From CNNs To VLMs: A Complete Review And Experimental Comparison of AI Models on Autonomous Vehicles

Abstract

This paper presents a critical analysis and experimental comparison of three AI paradigms in autonomous driving: convolutional neural networks (CNNs), large language models (LLMs), and vision-language models (VLMs). We reviewed more than thirty published papers spanning 2015- 2024 and conducted a practical benchmark using ResNet-50 (CNN), Flan-T5-base (LLM), and BLIP-2 OPT-2.7B (VLM) on 10 real-world driving scenes across 7 assessment categories, yielding 70 question-answer pairs. Our experiments show that VLMS achieve an overall accuracy of 84.3% (59/70), substantially outperforming CNN (61.4%, 43/70) and LLM (60.0%, 42/70) in visual object detection and natural-language scene understanding. The results indicate that VLMs represent the most promising next-generation paradigm for autonomous driving, unifying perception, reasoning, and explainability within a single model.