Accidental falls and their consequences are critical concerns for elderly people. Fatal injuries, when delayed in treatment, can lead to severe outcomes. Fall detection systems are crucial for the timely treatment of such injuries. Although sensor-based fall detection approaches are effective, video-based approaches are more useful because they assist in analyzing the fall scene and identifying the cause of the fall. However, privacy preservation is a major concern in video-based fall detection. The proposed system introduces a privacy-preserving mechanism that masks the identified human with a silhouette. A custom dataset, including 80 activities of daily living and 70 fall activities, is introduced. An LSTM variational autoencoder architecture is designed with a gradient clipping mechanism and a smooth variant of Adaptive Moment Estimation with Stochastic Gradient Descent (AMSGrad) optimizer to enhance the accuracy of fall detection. The reconstruction error between normal and fall activities is clearly identified with the help of a dynamic threshold. This results in a system performance that achieves accuracy, precision, and sensitivity of 99%, 97%, and 99%, respectively.