Vision-based Algorithms for UAV Mimicking Control System

Abstract

Vision-based algorithms designed to detect and track UAVs from an onboard moving platform have been the focus of active and extensive research over the last decade, and dozens of algorithms have been tested, compared and optimized. However, the existing approaches tend to rely on specific features such as color or edges which may not be able to detect and track various types of flying quadcopters. This thesis implements a modified version of an existing vision-based algorithm, the Cascade Classifier, originally designed to recognize facial features and humans, and demonstrates its capability of detecting and track any type of quadcopter with great accuracy over a variety of backgrounds, in both indoors and outdoors flight conditions. The Cascade Classifier algorithm is demonstrated on two specific quadcopter models used for this study, the 3DR Solo and the Parrot AR.Drone 2.0. This thesis introduces a novel method to reduce the amount of information which needs to be processed by vision-based algorithms when tracking physical objects undergoing non-random motion in 3D space. This method employs a Kalman filter to predict the estimated position, velocity and acceleration of the tracked object in order to reduce the image area in which the tracked quadcopter is believed to be. This enables the Cascade Classifier algorithm, or any other type of vision-based detection algorithm to track the target vehicle while greatly reducing the required image processing time. Experimental testing proves that the proposed algorithm obtains good detection and tracking performance in real-time for both quadcopter types in indoor and outdoor flight scenarios, as well as the successful performance of the mimicking control system design.