Novel Deep Learning Models for Optimizing Human Activity Recognition Using Wearable Sensors: An Analysis of Photoplethysmography and Accelerometer Signals

Abstract

Human activity recognition enables identifying the particular activity of an individual by analyzing sensor data. Wearable sensors are often utilized in this method to gather and categorize data. Many wearable devices include sensors for monitoring heart rate and detecting body posture. In this research, we experimented with the Photoplethysmography sensor for determining heart rate and accelerometer signals for recognizing body position to classify human activities such as squats, stepper, and resting. We have developed two novel deep learning models, ResTime and Minception, that can effectively recognize human activities through sensor data. These models identified the appropriate time intervals for activity recognition, which led to a decrease in false positives and false negatives. Our experiments on the PPG dataset yielded exceptional accuracy results, with ResTime and Mincep achieving 98.73$\%$ and 98.79$\%$, respectively, surpassing other existing models. We also discovered that by adjusting the window size and selecting the appropriate model, we were able to optimize accuracy and minimize false positives or negatives. This allows for a more sophisticated decision-making system for recognizing human activities utilizing wearable sensor sensors.