SrcSense: Robust WiFi-Based Motion Source Recognition via Signal-Informed Deep Learning

As WiFi has become a ubiquitous medium for communication, its role in sensing applications has expanded. However, the current WiFi sensing applications are limited by their assumption that any detected motion signifies human activity, overlooking the potential impact of nonhuman subjects. Existing attempts to recognize the interference from nonhuman motion impose stringent requirements regarding device positioning, data quality, environmental complexity, and nonhuman subject categories. In this study, we design a robust deep learning framework, SrcSense (“Source Sense”), to recognize the motion source with WiFi signals through the wall. SrcSense extracts environment-independent features from single-link WiFi. We investigate the performance of popular deep neural networks and explore the efficacy of transferring pretrained models to WiFi sensing tasks. We implement SrcSense and evaluate the performance in five real-world complex environments with commodity WiFi devices. With a challenging dataset considering large pets, diverse human activities and multiple subjects coexisting cases, SrcSense achieves an average validation accuracy of 95.84% across five distinct environments and an average testing accuracy of 91.71% in unseen environments without further model training or parameter tuning. By accumulating 20 s of WiFi data, SrcSense can achieve an elevated recognition accuracy of 99.77% with ResNet-50. These results underline the robustness of our approach and its readiness for integration into ubiquitous intelligent Internet of Things (IoT) systems and applications.