AIoT-Driven Human Activity Recognition for Versatile Framework on Multipurpose Applications

This paper presents the implementation of an AIoT-based Human Activity Recognition (HAR) framework designed for multipurpose applications. The framework integrates sensor data from wearable IoT devices, which is then processed by AI algorithms to classify and predict human activities in real-time. It utilizes machine learning models, particularly machine learning techniques, to analyze complex activity. Our framework uses all open-source IDEs, which is able to replicate and modify. For this paper we give examples of how to use our framework as gamification of a workout practice. The Idea is to recognize user activity including (bicep curl, shoulder press, and front rise) using accelerometer data, and then send recognized activity to our developed online game (pop-balloon). M5StickC Plus used as the hardware which already equipped with IMU sensor and power management. Furthermore, it has small form factor fit for wearable application. Before real-time performance was taken, we evaluated 5 different machine learning model and choose which one is more optimized. The five models include Naïve Bayes, Support Vector Machine (SVM), AdaBoost, ZeroR, and Random Forest. Accuracy given for offline analysis were 97.68%, 98.97%, 41.62%, 25%, 100% respectively to the previous model. In the final, for real-time performance we choose SVM model which most optimized even though the accuracy reduced to 89.67% for this task.