.png){kind=icon}
.png){kind=icon}
Return
Real-time human activity recognition on edge microcontrollers: dynamic hierarchical inference with multi-spectral sensor fusion
.png){kind=icon}
PRE
AI
B
Z
L
Y
DOI：10.1088/1361-6501/ae6469.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The demand for accurate on-device pattern recognition in edge applications intensifies, yet existing approaches struggle to reconcile accuracy with computational constraints. To address this critical challenge, a resource-aware hierarchical network based on multi-spectral fusion and interpretable modules, namely the hierarchical parallel pseudo-image enhanced fusion network (HPPI-Net), is proposed to enable real-time, on-device human activity recognition (HAR) tasks. Deployed on ARM Cortex-M4 MCU for low-power real-time inference, HPPI-Net achieves 96.70% accuracy while utilizing only 22.3 KiB of RAM and 439.5 KiB of ROM after optimization. HPPI-Net employs a two-layer architecture: the first layer extracts preliminary features using fast Fourier transform (FFT) spectrograms, while the second layer selectively activates either a dedicated module for stationary activity recognition or a parallel LSTM-MobileNet network (PLMN) for dynamic states. PLMN fuses FFT, Wavelet, and Gabor spectrograms through three parallel LSTM encoders and refines the concatenated features with efficient channel attention and depthwise separable convolution, thereby offering channel-level interpretability while substantially reducing multiply-accumulate operations. Compared to MobileNetV3, HPPI-Net notably increases accuracy by 1.22% and significantly reduces RAM usage by 71.2% and ROM usage by 42.1%. These results demonstrate that HPPI-Net realizes a favorable accuracy-efficiency trade-off and provides explainable predictions, establishing a practical solution for wearable, industrial, and smart-home HAR on memory-constrained edge platforms.
Keywords:
deep learning
edge computing
multi-feature fusion
explainable AI
Journal
IF:
3.4
Papers:
2.6K
Citations:
2.3W
Organization
Cited Papers