Enhancing Human Activity Recognition with IMU-based Motion Analysis

In this research, we delve into the domain of Human Activity Recognition (HAR) using deep learning techniques applied to data collected from a single wearable sensor. Our objective was to enhance the results obtained in previous studies and demonstrate the efficacy of automatic feature extraction methods. Employing an 8-layer Convolutional Neural Network (CNN) architecture, combined with innovative data augmentation techniques involving rotational and permutational transformations, we achieved notable improvements in classification accuracy. The augmented dataset, which incorporated both rotations and permutations, yielded the most significant enhancements, particularly benefiting underrepresented classes. Comparative analyses of various neural network architectures, including 1D and 2D convolutional networks, ResNets, and stacked denoising autoencoders, revealed that the optimal model featured ID and 2D convolutions capturing cross-sensor correlations. This model achieved an impressive F1-Score of 0.972. Our study emphasizes the potential for further exploration of recurrent neural networks and parameter fine-tuning. Furthermore, we underscore the critical role of data preprocessing and the need for a deep understanding of both machine learning and HAR concepts.