Falls among elderly individuals living alone are a major societal and health concern, particularly in smart home environments where safety and privacy must be preserved. This paper presents an artificial intelligence (AI)-driven, non-intrusive fall detection technique. A pose estimation model, combined with a TimeDistributed Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (BiLSTM) network, is presented. The UR Fall Detection (URFD) dataset, containing 30 fall events and 40 activities of daily living (ADLs), is used. The dataset is optimized by MoveNet, which extracts and defines 17 human body keypoints, which are processed to capture both spatial and temporal dependencies for accurate fall detection. The proposed hybrid architecture ensures real-time performance and minimizes false positives, as both Red-Green-Blue (RGB) and depth data were used in this study. Experimental results on the URFD Dataset demonstrate high accuracy, ranging from 95% to 100%, with minimal detection errors. This work provides an efficient and reliable solution for fall detection, laying the groundwork for enhanced elderly care in smart home environments.