Human Activity Recognition with an HMM-Based Generative Model

Narges Manouchehri, Nizar BouguilaSensors 2023 · DOI

TL;DR

This paper introduces a novel Human Activity Recognition (HAR) model called Scaled Dirichlet-based Hidden Markov Model (SD-HMM). By employing a more flexible distribution for emission probabilities, SD-HMM aims to improve the accuracy and robustness of activity recognition, particularly in scenarios with complex or proportional data, outperforming existing methods on benchmark datasets.

What Problem Does This Paper Address?

Human Activity Recognition (HAR) is a crucial area with applications ranging from healthcare monitoring to elder care and disease diagnosis. While various methods exist, challenges remain in accurately capturing the sequential nature of human activities, especially when dealing with noisy, complex, or proportional data. Traditional Hidden Markov Models (HMMs) often rely on Gaussian assumptions for emission probabilities, which may not be suitable for all types of data. Furthermore, deep learning approaches, while powerful, are often data-hungry and lack interpretability. This paper addresses the need for a more flexible and interpretable generative model for HAR that can handle proportional data effectively and overcome the limitations of existing methods.

Methods

The core contribution of this paper is the development of a Scaled Dirichlet-based Hidden Markov Model (SD-HMM). Unlike conventional HMMs that often assume Gaussian emission probabilities, SD-HMM utilizes a scaled Dirichlet distribution. This choice is motivated by the nature of proportional data, where the scaled Dirichlet distribution offers greater flexibility in modeling various data shapes.

To estimate the parameters of this complex model, the authors employ variational inference (VI). VI is presented as an efficient and accurate approximation technique that balances computational cost with performance, avoiding issues like local maxima and overfitting often encountered with maximum likelihood estimation. The paper details the variational learning procedure, including defining prior distributions, deriving a lower bound on the model's likelihood, and iteratively updating the model's parameters (transition matrix, emission matrix, and distribution parameters) using a mean-field approximation.

Datasets

The proposed SD-HMM model was evaluated on publicly available datasets:

  • Opportunity Dataset: This dataset was collected using wearable and object-based sensors. It includes data from individuals performing various activities. The authors specifically analyzed four daily activities: standing, walking, lying, and sitting. They also addressed challenges within this dataset, such as class imbalance (using SMOTE for oversampling) and missing values (imputed with the median).
  • UCI HAR Dataset: This dataset was collected from 30 individuals wearing a smartphone on their waist, capturing accelerometer and gyroscope data. The activities analyzed were standing, walking, lying, and sitting.

The paper does not introduce a new benchmark dataset but rather uses these established datasets to compare the performance of their proposed model against existing methods.

Key Insights and Ideas

  • Flexibility of Scaled Dirichlet: The key insight is that the scaled Dirichlet distribution offers a more adaptable way to model emission probabilities in HMMs, especially for proportional data, leading to improved HAR performance.
  • Variational Inference for Efficiency: Using variational inference provides a computationally efficient way to train complex generative models like SD-HMM, making them more practical for real-time applications.
  • Generative Models for Interpretability: By focusing on generative models like HMMs, the approach offers a degree of interpretability that is often lacking in purely discriminative deep learning models.
  • Addressing Data Challenges: The paper demonstrates a practical approach to handling common real-world data issues like class imbalance and missing values, which are critical for robust HAR systems.

Key Results

The SD-HMM model demonstrated superior performance across the evaluated datasets:

  • Opportunity Dataset (First Run): SD-HMM achieved an accuracy of 89.33%, outperforming D-HMM (85.96%) and GMM-HMM (86.08%). Similar improvements were observed in precision, recall, and F1-score.
  • Opportunity Dataset (Second Run): SD-HMM achieved an accuracy of 87.12%, again surpassing D-HMM (85.96%) and GMM-HMM (85.14%).
  • UCI HAR Dataset: SD-HMM achieved an accuracy of 86.17%, outperforming D-HMM (85.22%) and GMM-HMM (84.31%).

The authors attribute the improved performance to the increased flexibility of the scaled Dirichlet distribution, which allows for better modeling of the data's characteristics.

Limitations and Future Work

The authors acknowledge several areas for future work:

  • Exploring Other Emission Probabilities: While the scaled Dirichlet distribution showed promise, they plan to investigate other alternatives for emission probabilities.
  • Handling Outliers: The paper suggests that their model might be more robust to outliers, but further investigation and explicit methods for outlier handling could be explored.
  • Feature Selection: The authors intend to introduce feature selection techniques in future work to further enhance the model's performance and efficiency.
  • Testing on More Individuals: The current evaluation was performed on a limited number of individuals; future work will involve testing on a larger and more diverse population.

Summary updated

BibTeX
@article{manouchehri2023human,
  author = {Narges Manouchehri and Nizar Bouguila},
  title = {{Human Activity Recognition with an HMM-Based Generative Model}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23031390},
}
Figure