Lightweight Low-Rank Spatio-Temporal Fusion for mmWave Radar-Based Aquatic Human Activity Recognition

Abstract

For aquatic human activity recognition, millimeter-wave radar is less sensitive to illumination changes, visual occlusion, and privacy concerns than camera-based sensing. However, water-surface activities introduce additional challenges, including multispectrum heterogeneity, coupled temporal-motion and range/cadence structures, and unstable cross-spectrum interactions under lightweight network constraints. To address this challenge, we propose the adaptive low-rank spatio-temporal fusion network (ALRST-Net), a lightweight radar-task-driven architecture for water-surface multispectrum representations. The network adopts a shared-specialized spectrum encoder to reduce redundant low-level extraction while preserving discriminative cues in range–time spectrogram, Doppler–time spectrogram, cadence–velocity diagram, and cadence–range diagram. It then reorganizes the multispectrum features into a temporal-motion branch and a range/cadence-structural branch, avoiding heavy recurrent neural network (RNN) or Transformer modeling. Finally, an adaptive stable low-rank residual fusion module introduces compact cross-spectrum interaction as a controlled residual correction, improving fusion stability and discriminability. Experimental results on the aquatic human activity recognition-I (AHAR-I) dataset show that ALRST-Net achieves 99.58% recognition accuracy with only 81.5-K parameters, 47.8-M floating-point operations (FLOPs), and 3.20-ms network inference latency. Compared with existing methods, ALRST-Net provides a favorable performance–efficiency tradeoff.