Autonomous vehicle control based on deep learning methods

Tradicionalni sustavi autonomne vožnje sastoje se od percepcije, lokalizacije, mapiranja, planiranja i praćenja putanje, gdje svaki korak može biti vrlo računski zahtjevan pogotovo u kompleksnim i dinamičkim okruženjima. U diplomskom je radu potrebno ostvariti autonomnu vožnju automobila korištenjem dubokih modela, poput konvolucijskih neuronskih mreža koji će na temelju senzorskih podataka (iz kamere, lasera i radara) izravno odrediti upravljačke naredbe. Evaluaciju modela je potrebno napraviti u simulatoru tako da automobil tijekom vožnje uvijek ostane na stazi pri različitim vremenskim uvjetima i uvjetima osvjetljenja.Traditional autonomous driving systems consist of perception, localization, mapping, planning and trajectory tracking, where each step can be very computationally demanding, especially in complex and dynamic environments. In the thesis, it is necessary to achieve autonomous car driving using deep models, such as convolutional neural networks, which will directly determine control commands based on sensor data (from cameras, lasers and radars). The evaluation of the model must be done in the simulator so that the car always stays on the track while driving under different weather and lighting conditions