Efficient Voxel-Based mmWave Radar HAR With Early Spatio–Temporal Fusion and a Compact 3-D–2-D Hybrid Network

Millimeter-wave (mmWave) radar has emerged as a powerful sensing modality for human activity recognition (HAR) owing to its capability to capture 3D point cloud sequences without privacy concerns. However, effectively modeling the sparse, irregular, and non-uniform nature of radar data remains a major challenge. Existing approaches often rely on highly complex network architectures to improve accuracy, which leads to excessive computational overhead and poor scalability. To overcome these limitations, this paper proposes a compact hybrid feature extraction network for voxelized radar point cloud classification, which performs early-stage spatio-temporal fusion and is termed STFusionNet (Spatial-Temporal Fusion Network). The STFusionNet comprises (i) a lightweight 3D convolutional front-end, which treats consecutive temporal frames as input channels to encode motion dynamics into a compact volumetric representation and further aggregates features along the depth axis; (ii) a minimalist 2D convolutional backbone after a Depth-to-Channel Folding operation, which captures spatial features with minimal computational cost. Extensive experiments on the public MMActivity and MiliPoint datasets demonstrate that our model achieves competitive accuracy (e.g., 92.20% on MMActivity and 72.86% on MiliPoint) with only about 50K parameters and 46 MMac, outperforming or matching representative spatio-temporal baselines under a much lower computational budget. Extensive ablation experiments verify the contribution of key components, while statistical significance tests confirm the reliability of the performance improvements. These results confirm that STFusionNet offers a robust, efficient, and generalizable solution for mmWave radar-based HAR in Internet of Things (IoT) applications.