<p>Elderly human activity recognition is a critical component of smart healthcare systems, aiming to ensure safety and well-being by accurately monitoring and analyzing senior individuals’ daily movements and activities. Traditional methods may struggle with capturing the temporal continuity and spatial context of activities simultaneously, which is crucial for understanding the complex, nuanced motions characteristic of elderly behavior. To reduce misclassification among activities with similar spatial–temporal patterns, the model employs gated temporal memory and localization to emphasize discriminative motion regions across consecutive frames. In contrast, our proposed system is designed to address these challenges by integrating spatial and temporal feature extraction in a unified framework, enhancing the robustness of activity recognition. The study introduces a ConvLSTM2D model with localization for recognizing elderly human activities within smart home environments, demonstrating state-of-the-art performance on the ETRI-Activity3D dataset. The model’s unique architecture, combining convolutional neural networks with long short-term memory networks, allows it to capture and process spatial-temporal data effectively, significantly improving activity recognition accuracy. With preprocessing steps, including data normalization and Non-max Suppression, the model achieves up to 94% accuracy for Tooth brushing, 93% for Vacuuming, and over 90% F1-score for several activities. These results underscore the model’s capability to discern complex activity patterns with high precision and minimal error. The promising outcomes of this research suggest the model’s suitability for deployment in real-world elderly care applications, providing a robust tool for enhancing quality of life and safety in assisted living settings.</p>