Application and adoption of Human activity Recognition (HAR) capable wearable devices in industrial and academic research, has been increasing in recent years. One of the biggest challenges it faced has been to make it more efficient to be able to implement without the need for high-performance computing. The ability to train and compile the data in the edge devices without the need to connect to the remote servers, cloud, or powerful computing devices for processing and analyzing can provide an advantage in terms of response latency, data security and privacy, communication network efficiency, data transfer speed, and power consumption. In this research, we implemented the concept of TinyML for activity recognition with optimizing techniques to optimize the model, reducing the size of the model to be able to deploy the model in microcontroller units of portable wearable devices for on-device intelligence. For these experiments, we used two publicly available HAR datasets. The preliminary experimental results show that the size can be reduced up to 10 times using TinyML methods without much difference in the performance.