Human activity recognition has become a popular research topic in recent years, due to its many practical applications in areas such as smart homes, healthcare, robotics and sports. Researchers are interested in developing systems that can identify user activities from raw data with minimal resource usage. Deep learning is a promising technology for analyzing large datasets quickly and efficiently. However, previous methods that relied solely on smartphone accelerometers were limited in their ability to recognize complex, real-time human activities. In this study, we propose a transformer model that is trained on input datasets to extract low-level, high-level, and complex features and classify human activities. We use three publicly available datasets (UCI-HAR, WISDM, and MHealth) to test the system. Despite limited hardware resources, our proposed system performs well and provides satisfactory activity identification.