Abstract
Detecting human falls is essential to ensuring public safety, particularly in public areas like transit terminals. This study provides a precise and effective real-time fall detection system by utilising pose estimation and deep learning-based object detection approaches. The system correctly detects falls in dynamic circumstances by combining MediaPipe's pose estimate for in-depth body posture analysis with the YOLOv8 model for human recognition. The paper provides a novel method that improves the system's scalability and robustness in real-world scenarios by utilising position landmarks and activity identification algorithms. To enable accurate fall detection and reduce false positives, the system also uses anomaly detection techniques. The system uses Twilio to send real-time warnings as soon as a fall is detected, send out video footage of the incident, and alert the appropriate authorities. The system is an excellent option for enhancing safety in sizable public areas because to its effectiveness, scalability, and privacy-preserving features. Metrics like accuracy, precision, recall, and F1-score are used in the study to assess the system's performance and show its usefulness. The system outperformed existing fall detection approaches, achieving 96.06 % accuracy and 100 % recall, confirming its robustness in real-world scenarios.