High-Performance Real-Time Human Activity Recognition Using Machine Learning

Human Activity Recognition (HAR) is a vital technology in domains such as healthcare, fitness, and smart environments. This paper presents an innovative HAR system that leverages machine-learning algorithms deployed on the B-L475E-IOT01A Discovery Kit, a highly efficient microcontroller platform designed for low-power, real-time applications. The system utilizes wearable sensors (accelerometers and gyroscopes) integrated with the kit to enable seamless data acquisition and processing. Our model achieves outstanding performance in classifying dynamic activities, including walking, walking upstairs, and walking downstairs, with high precision and recall, demonstrating its reliability and robustness. However, distinguishing between static activities, such as sitting and standing, remains a challenge, with the model showing a lower recall for sitting due to subtle postural differences. To address these limitations, we implement advanced feature extraction, data augmentation, and sensor fusion techniques, which significantly improve classification accuracy. The ease of use of the B-L475E-IOT01A kit allows for real-time activity classification, validated through the Tera Term interface, making the system ideal for practical applications in wearable devices and embedded systems. The novelty of our approach lies in the seamless integration of real-time processing capabilities with advanced machine-learning techniques, providing immediate, actionable insights. With an overall classification accuracy of 90%, this system demonstrates great potential for deployment in health monitoring, fitness tracking, and eldercare applications. Future work will focus on enhancing the system’s performance in distinguishing static activities and broadening its real-world applicability.