Unmanned Aerial Vehicles (UAVs), commonly known as drones, have emerged as versatile platforms revolutionizing diverse fields such as precision agriculture, environmental monitoring, infrastructure inspection, and disaster management. Their growing impact is driven by advances in autonomy that enable efficient, scalable, and intelligent operations. Among the enabling technologies, computer vision plays a pivotal role by allowing drones to perceive, interpret, and interact with their environment through visual sensing. This capability has significantly improved tasks such as obstacle detection, localization, mapping, and scene understanding, even in GPS-denied or visually degraded conditions. Despite these advances, achieving full autonomy remains a major challenge. Vision-based navigation is often hindered by adverse weather, illumination changes, dynamic obstacles, and computational limitations, alongside issues of domain shift and long-tail edge cases that compromise reliability and safety. This review paper presents a comprehensive analysis of recent progress in vision-based autonomous drone navigation, spanning classical computer vision, deep learning, and emerging paradigms such as Vision Transformers, reinforcement learning, and self-supervised learning. It further highlights open challenges and outlines future research directions, including multi-sensor fusion, domain adaptation, collaborative perception, neuromorphic computing, and explainable AI—aiming to guide the development of resilient and robust UAV systems capable of dependable real-world autonomy.