Edge-Based Human Activity Recognition: A Novel Approach Using Spectral Analysis and Deep Learning

Human Activity Recognition plays a crucial role in various applications extending from healthcare to smart environments. In this paper, we present a novel approach for HAR using edge computing resources. Our method combines spectral analysis with deep learning techniques to efficiently extract relevant features from raw sensor data. By employing spectral analysis, we successfully reduce the feature space from 561 to 39, while maintaining high accuracy, low loss, and improved performance metrics such as F1 score and recall. Furthermore, we demonstrate the adaptability of our approach by deploying a quantized neural network model onto a resource-constrained edge device, specifically the NodeMCU microcontroller. This enables real-time HAR inference at the edge, making our solution suitable for applications where computational resources are limited. Experimental results on the UCU HAR dataset validate the effectiveness and efficiency of our proposed method for edge-based human activity recognition.