Lightweight mmWave radar human action recognition via knowledge distillation and physically enhanced representation
TL;DR
This paper introduces a physically enhanced input representation module that alleviates the representation limitations of sparse mmWave radar point clouds by explicitly incorporating inter-frame centroid differences and normalized timestamps as auxiliary motion and temporal cues and employs a high-accuracy dual-branch teacher model to guide a student model.
Abstract
Human activity recognition (HAR) based on millimeter-wave (mmWave) radar has attracted significant attention due to its advantages in non-contact sensing and privacy preservation. However, extracting robust fine-grained features from sparse and nonuniform mmWave point clouds remains challenging when relying solely on raw 3D coordinates, often leading to measurement uncertainty in complex dynamic scenarios. Furthermore, existing high-performance models incur high computational overhead, limiting their deployment on resource-constrained edge sensing devices. To address these challenges, this paper proposes a lightweight mmWave radar HAR framework based on knowledge distillation. First, we introduce a physically enhanced input representation module that alleviates the representation limitations of sparse mmWave radar point clouds by explicitly incorporating inter-frame centroid differences and normalized timestamps as auxiliary motion and temporal cues. Second, we employ a high-accuracy dual-branch teacher model to guide a student model that maintains macro-architectural consistency but utilizes simplified operators. A multi-granularity feature joint distillation strategy transfers representation capabilities to the student without incurring the teacher’s computational burden. Experimental results demonstrate that the student model achieves average recognition accuracies of 98.38% and 98.41% on the RadHAR and Pantomime datasets, respectively. Notably, the model requires only 0.27 M parameters and 0.029 GFLOPS.