• beam rider
• detection
• drone
• neural network
• rete neurale
• ROS
• spoofing
• tracking
• uav

The Object of this work is to design and develop a system for detection, tracking, relative
motion estimation and capture via spoofing of a non-cooperative Unmanned Aerial Vehicle
(UAV).
The hybrid vision system is composed by a YOLO Neural Network trained to detect UAV and
an artificial vision algorithm that allows the tracking of it in every frame of the video stream.
The overall system is composed by an UAV Evader which has a point target, a Ground Station
that detects the UAV and point it with a laser, an UAV Pursuer that, first of all adopts a beam
rider guidance to get close to the Evader, when the Neural Network detects with some confi-
dence the presence of an UAV inside the camera frame a mobile waypoint guidance is adopted
using an estimation algorithm to track the evader and follow it. When the Pursuer is close
enough to the UAV Evader, a Spoofing technique is used to capture the Target.
An Unscented Kalman Filter (UKF) is developed for make a relative motion estimation between
the two vehicles using only the information obtained by the image frame deriving by the De-
tection and Tracking system. The measure of the Subtended Angle, derived from Computer
Vision Techniques is used to improve the estimation.
The entire system was implemented in ROS inside the Jetson TX2 embedded system carried
on a Tarot 650 Quadcopter with a stereo camera D435i properly stabilised.
Initially a simulation of the entire system using Simulink/MATLAB is performed. A testing
Hardware In the Loop (HIL) was done to verify the performance of the entire architecture.

Il sistema complessivo `e composto da un APR Evader che ha un obiettivo puntuale, una
stazione a terra che lo punta con un laser e da un APR Pursuer che, per prima cosa si porta
nelle vicinanze del drone Target utilizzando una guida Beam Rider, dopo di che, quando `e
sufficientemente vicino, passa ad una guida basata sulle informazioni che provengono dalla
rete neurale, che utilizza come sorgente video una camera RealSense D435i montata a bordo
e sufficientemente stabilizzata. Quando si ha un informazione sufficiente precisa del drone
Evader, inizia la fase di cattura mediante spoofing. Un filtro di Kalman Uscented (UKF) `e
stato sviluppato per la parte di stima della posizione relativa fra i due velivoli, in questa fase
vengono utilizzati unicamente i dati provenienti dal sistema di visione.L’intero sistema `e stato
implementanto in ROS all’interno di una Jetson TX2 montata a bordo di un quadricottero Tarot
650. La parte iniziale `e stata prettamente simulativa e svolta in MATLAB/Simulink. E’ stata
poi svolta una parte Hardware In the Loop (HIL) per verificare il funzionamento dell’intera
architettura.