Learning robust general radio signal detection using computer vision methods

Abstract

We introduce a new method for radio signal detection and localization within the time-frequency spectrum based on the use of convolutional neural networks for bounding box regression. Recently, this class of approach has surpassed human-level performance on computer vision benchmarks for object detection, but similar techniques have not yet been adopted for radio applications. We introduce the basic approach explain how labeled training data containing wideband spectrum annotated with masks and bounding boxes can be used to train a highly effective radio signal detector which achieves higher levels of contextual understanding and improved sensitivity performance when compared with more traditional nave energy thresholding based signal detection schemes. We extend prior work from the computer vision field, employing a variation of the You Only Look Once (YOLO) architecture which is a fast and accurate visual object detector. Results are shown from illustrating the effectiveness from our entry into the DARPA Battle-of-the-ModRecs competition and over the air datasets.