A vision transformer with recurrent neural network-based fall activity recognition system for disabled persons in...

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Falls pose significant risks to autonomy, health, and life, especially for the elderly and disabled.
• Reliable fall detection is critical for healthcare and robotics to mitigate injuries and adverse effects.
• Deep learning (DL) and computer vision have advanced fall detection accuracy, overcoming limitations of traditional methods.

Purpose of the Study:

• To enhance fall detection and classification for individuals with disabilities in smart IoT environments.
• To introduce a novel Vision Transformer and Self-Attention Mechanism with Recurrent Neural Network-Based Fall Activity Recognition System (VTSAMRNN-FARS).
• To improve the accuracy and efficiency of fall recognition systems.

Main Methods:

• Image pre-processing using the bilateral filtering (BF) model to reduce noise.
• Feature extraction via the Vision Transformer (ViT) model for efficient data representation.
• Fall activity detection and classification using a bidirectional gated recurrent unit with a self-attention mechanism (BiGRU-SAM) model.
• Hyperparameter optimization of the BiGRU-SAM model with the enhanced wombat optimization algorithm (EWOA).

Main Results:

• The VTSAMRNN-FARS methodology demonstrated high performance in fall detection and classification.
• Simulation analysis was conducted using the UR_Fall_Dataset_Subset.
• The VTSAMRNN-FARS method achieved 99.67% effectiveness, outperforming existing fall detection models.

Conclusions:

• The VTSAMRNN-FARS system offers a significant advancement in fall detection technology.
• The proposed method shows high potential for application in smart IoT environments for vulnerable populations.
• The integration of ViT, BiGRU-SAM, and EWOA provides a robust and accurate fall recognition solution.