IMRadar: Bidirectional Velocity Mamba for Contactless Human Behavior Sensing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Intelligent human behavior sensing using channel state information (CSI) is crucial for contactless health monitoring.
• Existing methods struggle with global context perception, high computational costs, and unidirectional feature extraction.

Purpose of the Study:

• To develop a novel bidirectional velocity Mamba (BVMamba) model for enhanced human behavior sensing.
• To construct an intelligent behavior sensing system, IMRadar, for robust contactless perception.

Main Methods:

• Utilizing velocity information from CSI data for motion state characterization.
• Employing the BVMamba model, comprising forward (FVMamba), reverse (RVMamba), and fusion (FUBlock) components, for deep feature extraction.
• Analyzing global deep behavioral features from both forward and reverse directions to capture complex dynamic characteristics.

Main Results:

• The IMRadar system demonstrated excellent recognition performance across multiple datasets (ARIL, Widar, IM-HAR).
• Achieved accuracy rates exceeding 98% on all tested datasets, highlighting its effectiveness.
• The BVMamba model successfully captured dynamic characteristics of complex human behaviors.

Conclusions:

• IMRadar provides an efficient and robust solution for non-contact behavior perception technology.
• The proposed BVMamba model overcomes limitations of existing feature extraction networks.
• This approach significantly advances the field of intelligent sensing for health monitoring and beyond.