A Classification Algorithm for Blind UAV Detection in Wideband RF Systems

We consider the problem of detecting and localizing the time-frequency span of unmanned aerial vehicles (UAVs) which transmit Wi-Fi-type [1] signals in the ISM band. Interference is assumed to be present ubiquitously in the form of Bluetooth and microwave oven signals. We firstly expose that simple edge detection algorithms based on morphological processing, which are conventionally used to detect signals in spectrogram matrices, do not provide a scalable performance beyond a few RF captures because they suffer from limited hyperparameter generalizability. To overcome this limitation, we propose a support vector machine (SVM) classification algorithm, wherein, we run a sliding detection window across the spectrogram and extract the histogram of ordered gradient (HOG) features. The extracted feature vectors are used to classify the windows as positive or negative, depending on whether a signal of interest is present or not. For experimental evaluation, we generate a synthetic RF dataset containing 150 training captures and 50 test captures, each of duration 90ms, bandwidth 56MHz, and containing about 4.5 Wi-Fi signals on average. The proposed SVM algorithm achieves a mean average precision (mAP) of up to 0.4835, which is significantly higher than the state-of-the-art morphological processing algorithm [11] which only achieves an mAP of 0.3676.