A Lightweight Mobile Temporal Convolution Network for Multi-Location Human Activity Recognition based on Wi-Fi

Wi-Fi-based human activity recognition has been widely adopted in the field of the Internet of Things. Although recent works have made great progress for human activity recognition at multiple locations, most of them rely on high-resolution data, adequate training samples and large-scale networks to model human activities, which ignores serial floating-point operations on CPU-driven devices and memory consumption limitations. Therefore, to address above issue, this paper proposes a Lightweight Mobile Temporal Convolution Network (LM-TCN). On the one hand, the proposed approach uses the fully 1-D convolution framework to provide time-shift invariant inductive bias. On the other hand, the combination of invert bottleneck and gated mechanism optimizes the computational load of the conventional residual structure to prevent overfitting under few training samples. Experimental results show that the average accuracy of the proposed LM-TCN is 95.2% across all 24 predefined locations, which is 2.9% higher than the baseline TCN while the calculation cost is reduced to 6% of TCN. It is worth noting that only 10 samples and 15 subcarriers for each activity at each location are used for training.