Résumé
The increasing adoption of \mathbf{3 6 0}^{\circ} cameras has revolutionized the monitoring of large and complex spaces, providing comprehensive visual coverage that traditional cameras cannot match. However, this technology poses significant challenges for person detection and tracking, notably due to distortions caused by equirectangular projection. We present a lightweight method for tracking and fall detection from \mathbf{3 6 0}^{\circ} video streams captured by a Ricoh Theta X camera. Our approach combines dynamic generation of perspective views centered on targets, person detection using YOLOv8n, pose estimation via MediaPipe, and fall detection logic based on the skeleton aspect ratio. The entire pipeline operates in near real-time on standard CPU hardware, without requiring a GPU. Experimental results demonstrate the system's robustness in real-world conditions, paving the way for practical applications in constrained environments.