Enhanced Human Activity Recognition Using Inertial Sensor Data from Smart Wearables: A Neural Network Approach with Residual Connections

Human Activity Recognition (HAR) has gained significant interest in various research circles due to its wide-ranging applications, including patient monitoring, gaming, and education. While computer vision has traditionally played a key role in HAR, it encounters challenges like privacy concerns and environmental factors such as occlusion. To address these issues, inertial sensors like accelerometers and gyroscopic sensors have gained popularity, offering advantages such as cost-effectiveness and increased mobility. In this study, we propose a Convolutional Neural Network (CNN) architecture for HAR that incorporates residual connections. We rigorously evaluate our model using a freely available dataset WISDM (2011) from the wireless sensor data mining lab and compare its performance with state-of-the-art techniques. Our results indicate that our proposed model not only outperforms existing methods in terms of accuracy but also exhibits reduced complexity, requiring only 38,342 trainable parameters. Our HAR model achieves an impressive average accuracy of 98.32%, along with notable F1-score, Recall, and Precision values of 97.50%. These findings underscore the efficacy of our approach in advancing HAR technology while maintaining efficiency.