Abstract:
Inertial Measurement Units (IMU) or Inertial Sensors are among the most commonly
used wearable sensors. They can capture important features and patterns related to
movement and activity. Therefore, they play an essential role in many domains includ ing clinical investigations, navigation, activity recognition, and estimation of soft bio metrics such as age, gender, height, and age. Currently, human age estimation systems
relying on sensor based gait data either use classical machine learning methods such as
Support Vector Machines (SVMs), decision-trees based models such as Random Forest,
or deep learning techniques such as Convolutional Neural Networks (CNNs) on Recur rent Neural Networks(RNNs). However, there is a need to explore the state-of-the-art
Transformer model because of their ability to learn complex patterns and dependencies
from long, sequential data due to their attention mechanism. This research explored a
novel approach to estimating the human age using a custom Transformer model. Our
model leverages sensor-based inertial gait data to classify individuals into six distinct
age groups. The data was divided into twelve class-balanced datasets and twelve deep learning experiments were run using the proposed solution. It achieved an average test
accuracy of 88.91% and an average validation accuracy of 90.01%. The minimum and
maximum Precision-Recall Area Under the Curve (AUC) for any age group was 0.85
and 0.95, respectively. The results obtained from the proposed solution outperformed
many existing techniques for age estimation, such as those based on Convolutional Neu ral Networks (CNNs) on Recurrent Neural Networks (RNNs). Furthermore, the work
at hand shows the suitability of the Transformer model for human age estimation from
inertial sensor data.
