A Multi-Activity Fusion Approach for Gender Recognition based on Human Activity

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Gender recognition is increasingly important in healthcare, sports, and wearable technology.
• Wearable sensors offer a non-invasive method for collecting physiological data.
• Supervised machine learning models can analyze complex activity data for classification tasks.

Purpose of the Study:

• To develop and evaluate a gender recognition system using wearable inertial measurement units (IMUs).
• To identify optimal sensor placements and activity types for accurate gender classification.
• To compare the performance of different machine learning algorithms for this task.

Main Methods:

• Utilized a wearable sensor system with five IMUs placed on the upper and lower body.
• Recorded seven daily activities including standing, walking, and climbing exercises.
• Applied supervised machine learning, specifically Random Forest Classifier (RFC) and Support Vector Machines (SVM), for gender classification.

Main Results:

• Single activity classification achieved a maximum accuracy of 92.06% with RFC using ankle sensor data during walking.
• Multi-activity classification significantly improved accuracy, reaching 94.13% with RFC.
• The highest accuracy was obtained using a combination of Romberg test (eyes open), single leg stance (eyes open), and staircase climbing activities.

Conclusions:

• Wearable sensor data combined with machine learning provides an effective method for gender recognition.
• Analyzing multiple activities enhances classification accuracy compared to single activities.
• Optimal sensor placement and activity selection are crucial for maximizing gender recognition performance.