As the global population ages, falls represent a significant health risk for the elderly. This study aims to propose a high-performance, end-to-end fall detection model designed to serve as a core component for practical, vision-based monitoring systems in real-world environments. We introduce an optimized Transformer-based architecture that detects falls directly from raw RGB video streams, thereby obviating the need for extensive data pre-processing or wearable sensors. The model's generalizability and effectiveness were rigorously evaluated using the AIHUB dataset, which encompasses diverse scenarios, including varied locations and the use of assistive devices. The proposed model achieved an accuracy of 96.5% and an F1-score of 93.2%, demonstrating robust performance even under challenging conditions. The implications of this work are threefold. First, the system can be deployed on existing camera infrastructure, offering a scalable and cost-effective solution for continuous monitoring. Second, by enabling automated monitoring in residential and care facilities, it has the potential to reduce caregiving costs and address service gaps. Third, the non-intrusive nature of the system preserves the privacy and autonomy of individuals. This research contributes significantly to the development of technology-driven safety nets for vulnerable populations and offers practical considerations for senior welfare policies and the design of smart care infrastructure.