Human Activity Recognition (HAR) has emerged as a critical area of research, with applications spanning video surveillance, healthcare, sports analytics, and human-computer interaction. Deep learning was used in this work because it provides superior feature extraction and temporal modeling compared to hand-crafted features. This work employs the Inflated 3D Convolutional Network (I3D) to capture low-level spatiotemporal features from video frames. Subsequently, various configurations of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks are utilized to model high-level temporal dependencies, addressing challenges such as vanishing and exploding gradients. The experimental evaluation is conducted on benchmark datasets: UCF-101 and HMDB-51. The results demonstrate that bidirectional variants of LSTM and GRU outperform their unidirectional counterparts by effectively capturing both forward and backward temporal dependencies, leading to enhanced recognition accuracy in Human Activity Recognition tasks. Moreover, GRU outperforms LSTM, indicating its efficiency in modeling temporal dependencies with fewer parameters. Notably, the I3D model combined with Bi-GRU achieves the best results across both datasets, demonstrating its effectiveness in capturing spatiotemporal features, with the highest possible top-5 accuracy of 99.8% on the UCF-101 dataset and 92.0% on the HMDB-51 dataset.