Answers are grounded in this paper and cite their sources.
Smartphone sensor-based human activity recognition is a practical sensing problem for mobile health, rehabilitation support, and daily behavior analysis. A useful model for this task must recognize temporal motion patterns from inertial sensor streams while remaining small enough for mobile inference. This paper proposes LightHARNet, a lightweight temporal convolutional neural network for classifying human activities from smartphone accelerometer and gyroscope windows. The method separates temporal filtering, channel mixing, and channel attention. Depthwise temporal convolution reduces trainable weights, while squeeze and excitation attention recalibrates accelerometer and gyroscope channels. Complete parameter and bounded modulation propositions are provided to explain why the design reduces model size while preserving stable feature scaling. Experiments are conducted on the public UCI Human Activity Recognition Using Smartphones dataset with the official subject independent test split. LightHARNet gives a stronger accuracy and macro F-score balance than compact neural baselines while keeping a small memory footprint. Paired significance testing indicates that the improvement is not explained by random paired disagreement. The class level analysis shows strong recognition of dynamic activities and laying, while the remaining errors mainly occur between sitting and standing. These results indicate that the proposed depthwise attention design gives a meaningful accuracy and efficiency balance for smartphone inertial sensing applications.
No takes yet. Share an insight, caveat, or question.
Yin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: