Human activity recognition (HAR) leverages artificial intelligence (AI) to classify activities using raw data obtained from wearable inertial measurement unit (IMU) sensors. Applications can be found in healthcare, surveillance, and remote elderly care. However, the inherent complexity of HAR systems, requiring advanced algorithms and substantial computational resources, challenges its scalability and real-time processing, especially in resource-constrained settings. This paper proposes an approach to alleviate these limitations by employing AI on the edge for HAR, where machine learning models are deployed onto a microcontroller, shifting the computational workload closer to the data source. Throughout this study, the performances of different input variations are compared and analyzed. In the result section, it shows clearly that the model trained on statistical features outperformed the one trained on the raw IMU sensor data. In addition, experiments are performed to demonstrate the viability and effectiveness of the on-edge implementation of both inputs. Finally, conclusions and directions for future improvements are discussed.