PhysioFit: A Inline Activity Recognition Leveraging Physiological Sensors

This research presents a novel approach to fitness monitoring using multi-modal data fusion and deep learning techniques implemented on a Raspberry Pi platform. The system utilizes four Inertial Measurement Unit (IMU) sensors and a microphone to capture motion and audio data during exercise. A custom Convolutional Neural Network (CNN) model processes this multi-modal data to classify exercises and estimate key health metrics. The study involved creating an open-source dataset comprising 11 subjects performing 25 different exercises, which was used to train and evaluate the model. The system achieved a 93.2% accuracy in exercise classification, significantly outperforming single-modality approaches. Health metric estimations, including blood pressure, oxygen saturation, and calorie burn, showed promising results with low Mean Absolute Errors. Real-time performance tests on the Raspberry Pi demonstrated the system's capability for immediate feedback, processing an average of 30 inferences per second with a latency of 0.3 seconds. User experience studies indicated high satisfaction with the system's usability and perceived accuracy. This research contributes to the field of fitness monitoring by demonstrating the effectiveness of a multimodal approach and the feasibility of implementing complex deep learning models on accessible hardware, paving the way for more comprehensive and widely available fitness tracking solutions.