Ευφυή αυτόνομα μη επανδρωμένα αεροσκάφη για την ανίχνευση και τον εντοπισμό αντικειμένων σε ακατάστατα άγνωστα περιβάλλοντα: Αλγόριθμοι ανίχνευσης αντικειμένων

Abstract

The use of unmanned aerial vehicles (UAV), also known as drones, has increased enormously in the last decade. Other than military purposes, special kind of drones have been developed and widely used in new innovative categories such as aerial surveillance, in video film production, for protection and law enforcement, for search and rescue purposes and in many other categories. This thesis research focuses on object detection on a variety of pictures. The unmanned aerial vehicle, processes a picture with the relevant algorithm and detects certain items that it includes. The results of the detection will be used by the aerial vehicle to navigate correctly in an indoors environment. More specifically, the thesis focuses on the detection of survivors inside of a burning building. My approach involves the use of the system “R-CNN: Regions with Convolutional Neural Network Features” [1], a convolutional neural network that has been trained with a large volume of photos from various known datasets related to visual recognition. What the algorithm does by the end of one execution, it displays on the photo we give as input, each object individually in ascending order based on the degree of confidence that has been identified by MATLAB. Studying and experimenting with the code of the system described above, in this thesis an extra module has been written that limits the detection on a specific object on command. Then produces a vector of 100 sections, which each one of them will get a value of 1 or 0 accordingly if the item found in the photo, intersects with the section. The vector can subsequently be used by a program that gives commands for future movements to the UAV. Besides the vector, the output includes the corresponding ‘score’ which is taken from the program, to be used as a reward measure in the thesis work of Michael Panagiotou, employing “Reinforcement Learning” [12] using the symbolic language programming THEANO, to improve the decision path to be followed by the aircraft after a history of steps decided in one flight. The unmanned vehicle used for this thesis is the model AR.Drone 2.0 by Parrot and the code is written in MATLAB.