WaveSentry: Detecting Drone Signals with a HackRF, a Neural Network and an ESP32 Handheld
Thank you to Minsu (Gabriel) Kim for writing in and sharing WaveSentry, a drone RF detection prototype built with teammates Hyeongjun Kim and Junghwan Kim for D4D, a defense hackathon in South Korea. Their project uses a HackRF to feed I/Q samples to a Windows app, which converts them into spectrograms and uses a YOLO model to spot drone signal patterns. YOLO ("You Only Look Once") is a popular family of fast neural network models for detecting objects in images, and here it treats drone transmissions in the spectrogram as objects to find and draw boxes around. An "RF Observatory" view shows the spectrum, waterfall, and detections, and alerts can be sent to "Halo", an ESP32-S3 handheld with a circular LVGL display.
In their write-up on LinkedIn, the team explains that their first model, trained on synthetic backgrounds, raised alerts on recordings made with the drone switched off. They fixed this by recording real background I/Q with the HackRF and inserting segments of recorded drone signals into it to generate labeled training data. Minsu also notes that on November 14 they recorded the drone link on the platform and aboard a crowded train at Suwon Station, a busy rail station full of Wi-Fi and Bluetooth traffic, and they are now preparing a paper on the work.
We note that WaveSentry is still a prototype. The RF detection and the handheld alerts were tested separately, and alert latency and false alarm rates over the full path from receiver to handheld have yet to be measured.
