This paper presents a comprehensive approach to real-time fall detection using advanced Transformer-based architectures tailored for deployment on resource-constrained devices. Our dataset, collected over four months using the WitMotion BWT61CL IMU and complemented by smartphone video recordings, provides a rich, multi-modal source for modeling fall and non-fall events in diverse environments. Our work focuses on the deployment and performance evaluation of three Transformer-based models—Standard Transformer, Performer, and Linformer—each optimized for latency and accuracy in processing time-series accelerometer data. Rigorous data preprocessing, including noise filtering and feature extraction, was applied to enhance signal quality. We evaluate the models on a dataset comprising 403 samples, achieving a peak accuracy of 98% with the Standard Transformer, and competitive results of 96% with the Performer and Linformer. The Performer model emerges as the most efficient latency, achieving an average response time of 34ms, while the Standard Transformer and Linformer require 350ms and 110ms, respectively. This efficiency, combined with high sensitivity and specificity, underscores the Performer model’s suitability for real-time embedded systems. Our findings demonstrate that advanced Transformer models, with optimized hyperparameters and efficient architectures, can deliver accurate, low-latency fall detection solutions, paving the way for enhanced safety in applications requiring real-time monitoring on compact hardware.