Graph based Hierarchical Temporal Network with hybrid optimization models for accurate Fall Detection

Abstract

This research addresses the critical health risk of falls in the elderly through the development of an innovative fall recognition system. The proposed methodology incorporates data augmentation, joint angle extraction, velocity, acceleration, and spatial feature analysis. RNN embeddings capture temporal dependencies, while a hybrid optimization model Minkowski Distance in Locust Movement with Aquila (MDLMA), combining Artificial Locust Swarm Optimization (ALSO) and Aquila Optimizer (AO) facilitates effective feature selection. The Graph-Based Hierarchical Temporal Network (GHTNet) integrates graph convolutional networks, bidirectional Long Short-Term Memory (Bi-LSTM) layers with Residual Blocks, and Attention-Weighted Fusion for adaptive spatial and bidirectional temporal modelling. The system culminates in a fully connected layer for precise fall classification, ensuring robustness and efficiency in detection.