Abstract— Different applications used cloud computing, machine learning, and the Internet of Things (IoT). Transferring data from the local network to the cloud for processing causes huge traffic and delay. IoT services, like Human Activity Recognition (HAR), use IoT edge options to be near the place of telemetry data generation that decreases traffic and speeds up the results. This study used three smartphones with built-in accelerometers; three parameters from each accelerometer to predict human activities. While building the models at the Raspberry PI edge, the most important features were determined using Principal Component Analysis (PCA). Light GBM, Extra Trees, and Random Forest algorithms were employed to evaluate the best models. Significant performance improvements in training and real-time results were achieved using the top related features at the IoT edge. The Light GBM recognized four different activities with 99.6% accuracy when all nine features were used, and with more than 98% accuracy when less than half of the features were used. To process one prediction, Raspberry PI 3 took 6.1 milliseconds, Raspberry PI 4 took less than 3 milliseconds if all features are used, while Microsoft Azure cloud took 5.8 seconds, including prediction time and network latency.