Characterizing Dynamic Walking Patterns and Detecting Falls with Wearable Sensors Using Gaussian Process Methods

Taehwan Kim, Jeongho Park, Seongman Heo, Keehoon Sung, Jooyoung Park2017 · DOI

TL;DR

This paper introduces an "auto-encoded Gaussian Process Dynamical Model" (auto-encoded GPDM) to characterize dynamic walking patterns from wearable sensor data by mapping high-dimensional sensor inputs to a low-dimensional latent space. This learned representation is then used with Gaussian Process (GP) based novelty detection to effectively identify falls before impact, offering a seamless and robust solution for fall detection.

What Problem Does This Paper Address?

The increasing prevalence of wearable sensors in smartphones and smartwatches presents a significant opportunity for health monitoring. A critical application in this domain is the automatic and prompt detection of falls, which can lead to serious injuries. Traditional methods relying on simple sensor value thresholds are prone to errors and false alarms. Characterizing the subtle nuances of human walking patterns is essential for distinguishing normal activity from falls. However, the high dimensionality and noisy nature of raw sensor data from multiple devices make direct modeling challenging. This paper aims to address this by developing a method to effectively represent dynamic walking patterns in a low-dimensional latent space and then leverage this representation for reliable fall detection.

Methods

The core of this paper lies in the novel combination of two key techniques:

  • Auto-encoded Gaussian Process Dynamical Model (auto-encoded GPDM): This approach extends the Gaussian Process Dynamical Model (GPDM) by incorporating an auto-encoder architecture. The GPDM is a GP-based state-space model adept at handling high-dimensional dynamical systems and quantifying uncertainty. The auto-encoder component, specifically a GP encoder, is trained to map high-dimensional sensor observations to a low-dimensional latent space. This process effectively learns an intrinsic representation of walking dynamics without requiring computationally intensive optimization for each new observation, unlike the original GPDM. The GP encoder provides latent samples that capture the essence of the walking patterns.
  • GP-based Novelty Detection: For the fall detection task, the paper leverages the latent samples generated during the training of the auto-encoded GPDM. These latent samples, representing normal walking behavior, are used to train a GP-based novelty detection model. When new, unseen latent samples are fed into this model, deviations from the learned normal patterns are identified, signaling a potential fall. This approach is efficient because it reuses the latent representations already computed for characterization.

The paper also details the specific kernel functions used for the GP models, including squared exponential, linear, and white noise kernels, chosen to capture smooth interpolations, periodic dynamics, and noise, respectively.

Datasets

The experiments were conducted using data collected from five male volunteers. Each volunteer wore two smartphones (iPhone 7) – one in their right pants pocket and another attached to their left wrist. These phones captured acceleration and angular velocity data along three orthogonal axes, along with their total magnitudes, resulting in a 16-dimensional feature vector per sensor. The data was collected at a sampling rate of 10 Hz. The experiments involved subjects walking predefined courses and then falling at random moments. The paper describes a five-fold cross-validation procedure where four subjects' data were used for training and one for testing.

Key Insights and Ideas

  • Low-Dimensional Latent Space for Walking: The auto-encoded GPDM successfully transforms complex, high-dimensional sensor data into a meaningful low-dimensional latent space. This space effectively captures the intrinsic characteristics of dynamic walking patterns, making them easier to analyze and distinguish.
  • Seamless Integration of Characterization and Detection: By using the latent samples generated by the auto-encoder for novelty detection, the paper achieves a seamless and efficient pipeline. The same learned representation serves both to understand walking and to detect anomalies (falls).
  • Pre-Impact Fall Detection: The method is capable of detecting falls before the body hits the ground, which is crucial for timely intervention and potentially mitigating injury severity.
  • Uncertainty Quantification: The inherent strength of Gaussian Processes in quantifying uncertainty is leveraged, which is vital for robust decision-making in dynamic systems and anomaly detection.

Key Results

The paper demonstrates that the auto-encoded GPDM effectively characterizes dynamic walking patterns, as evidenced by the cross-validation results where latent trajectories from test data closely matched those from training data. This indicates the model's ability to generalize and capture consistent walking signatures. For fall detection, the GP-based novelty detection on these latent trajectories showed promising results, with abnormal deviations clearly indicating falls. The paper visually presents these results through contour plots of the novelty detection scores and latent trajectories of fall events, showing clear separation from normal walking patterns. While specific quantitative metrics like accuracy are not explicitly detailed in the provided text, the qualitative results and visual evidence suggest a high degree of effectiveness.

Limitations and Future Work

The authors acknowledge several areas for future work:

  • More Extensive Experiments: The current study includes experiments with a limited number of subjects and falls. Future work should involve more extensive experiments and comparative studies to thoroughly reveal the strengths and weaknesses of the proposed approach.
  • Statistical Results: The paper mentions the intention to report statistical results for applicability verification, suggesting that detailed statistical analysis might be a future step.
  • Alternative Auto-Encoders: Exploring other types of auto-encoders for nonlinear dimensionality reduction could be beneficial.
  • More Wearable Sensors: Investigating the use of a larger number of wearable sensors could potentially enhance the robustness and accuracy of the system.
  • Deployment: The authors express a desire to deploy these methods in current smartphone and smartwatch systems, indicating a focus on practical implementation.

Summary updated

BibTeX
@misc{kim2017characterizing,
  author = {Taehwan Kim and Jeongho Park and Seongman Heo and Keehoon Sung and Jooyoung Park},
  title = {{Characterizing Dynamic Walking Patterns and Detecting Falls with Wearable Sensors Using Gaussian Process Methods}},
  year = {2017},
  doi = {10.3390/s17051172},
}
Figure