Lightweight Deep Neural Network Embedded with Stochastic Variational Inference Loss Function for Fast Detection of...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Lightweight neural network models are crucial for efficient real-time applications.
• Human posture identification requires accurate and fast inference.
• Existing models often face challenges with model size and computational complexity.

Purpose of the Study:

• To develop a novel scheme for lightweight neural network models.
• To improve inference speed and reduce model size.
• To apply the technique for fast human posture identification and fall detection.

Main Methods:

• Fusing object detection with stochastic variational inference.
• Utilizing an integer-arithmetic-only algorithm and feature pyramid network.
• Employing a self-attention mechanism for sequential frame feature extraction.
• Applying Bayesian neural networks and Gaussian mixture models for classification.

Main Results:

• Achieved superior mean average precision (34.6) compared to ResNet (32.5).
• Demonstrated significantly faster inference speed (27 ms vs. 48 ms).
• Reduced model size substantially (46.2 MB vs. 227.8 MB).
• Enabled advance alerts for suspected human falling events (0.66 s).

Conclusions:

• The proposed lightweight neural network scheme offers improved performance in human posture identification.
• The model's efficiency in speed and size makes it suitable for real-time applications.
• The technique shows promise for early fall detection systems.