Radio-based Object Detection using Deep Learning

We present a novel radio-signal based, short-range object detection system for controlled environments, which substitutes complex signal processing and expensive hardware with deep learning networks to detect patterns from low-quality, inexpensive sensors. The models are trained on forms of I/Q data, sampled via software defined radios. Our system operates in the less crowded low-frequency range of 433MHz in contrast to existing RF-based sensing methods, allowing us to use cheap, off-the-shelf hardware. We demonstrate a proof-of-concept prototype, in a controlled environment, which can be scaled to build more complex detection systems relying on higher frequencies. Our system can handle occlusions, and, since it does not use visual data, is not hampered by bad lighting. The core of our system is a VGG-16 based CNN architecture trained on spectrograms, obtained by transforming I/Q data via Fourier methods. We achieve an accuracy of 0.96 on a binary classification task of detecting the presence or absence of an object in an enclosed space. Our prototype demonstrates that convolutional networks can learn features important enough from spectrograms that enable it to distinguish the presence of objects, thereby eliminating the need of sophisticated signal processing methods to do the same.