Lightweight Deep Learning Framework for Human Activity Recognition: A MobileNet-LSTM Approach with WHOA Optimization

Medical informatics, HCI, surveillance, and task monitoring systems use computer vision for human activity classification in novel ways. To aid doctors in medication reactions and diagnosis, patients' activities must be analyzed and categorized. Machine learning and soft computational algorithms are used for human activity recognition from movies and photographs, but more advanced computer vision approaches are being investigated. This study proposed automated HAR classification using MobileNet V2 and LSTM. MobileNet V2 beat its predecessors in efficiency and accuracy testing on low-power computing devices. For different action recognition scenarios, this study uses RGB, depth, and skeletal joint data from UTKinect-Action3D and CMU Mocap3D. A systematic preprocessing pipeline includes frame segmentation, temporal sampling, optical flow computation, Gaussian blur, color jittering, and random flips to improve model performance. Interpolation and mean imputation provide dataset completeness, while quality control procedures improve dataset dependability. MobileNet, intended for low-computation contexts, captures action sequence temporal dependencies with LSTM. Fine-tuning model hyperparameters with the Wildebeest Herd Optimization Algorithm (WHOA) improves efficiency and accuracy. Experimental results show the proposed model can achieve excellent recognition performance while being computationally efficient, making it suitable for real-time HAR applications.