HAR (human activity recognition) is a pivotal task in computer vision with diverse applications ranging from surveillance to healthcare. The objective of our project is to develop a system for the classification of human activities using vision-based techniques. We specifically use video frames and feed them into 2D Convolutional Neural Networks (CNNs) following Bidirectional Long Short-Term Memory (Bi-LSTM) networks. We aim to address the challenges of human activity recognition by designing a hybrid architecture capable of effectively capturing both spatial and temporal features from video data. Our framework aims to capture both spatial and temporal dependencies in activity sequences extracted from video data. Initially, a 2D CNN is employed to extract hierarchical features from video frames, effectively encoding spatial information. Subsequently, the extracted features are fed into Bidirectional LSTM layers, capturing temporal dynamics and dependencies among sequential frames. This combination allows for robust feature representation and contextual understanding of human activities over time. Experiments conducted on a subset of five categories from the UCF50 dataset demonstrate the efficacy of the proposed approach, achieving an accuracy of 99.22% in recognizing activities.