise TO and MS detection. When these events are detected correctly, the complementary filter can effectively model errors, even in complex gaits.
  • Challenges of Non-Stride Movements: Activities involving stopping, sitting, or lateral/backward steps pose a challenge because their generated signals differ from typical walking/running templates. This suggests a need for more sophisticated segmentation strategies or specialized templates for such movements.
  • Dataset Design Matters: The study emphasizes the importance of well-labeled datasets (activity and stride/step levels) and precise start/end points for participants to enable fair and accurate evaluation of IPDR systems.

Key Results

  • Activity Classification: The SVM classifier achieved high accuracy (90% for Unicauca, 93% for FAU) in classifying walking, jogging, and running activities, with most misclassifications occurring between jogging and running.
  • TO and MS Detection: The proposed TO and MS detection algorithms achieved high F-scores, particularly for jogging and running activities, outperforming previous methods. For walking, a perfect F-score was obtained. For non-stopping activities, TO detection F-scores ranged from 92% to 100%, and for activities involving stopping, they were between 67% and 70%.
  • Trajectory Reconstruction:
    • Unicauca Dataset: Return Position Errors (RPE) were generally low, less than 1% for walking and jogging, and around 1.4% for running. Strides out of trajectory (SOT) were also kept to a minimum.
  • FAU Dataset: RPEs were higher than in the Unicauca dataset, attributed to limitations in the dataset's ground truth and continuous activity performance. However, the proposed system still showed improved performance over previous methods for most activities. For non-stopping activities, RPEs ranged from 0.51% to 8.67%, and for activities with stopping, they were between 8.79% and 27.36%.

Limitations and Future Work

  • Non-Stride Movements: The current system struggles with activities involving frequent stopping, sitting, or lateral/backward steps due to signal differences from walking/running templates. Future work could explore using specialized templates for these movements or employing threshold-free methods like hierarchical hidden Markov models (hHMM).
  • Jogging vs. Running Classification: The SVM classifier occasionally confused jogging and running. Future research could investigate personalized models to improve this distinction.
  • FAU Dataset Limitations: The FAU dataset's inherent limitations (imprecise start/end points, continuous activity) impacted RPE accuracy. Future datasets should ensure precise participant positioning and potentially include periods of stillness for gyroscope bias calculation.
  • Future Directions: The authors suggest exploring hHMM for improved stride segmentation, fusing symbolic location data (e.g., from RSSI signals) to update indoor localization, and developing more robust methods for activities involving non-stride movements.

Summary updated

BibTeX
@inproceedings{ceron2020indoor,
  author = {Jesus D. Ceron and Christine F. Martindale and Diego M. López and Felix Kluge and Björn M. Eskofier},
  title = {{Indoor Trajectory Reconstruction of Walking, Jogging, and Running Activities Based on a Foot-Mounted Inertial Pedestrian Dead-Reckoning System}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=_Yo7-4cfyZs},
}
Figure