Wearable Sensor-Based Gait Analysis for Age and Gender Estimation

Md Atiqur Rahman Ahad, Trung Ngo Thanh, Anindya Das Antar, Masud Ahmed, Tahera Hossain, Daigo Muramatsu, Yasushi Makihara, Sozo Inoue, Yasushi Yagi

TL;DR

This paper presents an analysis of a large-scale wearable sensor-based gait analysis challenge focused on age and gender estimation. The study highlights that deep learning approaches, particularly those leveraging angle-embedded gait dynamic images and temporal convolutional networks, achieved the best performance, outperforming traditional handcrafted methods. This work underscores the potential of gait analysis for non-intrusive demographic estimation in healthcare and beyond.

What Problem Does This Paper Address?

As global populations age, there's a growing need for automated healthcare solutions that can provide personalized support. Understanding a user's age and gender is crucial for tailoring these systems. While gait is a rich biometric that can reveal information about a person's physical and medical condition, accurately estimating age and gender from gait data, especially using wearable sensors, remains a challenge. Existing research often relies on small datasets, lacks diverse age ranges, or has imbalanced gender distributions, limiting generalizability. This paper addresses this gap by analyzing a comprehensive challenge that utilized a large, diverse dataset to push the boundaries of sensor-based gait analysis for age and gender estimation.

Methods

The core of this paper is an analysis of the methods submitted by participants in the 12th IAPR International Conference on Biometrics (ICB) gait analysis challenge. The challenge encouraged the development of smart approaches to estimate age and gender from wearable sensor data. The paper categorizes the submitted algorithms based on their preprocessing steps, feature extraction techniques, and classification models. Key approaches explored include:

  • Sensor Orientation Invariance: Several teams tackled the challenge of inconsistent sensor placement and orientation by employing methods like using the magnitude of sensor signals, PCA-based rotation matrices, or random rotations.
  • Feature Extraction: A wide array of features were extracted, ranging from basic statistical measures (mean, standard deviation, variance) to more complex representations like Gait Dynamic Images (GDI) and Angle Embedded Gait Dynamic Images (AE-GDI). Temporal features like gait cycle time and step length were also utilized.
  • Classification Models: Both classical machine learning algorithms (SVM, Random Forest, KNN) and deep learning architectures (CNN, LSTM, BLSTM, TCN) were employed. Deep learning models, particularly those that could learn complex spatio-temporal patterns, showed a clear advantage.

Datasets

The study utilized a large, multi-modal gait dataset collected using three IMUZ inertial measurement units (IMUs) placed on a waist belt.

  • Training Dataset: Consisted of gait data from 745 subjects (388 males, 357 females) with an age range of 2 to 78 years. This dataset was composed of two subsets: the OU-ISIR Gait Database (Inertial Sensor Dataset) and the OU-ISIR Gait Database (Similar Action Inertial Dataset). While the age distribution was non-uniform, with a higher concentration of young children, it offered a broad age spectrum and a relatively balanced gender distribution.
  • Test Dataset: Comprised data from 58 subjects (39 males, 19 females). This dataset presented increased challenges, including variations in sensor location (backpack vs. waist belt) and orientations, and importantly, lacked labels for age and gender.

Key Insights and Ideas

  • Deep Learning Dominance: Deep learning-based solutions consistently outperformed traditional handcrafted feature-based methods, especially when dealing with the complexities of sensor data.
  • AE-GDI as a Powerful Representation: The Angle Embedded Gait Dynamic Image (AE-GDI) representation, which is invariant to rotation and translation, proved highly effective when combined with temporal convolutional networks.
  • Sensor Inconsistency is a Major Hurdle: The significant variation in sensor placement and orientation across subjects and even within a single subject's data posed a major challenge. Teams that explicitly addressed this inconsistency generally achieved better results.
  • Age vs. Gender Performance: Younger subjects generally yielded better age estimation performance compared to older age groups, suggesting that gait patterns might change more drastically with age in later life.
  • Dataset Diversity is Key: The large and diverse dataset was crucial for evaluating the generalizability of the proposed methods, revealing that some algorithms struggled with robustness and generalizability.

Key Results

The competition saw 18 teams register, with 10 ultimately submitting results. A total of 67 algorithms were submitted (32 for gender estimation, 35 for age estimation).

  • Best Performance: The top-performing algorithm, achieving the best results for both age and gender estimation, was an approach that utilized the Angle Embedded Gait Dynamic Image (AE-GDI) representation and a temporal convolutional network.
  • Gender Estimation: Achieved a prediction error of 24.23%.
  • Age Estimation: Achieved a mean absolute error of 5.39 years.
  • Other Top Performers:
    • For gender estimation, the second-best result was 24.74% (Team T8, Alg.1), and the third-best was 30.41% (Team T5, Alg.1 & Alg.2).
  • For age estimation, another algorithm from Team T7 (Alg.7) achieved the second-best result with a mean absolute error of 5.94 years. Team T5 (Alg.1) and Team T8 (Alg.1) also performed well with mean absolute errors of 6.44 years and 6.62 years, respectively.
  • Worst Performers: Teams T1 and T10 generally produced the poorest results, with some algorithms showing very inaccurate predictions.

Limitations and Future Work

  • Sensor Inconsistency: While addressed by some, the challenge of sensor location and orientation inconsistency remains a significant limitation for many approaches. Methods that did not explicitly handle this often achieved inferior results.
  • Dataset Limitations: Despite its size, the training dataset exhibited a non-uniform age distribution, with fewer participants in the very young and very old age groups. The test dataset also presented challenges with unlabeled data and varied sensor placements.
  • Generalizability: The paper notes that some algorithms showed poor generalizability and robustness, indicating that further work is needed to develop models that perform consistently across different sensor setups and subject demographics.
  • Future Directions: The authors suggest exploring further strategies to improve results, potentially by introducing similar challenges to foster research development. They also hint at the possibility of exploring multi-task learning and more robust sensor-invariant techniques.

Summary updated

BibTeX
@inproceedings{ahad2020wearable,
  author = {Md Atiqur Rahman Ahad and Trung Ngo Thanh and Anindya Das Antar and Masud Ahmed and Tahera Hossain and Daigo Muramatsu and Yasushi Makihara and Sozo Inoue and Yasushi Yagi},
  title = {{Wearable Sensor-Based Gait Analysis for Age and Gender Estimation}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=WfxKeQdif_},
}
Figure