RadarSSM: A Lightweight Spatiotemporal State Space Network for Efficient Radar-Based Human Activity Recognition.

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• Millimeter-wave radar is increasingly used for Human Activity Recognition (HAR) due to its privacy preservation and environmental robustness.
• Processing high-dimensional 4D radar data for fast inference on resource-constrained edge devices remains a challenge.
• Existing 3D Convolutional Neural Networks and Transformer models suffer from high parameter overhead and computational complexity.

Purpose of the Study:

• To introduce RadarSSM, a lightweight spatiotemporal hybrid network for efficient radar-based HAR.
• To address the limitations of current models in terms of computational complexity and parameter overhead for edge applications.
• To enable accurate and efficient human activity recognition on low-resource edge hardware using radar sensing.

Main Methods:

• Developed RadarSSM, a hybrid network explicitly separating spatial feature extraction and temporal dependency modeling.
• Employed depthwise separable 3D convolutions for efficient spatial feature extraction from voxelized radar data.
• Utilized a bidirectional State Space Model for capturing long-range temporal dependencies with linear time complexity O(T).

Main Results:

• RadarSSM demonstrated accuracy competitive with state-of-the-art methods on public radar HAR datasets.
• The proposed network significantly reduced parameter count and computational cost compared to convolutional baselines.
• Achieved efficient radar sensing suitable for edge hardware applications.

Conclusions:

• RadarSSM effectively addresses the computational challenges of radar-based HAR on edge devices.
• The network's lightweight design and efficient temporal modeling make it suitable for real-world, privacy-preserving sensing applications.
• Validates the potential of spatiotemporal hybrid networks for resource-constrained radar sensing tasks.