In this study, we introduce a robust hybrid system for Human Activity Recognition (HAR) using computer vision. By combining ResNet50 with a BLSTM network enhanced by an Attention mechanism, we aim to address challenges such as variations in human movement and lighting conditions. ResNet50 facilitates precise representations and mitigates information degradation concerns through its residual connections, ensuring efficient information flow. Its ability to extract deep features significantly enhances human activity recognition accuracy. Integrating BLSTM with ResNet50 improves sequence processing and prediction accuracy. The bidirectional data processing by BLSTM enables more accurate pattern recognition, refining the system’s overall accuracy. Our method demonstrates a substantial improvement in accuracy compared to traditional models, achieving a high Top5 accuracy rate of 96% in human activity classification. This hybrid approach exemplifies the potential of advanced techniques in handling the complexities of Human Activity Recognition in computer vision.