YOLOv5 Meets ISOA: A Bio-Inspired Optimization Framework for Fall Detection and Localization

Fall detection using deep learning leverages neural networks to accurately identify and classify fall events from sensor or video data. By learning complex spatial and temporal patterns, models such as CNNs, LSTMs, or hybrid architectures can distinguish among normal activities and falls in real-time. This technology enhances safety in healthcare and eldercare settings by enabling timely alerts and interventions. Falls are a critical health concern, especially among the elderly, with timely detection playing a vital role in reducing injuries and fatalities. This study proposes an optimized deep learning framework for fall detection and segmentation using the KFall dataset. The approach integrates YOLOv5, a state-of-the-art real-time object detection model, for both classification (fall vs. non-fall) and bounding box-based segmentation. To enhance detection accuracy and localization precision, the model’s hyperparameters are fine-tuned using the Improved Snake Optimization Algorithm (ISOA), a bio-inspired metaheuristic optimization technique. The proposed method outperforms the default YOLOv5 across multiple performance metrics. Classification results show significant gains, with accuracy improving from 92.84% to 96.13% and AUC-ROC increasing from 94.30% to 97.10%. Similarly, segmentation metrics such as [email protected] and IoU also show noticeable enhancements. Despite a slight increase in inference time (from 21.3 ms to 23.1 ms), the model maintains real-time capability with minimal computational overhead. The proposed ISOA-tuned YOLOv5 framework demonstrates robust detection and efficient segmentation, providing a reliable foundation for fall detection systems in real-world smart surveillance and healthcare applications.