Efficient Human Activity Classification Via 1D Capsule Networks and Chaotic Optimization
Rajanikanth Aluvalu, K Reshma Reddy, Anitha Patil, V. Uma Maheswari, Gangula Vijay Kumar, K. S. Raghavendra Reddy · 2025
Research into Human Activity Recognition (HAR) is becoming essential for healthcare monitoring organisations and pervasive computing. HAR uses mobile devices or wearable sensors to automatically detect and analyse physical activities. Supporting early identification of health disorders, monitoring patient behaviours and identifying odd motions are all made easier with its guidance. Because it allows for continuous, realtime monitoring even when patients are not in a clinical setting, HAR improves personalised care. This article presents a new method for HAR classification using the UCI-HAR dataset. It takes advantage of hyperparameter tuning with a proposed 1D capsule neural network that is augmented by the Chaotic Satin Bowerbird Optimisation Algorithm (CSBOA). When it comes to dynamic feature extraction and appropriate parameter selection, traditional models such as CNN-LSTM, conventional Capsule Networks, and Long Short-Term Memory (LSTM) frequently fail, resulting in subpar classification results. On the other hand, the suggested approach makes use of capsules' dynamic routing capabilities to gather hierarchical spatial-temporal information, and CSBOA efficiently finds the optimal hyperparameters, guaranteeing better model convergence and application. An impressive$\mathbf{9 8. 4 9 \%}$accuracy,$\mathbf{9 9. 9 0 \%}$precision,$\mathbf{9 9. 5 5 \%}$recall,$\mathbf{9 9. 7 2 \%}$F1-score, and 99.78% AUC were achieved by the proposed model. On the UCI-HAR benchmark, these metrics considerably surpass state-of-the-art deep learning approaches. This research highlights the promising future of human activity recognition systems that combine intelligent optimisation techniques with capsule-based designs.