Human Activity Recognition-Based Fall Detection with GCRNN via Ninja Optimizer
S. Banushri, R Jagadeesha · 2025
Smart sensors have become popular due to the fast advancements in wireless communication networks, information, and communication technology. With the use of smart sensors, healthcare providers and patients can work together to intelligently and automatically track the daily activities of the elderly. The usage of smartphones and smart body sensors to track health and wellness is on the rise. When it comes to healthcare monitoring systems, wearable sensor technology is one of the most important advancements in smart sensor technology. Elderly fall detection is a critical task in healthcare and assisted living systems, aiming to prevent severe injuries and provide timely medical assistance. In this study, an advanced fall detection framework utilizing Human Activity Recognition (HAR) on the K-Fall dataset is proposed. Our approach employs a Graph Convolutional Recurrent Neural Network (GCRNN) to effectively capture both spatial and temporal dependencies in motion data collected from wearable sensors. The GCRNN model leverages Graph Convolutional Networks (GCN) to learn spatial relationships among sensor nodes while integrating recurrent layers to model sequential dependencies, improving the classification of fall-related activities. To further optimize model performance, incorporate the Ninja Optimizer, a novel optimization technique designed to enhance learning rate adaptation and parameter fine-tuning. The Ninja Optimizer aids in accelerating convergence, reducing overfitting, and improving generalization, leading to more reliable fall detection outcomes. Experimental results on the K-Fall dataset demonstrate that our GCRNN model, fine-tuned with the Ninja Optimizer, achieves superior performance compared to conventional machine learning classifiers and baseline deep learning models. The proposed method significantly improves key presentation metrics and score, making it a robust solution for real-time elderly fall detection. The findings of this research highlight the potential of GCRNN-based HAR models in enhancing fall detection accuracy and reliability.