Human activity recognition (HAR) in video data is critical for a wide range of applications and has attracted extensive scholarly interest. In this work, a novel deep learning model, iPose-H, has been developed to enhance posture detection in video-based HAR via an innovative tuning methodology. The model mathematically integrates optimized feature selection and systematic hyperparameter tuning to robustly capture the spatiotemporal dynamics inherent in video frames. Experimental evaluations on benchmark datasets reveal that iPose-H achieves a validation accuracy of 82.11% on HMDB-1, significantly outperforming the ConvLSTM model (63.01%). On HMDB-2, iPose-H attains an accuracy of 85.71%, yielding improvements of 5% over LRCN (80.22%) and 70% over MobileViT (13.95%). Quantitative analyses underscore the model’s capacity to extract salient features and effectively model temporal variations, while a comparative assessment of deep convolutional neural networks and vision transformers elucidates distinct advantages in feature extraction and classification. Rigorous statistical validation substantiates these findings, establishing iPose-H as a potential benchmark for future research in video-based HAR. The trained weights and algorithm are publicly available at https://github.com/ernamangoyal/DNN-HPEAR , thereby facilitating further investigation and replication by the research community.