A Lightweight Model for Human Activity Recognition Based on Two-Level Classifier and Compact CNN Model

Wearable devices for human activity recognition (HAR) provide continuous remote health monitoring, which can offer many benefits for patients in rehabilitation, the elderly, and even the population in general. Those devices handle a big amount of data and run complex algorithms for delivering information with precision to the user. However, to achieve a satisfactory performance in terms of energy consumption, real-time response and privacy, the solution needs to be efficient and suitable for running in resource constrained devices. Many HAR proposals use deep-learning approaches for achieving almost perfect performance, nonetheless those techniques are not convenient for embedded solutions. To meet the demands of high precision and low power, more efficient strategies can be developed and optimization techniques can be applied to the usual deep learning approaches. We present in this paper a lightweight model for HAR based on two-level classifier and using a compact, pruned and quantized convolutional neural network. We obtained accuracy and F1-score above 90% with an extremely lightweight solution and intended for use in embedded systems.