Transparent and Resilient Activity Recognition via Attention-Based Distributed Radar Sensing

TL;DR

This work tackles scalable, interpretable activity recognition with distributed radar by proposing an end-to-end framework that processes raw radar data at each node using a lightweight 2D CNN, followed by a self-attention fusion block to model inter-node dependencies. Training combines a cross-entropy loss with a supervised contrastive objective to boost discriminability, while the attention mechanism provides inherent interpretability through node-level weights $\alpha$ and the derived importance $I(j)$. Empirical results on real-world UWB radar data show higher accuracy than a CNN–RNN baseline, with substantially reduced model size (3.4×) and faster inference ($0.013$ s per sample), and encoder-based feature transmission yields superior compression over traditional downsampling. The proposed method enables transparent, efficient distributed sensing suitable for real-time ambient intelligence in sensitive or industrial environments.

Abstract

Distributed radar sensors enable robust human activity recognition. However, scaling the number of coordinated nodes introduces challenges in feature extraction from large datasets, and transparent data fusion. We propose an end-to-end framework that operates directly on raw radar data. Each radar node employs a lightweight 2D Convolutional Neural Network (CNN) to extract local features. A self-attention fusion block then models inter-node relationships and performs adaptive information fusion. Local feature extraction reduces the input dimensionality by up to 480x. This significantly lowers communication overhead and latency. The attention mechanism provides inherent interpretability by quantifying the contribution of each radar node. A hybrid supervised contrastive loss further improves feature separability, especially for fine-grained and imbalanced activity classes. Experiments on real-world distributed Ultra Wide Band (UWB) radar data demonstrate that the proposed method reduces model complexity by 70.8\%, while achieving higher average accuracy than baseline approaches. Overall, the framework enables transparent, efficient, and low-overhead distributed radar sensing.

Figures (5)

• Figure 1: The proposed distributed radar recognition model consists of a 2D CNN block for local feature extraction, a self-attention block for adaptive and interpretable fusion, and a final classification block.
• Figure 2: Confusion matrices (in percentage) for (a) the proposed model and (b) the baseline CNN-RNNzhu2022continuous.
• Figure 3: t-SNE visualization of (a) raw input features, (b) fused representation, and (c) output of the final dense layer.
• Figure 4: Encoder-based feature compression versus conventional downsampling, evaluated under varying SNR conditions.
• Figure 5: Ablation study of node importance, with the bar plot showing the distribution of the most important nodes across the radars.