An Efficient Human Walking vs. Running Activity Recognition Exploring Small-Scale Dynamic Neural Networks

This study presents a comprehensive examination of small-scale dynamic neural network (NN) models for human activity recognition using smartphone data. We investigate five dynamic NNs: Gamma Memory (GMNN), Lattice Ladder (LLNN), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRUNN), and Long Short-Term Memory (LSTMNN), on an open-source Kaggle dataset. Our research focuses on classifying human walking and running activities based on temporal features captured by accelerometer sensors. We explore the NN architecture and synapse configurations, assessing their performance in terms of accuracy and computational complexity. The results demonstrate significant potential for dynamic NN in human activity recognition tasks. In this study, an RNN achieved the highest accuracy, at 99.77%, for distinguishing between considered human activities. Moreover, a small-scale RNN, with 2 times fewer computational for binary additions, multiplications, and activation functions, performed remarkably well, achieving 99.72% accuracy.