A Wearable Multi-Sensor Fusion Approach for Gender Recognition based on Deep Learning

Human activity recognition (HAR) has gained significant attention over the last decade due to its usefulness in various fields, including healthcare, sports, rehabilitation, and wearable technology. HAR involves using sensors, such as wearables, to automatically identify human activity. Recently, researchers have started using HAR data to recognize subject attributes like age and gender, making biometric analysis a critical complement to activity recognition tools. This study presents a new and adaptable deep learning approach to recognize gender based on a variety of activities utilizing wearable sensor systems equipped with Inertial Measurement Units (IMU). The system includes five sensors placed on the upper and lower body during seven standing, walking, and climbing-related tasks that mimic daily activities. Using both single and multi-head Convolutional Neural Networks (CNN) with standalone and fused body location sensors, we conducted a comprehensive study to build a gender recognition model. The study identifies a set of sensor placements and activities that result in more accurate gender detection. Our results are compared to previous studies that used classical machine learning and deep learning models for gender recognition, considering both simple and complex activities on three different datasets - two public datasets and our collected dataset. Our proposed CNN model exhibits accurate gender detection in simpler activities, such as walking and Romberg tests, with almost 90% accuracy when using the chest sensor. Furthermore, our experimental evaluation demonstrates excellent gender detection performance for more complex activities, such as timed-up-and-go and climbing stairs. By utilizing a multi-head CNN and merging data from both chest and waist sensors, the model achieves a prediction accuracy of up to 97%.