A Light-weight Deep Human Activity Recognition Algorithm Using Multi-knowledge Distillation

Human Activity Recognition (HAR) is crucial in fields such as human-computer interaction, motion estimation, and intelligent transportation. Yet, attaining high accuracy in HAR, especially in scenarios limited by computing resources, poses a considerable challenge. This paper presents SMLDist (Stage-Memory-Logits Distillation), a framework designed to build highly customizable HAR models that achieve optimal performance under various resource constraints. SMLDist prioritizes frequency-related features in its distillation process to bolster HAR classification robustness. We also introduce an auto-search mechanism within heterogeneous classifiers to boost performance further. Our evaluation addresses the challenges of generalizing across users, sensor placements, and recognizing a wide array of activity modes. Models crafted with SMLDist, leveraging a teacher-based approach that achieves a 40%-50% reduction in operational expenditure, surpass the performance of existing state-of-the-art architectures. When assessing computational costs and energy consumption on the Jetson Xavier AGX platform, SMLDist-based models show strong economic and environmental sustainability advantages. Our results indicate that SMLDist effectively alleviates the performance degradation typically associated with limited computational resources, underscoring its significant theoretical and practical contributions to the field of HAR.