Fixed-wing UAV for real-time vehicle tracking in GPS-denied environments
Undergraduate thesis on the study, design, and construction of an autonomous fixed-wing Unmanned Aerial Vehicle (UAV) for real-time detection and continuous tracking of moving vehicles, with emphasis on GPS-denied environments.
The airframe is based on a modified 3D-printed PETG skeleton (Titan Dynamics Slinger), reinforced with PA12-CF for structural integrity. A pusher configuration with a brushless outrunner motor provides propulsion. The onboard computing stack includes an NVIDIA Jetson Nano for vision processing, a Pixhawk-compatible flight controller running PX4, and a Raspberry Pi for telemetry relay. A custom power distribution board manages two separate LiPo batteries — one for the propulsion system and one for the computing payload — with LC filters to suppress BEC noise.
Vehicle detection uses YOLO11 Nano, fine-tuned on a custom aerial dataset of approximately 1,500 annotated frames collected from a DJI Phantom 4. The dataset was augmented with rotations, flips, HSV shifts, and mosaic transforms. The trained model was exported to TensorRT with FP16 precision, achieving 25–30 FPS inference on the Jetson Nano — sufficient for real-time operation. A tracking layer built on Kalman Filters and the Hungarian Algorithm maintains consistent target identities across frames.
When GPS is unavailable or degraded, the system switches to a visual localization pipeline. Camera frames are matched against geo-referenced map tiles using SuperPoint (feature detection) and SuperGlue (feature matching). The resulting position estimates are fused with IMU readings through an Extended Kalman Filter (EKF), providing continuous 6-DOF pose estimation without external positioning signals.
A Dynamic Resolution Scaling mechanism adjusts camera resolution based on the current computational load. When visual localization is active alongside detection and tracking, the system reduces resolution to maintain frame rate. The ground control station runs a custom Rust-based RPV communication system for real-time video downlink, telemetry, and RC override over raw 802.11 links.
