A team built and tested a compact radar system that can spot small unmanned aerial systems (sUAS), or drones, even when the radar cannot see them directly. The radar works at about 2.47 gigahertz in the 2.4 GHz ISM band. Using signals from one transmitter and four receivers, the authors feed a special radar image into a machine learning model and report an overall detection accuracy of 86.11% across five drone types in mixed line-of-sight and non-line-of-sight conditions.

The hardware is a continuous‑wave multiple‑input multiple‑output (MIMO) Doppler radar. Continuous‑wave here means the radar transmits a steady radio tone rather than pulses. The system uses a five‑element microstrip patch transmit array tuned near 2.47 GHz (about 4.1 dB peak gain and a half‑power beamwidth near 21°). It is designed to detect targets up to about 30 meters, with a transmitter power of 23 dBm and a target signal‑to‑noise goal of 30 dB. Each measurement records the complex baseband signal from all four receive channels at 1 kHz sampling.

The detection method looks for micro‑Doppler signatures. Micro‑Doppler refers to small, fast changes in the returned radio frequency caused by moving parts — in this case the rotating propeller blades. Those rotations create periodic patterns in the radar signal. The team computes spectral correlation densities (SCDs), which capture these periodic, or “cyclostationary,” features. SCD images from the four receiver channels form a 4×512×512 input that feeds a convolutional neural network (EfficientNet‑B0). The radar data are processed in roughly 2‑second observation windows, and the authors report that SCD calculation and model inference take only a few milliseconds, making near real‑time use feasible.

The system was tested with five drone platforms, three using plastic propellers and two using carbon‑fiber propellers. Experiments included both clear line‑of‑sight setups and deliberately obstructed, non‑line‑of‑sight measurements created with cabinets, doors, and walls. During tests the drones were held stationary while only their propellers rotated. Using data augmentation and a two‑stage training procedure (initially training custom layers, then fine‑tuning the full network), the model reached the reported 86.11% detection accuracy on this dataset.