Hybrid Deep Learning for Human Fall Detection: A Synergistic Approach Using YOLOv8 and Time-Space Transformers

Falls are a major concern, particularly for elderly individuals and vulnerable populations, often leading to severe injuries if not detected promptly. Traditional sensor-based fall detection methods suffer from limitations such as discomfort, maintenance challenges, and susceptibility to false readings. To address these issues, we propose a hybrid deep learning-based system that leverages video-based techniques for accurate and efficient fall detection. Our approach integrates YOLOv8 for real-time human detection with Time-Space Transformers for temporal motion analysis, ensuring robust recognition of fall incidents while minimizing false positives and negatives. The proposed system is evaluated on two benchmark datasets, CUCAFall and DiverseFALL10500, which contain diverse and challenging environmental scenarios. Experimental results demonstrate that our model outperforms existing state-of-the-art methods, achieving a mean Average Precision (mAP) of 0.9955 on CUCAFall and 0.950 on DiverseFALL10500, along with F1-scores exceeding 0.998 for all classes in certain configurations. These results confirm the model's ability to distinguish fall events from normal activities accurately. Furthermore, the system maintains computational efficiency, making it suitable for real-time deployment in healthcare and smart home environments. By enhancing spatial and temporal awareness, our hybrid approach significantly improves fall detection reliability, ensuring timely intervention and enhancing safety for at-risk individuals.