Computer vision vs. deep learning: the different types of training in autonomous vehicle development

Abstract

The development of autonomous vehicles relies on advanced image processing and artificial intelligence techniques, with computer vision and deep learning being two fundamental approaches. While computer vision uses traditional algorithms to interpret images and extract specific features, deep learning employs neural networks to recognize complex patterns and make decisions autonomously. In this study, different training models applied to an autonomous car were compared: one based on computer vision, programmed to follow a central yellow line on the track, and another based on deep learning, featuring five types of behavioral cloning training. Tests showed that computer vision struggled with varying lighting conditions, resulting in navigation errors, while deep learning demonstrated greater precision and adaptability, successfully completing laps on the track consistently. The results indicate that although computer vision is effective for specific tasks, its rigidity makes it less efficient in dynamic environments, whereas deep learning proves to be a more robust and promising approach for autonomous vehicle training