Convolutional Neural Network Bootstrapped by Dynamic Segmentation and Stigmergy-Based Encoding for Real-Time Human...

Area of Science:

• Computer Science
• Artificial Intelligence
• Smart Environments

Background:

• Deep learning (DL) is increasingly used for human activity recognition (HAR) in smart buildings.
• Existing Convolutional Neural Network (CNN) methods often analyze pre-segmented data, limiting real-time applications.
• Real-time HAR is crucial for many smart building functionalities.

Purpose of the Study:

• To propose an online human activity recognition framework for real-time sensor data processing.
• To address the limitations of traditional methods in handling continuous data streams.
• To enable accurate activity classification in dynamic smart building environments.

Main Methods:

• The framework employs real-time dynamic segmentation to identify activity segments.
• Stigmergy-based encoding, using directed weighted networks (DWN), transforms spatio-temporal tracks into a multi-dimensional format suitable for CNNs.
• A 2D Convolutional Neural Network (CNN2D) classifies activities based on the encoded DWN representations.

Main Results:

• The proposed online HAR framework successfully processes streaming sensor data.
• The integration of dynamic segmentation and stigmergy-based encoding enables real-time feature extraction.
• The CNN2D effectively classifies human activities using the generated DWN features.

Conclusions:

• The developed framework provides an effective solution for online human activity recognition in smart buildings.
• This approach enhances the capability of smart environments to understand and respond to human actions in real-time.
• The methodology demonstrates the potential of DL for real-time HAR applications using continuous sensor streams.