Zasnova in implementacija odločitvenega sistema na osnovi umetne inteligence za demonstratorsko vozilo na platformi Raspberry Pi

Autonomous driving increasingly depends on artificial intelligence, where perception, reliability assessment, and decision making must operate safely under limited computational resources. Convolutional neural networks (CNNs) achieve high accuracy in lane detection and steering prediction, but their deployment on embedded platforms is hindered by processing limits and the lack of mechanisms for evaluating output trustworthiness. In this master’s thesis, I address these shortcomings by developing a decision making system for a Raspberry Pi based vehicle demonstrator. The main challenge is how can trust aware decision making be implemented on an embedded autonomous vehicle platform using convolutional neural networks and sensor fusion to improve the reliability of autonomous driving behaviors. The work presents an architecture that combines two perception pipelines (LaneNet/Pilot- Net and SegNet/PilotNet), a set of computationally efficient trust indicators and a rule based policy layer for safe control. The trust indicators capture geometric quality, temporal stability and scene conditions, serving either as inputs to a voting model that selects the more reliable steering command or as triggers for the policy mechanisms. Experiments on recorded datasets and real time runs show that the voting model generally chooses the safer prediction, achieving smoother steering and preventing unsafe behavior. The results confirm that combining trust indicators and policy constraints enables reliable autonomous driving on embedded hardware.