Multiscale Residual Weighted Classification Network for Human Activity Recognition in Microwave Radar.

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Human activity recognition (HAR) using radar sensors is crucial for healthcare and smart homes.
• Labeling large radar datasets is time-consuming and hinders model performance.
• Existing models struggle with classification accuracy due to insufficient labeled data.

Purpose of the Study:

• To propose a novel multiscale residual weighted classification network (MRW-CN) for efficient HAR.
• To address the challenge of limited labeled data in radar HAR.
• To improve classification accuracy in radar-based activity recognition.

Main Methods:

• Utilized a multiscale residual weighted (MRW) image encoder with contrastive learning for feature extraction.
• Employed large, medium, and small-scale residual networks for global, texture, and semantic information.
• Incorporated a time-channel weighting mechanism for enhanced feature extraction.
• Pre-trained the MRW encoder, froze parameters, and fine-tuned a classifier with limited labeled data.

Main Results:

• Achieved a classification accuracy of 96.9% on a newly constructed dataset of eight dangerous activities.
• Demonstrated state-of-the-art performance in radar-based human activity recognition.
• Ablation studies confirmed the effectiveness of multi-scale kernels and time-channel weighting.

Conclusions:

• The proposed MRW-CN model effectively addresses the limitations of insufficient labeled data in radar HAR.
• The multiscale approach and time-channel weighting significantly enhance feature representation and classification accuracy.
• This method offers a promising solution for reliable human activity recognition in real-world applications.