This paper outlines a study aimed at enhancing elderly care through an intelligent video surveillance system that leverages deep learning for detecting mobility anomalies, specifically near-falls. Identifying near-falls is essential because people who experience frequent near-falls while carrying out their daily activities are at risk of future falls. We successfully developed an autoencoder to detect these anomalies, particularly near-falls, by identifying high reconstruction errors throughout five consecutive frames. To extract a person's skeleton, we utilized MoveNet and narrowed it down to only seven keypoints. We then used a set of 20 features, encompassing joint positions, velocities, accelerations, angles, and angular accelerations, to train the model. Our model was tested on 100 videos of simulated daily activities recorded in an apartment laboratory, where 50 of them contained a near-fall. Results show that our model can successfully detect near-falls with 90% sensitivity, specificity, and accuracy, highlighting its potential to enhance elderly care in their living environments.