VBCNet: A Hybird Network for Human Activity Recognition.

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• Human activity recognition (HAR) is crucial for various applications.
• Existing HAR methods often require dedicated sensors, raising privacy concerns and deployment costs.
• Channel State Information (CSI) from Wi-Fi signals offers a privacy-preserving alternative for HAR.

Purpose of the Study:

• To propose VBCNet, a hybrid network architecture for effective human activity posture identification using Wi-Fi CSI.
• To enhance temporal feature extraction and local/multi-scale feature processing for improved HAR accuracy.
• To demonstrate the robustness of VBCNet in complex scenarios without additional device deployment.

Main Methods:

• Extraction and preprocessing of CSI sequences from Wi-Fi signals.
• Implementation of a hybrid network (VBCNet) incorporating Long Short-Term Memory (LSTM) for temporal features.
• Utilization of a convolutional feed-forward network for enhanced local and multi-scale feature processing.
• Classification of extracted features to identify human behaviors.

Main Results:

• VBCNet achieved high accuracy on benchmark datasets: 98.65% on UT-HAR and 77.92% on Widar3.0.
• The hybrid architecture effectively captures temporal and multi-scale features from CSI data.
• The model demonstrates strong performance in complex environments, validating its robustness.

Conclusions:

• VBCNet is a highly effective and robust method for human activity recognition using Wi-Fi CSI.
• The proposed approach minimizes privacy leakage and avoids the need for extra hardware.
• VBCNet shows significant potential for practical HAR applications in diverse settings.