Background.

The proliferation of Unmanned Aerial Vehicles (UAVs) presents significant security challenges due to their ability to bypass traditional perimeters. Detecting and tracking small, low-flying UAVs with radar is difficult due to low RCS, clutter interference, and slow speeds. Real-time processing and trajectory prediction on resource-constrained embedded platforms are critical for effective counter-UAV systems but remain significant challenges.

Objectives.

This thesis aimed to develop, implement, and validate a radar-based framework capable of real-time UAV detection and trajectory prediction in real-time, specifically optimized for execution on an NVIDIA Jetson AGX Orin embedded platform.

Methods.

A literature review was first conducted to understand existing approaches in radar-based tracking, trajectory prediction, and drone detection. Following this, a system was developed using a COTS X-band phased-array radar and a Jetson AGX Orin. Synchronized radar and GPS ground-truth data were collected using a 10-inch FPV drone in an open-field environment. A processing pipeline involving MTI filtering, Hann Windowing, 2D FFT, CFAR detection, DBSCAN clustering, and Kalman filtering was implemented to generate target tracks. A custom Transformer model utilized features from these tracks (position, velocity) to perform next-frame trajectory prediction and implicit drone presence classification (zero vector output indicating absence). Performance was evaluated on a held-out test set.

Results.

The software pipeline executed efficiently on the Jetson platform (374 ms/frame), but the overall system throughput was limited by the radar hardware’s slow frame rate (0.75 FPS). The classification achieved perfect precision (1.0) but low recall (0.53), indicating high reliability when a drone was detected but frequent missed detections. The Transformer model predicted the next-frame position with a median Euclidean error of approximately 11.07 meters.

Conclusions.

The feasibility of integrating classical radar processing with AI-based trajectory prediction on embedded hardware was demonstrated. While software efficiency was high, system performance was bottlenecked by sensor hardware speed. The implemented pipeline showed a trade-off favouring high precision over detection sensitivity. The Transformer architecture shows promise for embedded trajectory prediction, we expect that the accuracy is currently limited by the temporal sparsity of the input data.