Hybrid Model and Data Driven WiFi Sensing Using Data-Traffic-Informed Augmentation

This paper considers a particular use case for car density estimation by reusing WiFi downlink (DL) signals of opportunity to demonstrate how sensing performance can be maximized through the exploration of signal models and data-traffic-informed augmentation. Despite the success of reusing WiFi signals opportunistically for various sensing tasks, the optimal utilization of raw WiFi waveforms for outdoor target sensing has received limited attention. The car density inference is based on the monitoring of the average power of the signals reflected by the cars. However, a raw WiFi signal is a mix of time-varying components, including DL and uplink (UL) portions, with the UL portion being user-dependent, highly random, and therefore unreliable. Additionally, the DL portion fluctuates in average power, mainly due to the randomness of packet lengths and their arrival rates. Based on the DL traffic structure, the data augmentation mitigates disruptive factors by applying DL signal extraction, packet length equalization, and packet arrival rate normalization. Car density estimation uses a hybrid convolutional neural network (CNN) and support vector machine (SVM) for classification, combined with the least squares estimate (LSE) for fine-grained accuracy. Two physical models are considered in the process: the structure of the mixed UL-DL waveform is used to identify the DL portion, and a linear relation between car density and classification similarity scores are assumed. The results demonstrate that the proposed scheme significantly improves the estimation accuracy.