Autonomous navigation of mobile robots based on neural networks

Rad se bavi aktualnim problemima autonomne navigacije vozila uz pomoć alata dubokog učenja. Implementacija i optimizacija algoritama detekcije objekata na slikama ili video zapisima, te njihova aplikacija na robotima jedna je od gorućih tema moderne tehnologije. Rad sadrži kratki pregled povijesti robotike i strojnog učenja, dotiče se utjecaja "pametnih" robota na moderno društvo te daje teorijski uvod u razvoj umjetnih neuronskih mreža. Poseban naglasak stavljen je na konvolucijske neuronske mreže iz razloga što se danas najviše koriste pri prepoznavanju i klasifikaciji objekata. Za projekt je korišten TurtleBot3 Waffle model robota. Na kraju je dan pregled raznih algoritama detekcije objekata (R-CNN, Fast R-CNN, Faster R-CNN itd.) te je opisan način treninga već postojećih mreža (Inception v2, Darknet) korištenih u ostvarivanju autonomne navigacije TurtleBot3 robota.The paper deals with current problems in field of autonomous navigation with the help of deep learning algorithms. Implementation and optimizaton of object detection algorithms, together with their application in the field of robotics, has become one of the most important tasks of modern technology. The paper gives a quick overview of the history of robotics and deep learning, deals with the impact of "smart" robots on modern society and provides theoretical background of artifical neural networks, with the emphasis on convolutional neural networks as on the most important tools in object clasification and detection. TurtleBot3 Waffel has been used as the robot model for this project. In the end, the paper provides a short overview of different object detection algorithms (R-CNN, Fast R-CNN, Faster R-CNN etc.) and the way different neural networks (Inception v2, Darknet) have been retrained in order to help achieve autonomous navigation of TurtleBot3 robot