Ambulatory Human Gait Phase Detection Using Wearable Inertial Sensors and Hidden Markov Model

Long Liu, Huihui Wang, Haorui Li, Jiayi Liu, Sen Qiu, Hongyu Zhao, Xiangyang Guo2021 · DOI

TL;DR

This paper presents a novel method for accurately detecting human gait phases using wearable inertial sensors and a Hidden Markov Model (HMM). By adapting HMM parameters to individual gait patterns, the system achieves high recognition rates, paving the way for more reliable and personalized gait analysis in healthcare and rehabilitation.

What Problem Does This Paper Address?

Walking is a fundamental human activity, but gait abnormalities can signal underlying neurological or musculoskeletal issues. Accurately identifying different phases of the gait cycle (e.g., heel-strike, toe-off) is crucial for diagnosing diseases like Alzheimer's and Parkinson's, assessing recovery after surgery, and guiding rehabilitation. Traditional gait analysis methods often rely on complex lab equipment or have limitations in real-world ambulatory settings. Existing wearable sensor-based approaches, while promising, struggle with the inherent variability of human gait due to individual differences, environments, and changing physical states, leading to reduced accuracy and adaptability. This paper aims to bridge this gap by developing a robust and adaptive gait phase detection system using readily available inertial sensors.

Methods

The core of this research lies in the application of a Hidden Markov Model (HMM) for gait phase segmentation. HMMs are well-suited for time-series data with hidden states, making them ideal for modeling the sequential nature of gait. The authors propose a 15-dimensional temporal characteristic feature extracted from smoothed and filtered angular velocity data obtained from inertial sensors.

A key innovation is the parameter adaptation of the HMM. Recognizing that a fixed HMM model struggles with gait variability, the researchers employ Maximum a Posteriori (MAP) and Minimum Likelihood Linear Regression (MLLR) techniques. These methods adjust the HMM's model parameters (mean and variance) based on the specific gait data being analyzed, allowing the model to adapt to individual walking patterns and improve recognition accuracy. The data acquisition system involves a wearable sensor placed on the toe, transmitting data wirelessly to a mobile phone for real-time processing and storage.

Datasets

The study utilized data collected from sixteen healthy volunteers (eleven males, five females) aged 30-60, with varying heights and weights. Gait data was collected while participants walked normally on level ground. The inertial sensor data (angular velocity) was sampled at 100 Hz. The paper does not mention using any publicly available datasets for training or evaluation, indicating that the dataset was collected specifically for this research.

Key Insights and Ideas

  • HMMs are powerful for sequential data: The inherent sequential and periodic nature of human gait makes HMMs a natural fit for gait phase recognition.
  • Adaptation is crucial for real-world gait analysis: A static HMM model performs poorly when faced with the diverse gait patterns of different individuals and conditions. Adaptive parameter estimation is essential for achieving high accuracy in ambulatory settings.
  • Wearable sensors offer practical gait analysis: The use of a single inertial sensor placed on the toe provides a convenient and unobtrusive way to collect gait data, enabling ambulatory and outdoor monitoring.
  • Feature engineering matters: Extracting relevant temporal features from angular velocity data is critical for the HMM's performance.
  • MAP and MLLR offer effective adaptation: These parameter adaptation techniques significantly improve the HMM's ability to recognize gait phases across different individuals.

Key Results

The proposed HMM-based gait phase detection system, particularly with parameter adaptation, demonstrated strong performance.

  • The best recognition rate achieved was 91.88% using the MAP adaptation method.
  • Without parameter adaptation, the HMM achieved a recognition rate of 91.59%.
  • The study found that a sliding window size between 5 and 15 was optimal for feature extraction, and a window size of 14 yielded the best results for 3D features.
  • When comparing different feature dimensions, 15 features resulted in the highest recognition rate of 89.54% with a sliding window size of 9.
  • The adaptive HMM model showed improved accuracy in correctly identifying gait states S1 and S4, with a slight decrease in S3 but an overall increase in identification rate.
  • Comparison with other algorithms (KNN, LR, LDA, RF, NBM, SVM) showed that while some achieved comparable recognition rates (80-92%), they did not account for the crucial state transition dynamics, which the HMM inherently handles.

Limitations and Future Work

The authors acknowledge several limitations:

  • HMM's memoryless nature: The standard HMM model does not inherently capture long-term dependencies or contextual information, which can be important for complex gait analysis. The authors suggest that building a "high-level HMM model" could address this.
  • Computational cost: While the proposed methods are effective, deep learning models like CNNs and RNNs, which have shown promise in clinical gait analysis, require significant computation, potentially hindering real-time monitoring.
  • Future work: The authors plan to conduct more comparative trials to understand the deep mechanisms affecting parameter adaptation performance. They also aim to explore deep learning models and address the real-time monitoring challenge. Additionally, they intend to investigate how to find the deep mechanism of what kind of factors affect the performance of parameter adaptation.

Summary updated

BibTeX
@misc{liu2021ambulatory,
  author = {Long Liu and Huihui Wang and Haorui Li and Jiayi Liu and Sen Qiu and Hongyu Zhao and Xiangyang Guo},
  title = {{Ambulatory Human Gait Phase Detection Using Wearable Inertial Sensors and Hidden Markov Model}},
  year = {2021},
  doi = {10.3390/s21041347},
}
Figure