Computer Vision Based Guidance, Navigation and Control for Autonomous Aerial Vehicles: A Systematic Survey

This systematic survey examines advances in computer vision techniques that support guidance, navigation and control (GNC) for autonomous aerial vehicles (AAVs) - including multirotor UAVs, hybrid VTOL/eVTOL platforms and fixed wing drones - operating in unstructured, GNSS‑denied environments. Guided by the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta‑Analyses) protocol, a targeted query was used to perform a literature search across six academic databases for the period 2005–2025. Each selected contribution was mapped onto the five functional stages of autonomous flight operations - sensing, perception, navigation, guidance and control - producing a comprehensive comparison matrix that links computer vision algorithms, sensor configurations and vehicle classes. The synthesis shows that current aerial vehicle autonomy relies mainly on visual inertial odometry (VIO) and visual SLAM (vSLAM) pipelines. Among the simultaneous localization and mapping (SLAM) solutions, ORB‑SLAM3 stands out for multi‑map capability and support for monocular, stereo and RGB‑D inputs, making it a robust backbone for GNSS‑denied localization. For real-time obstacle detection, one stage deep learning detectors such as YOLO‑v8 and YOLO‑Tiny achieve frame rates exceeding 30 FPS on embedded GPUs, while two stage architectures (R‑CNN, Faster R‑CNN) provide higher detection accuracy at the expense of latency. Adaptive multimodal sensor fusion - combining cameras with inertial measurement unit (IMU), light detection and ranging (LiDAR), radar or event‑based sensors - substantially mitigates the impact of adverse illumination, weather and texture‑poor scenes, thereby extending operational envelopes. However, three main gaps limit the transition from prototypes to certifiable flight systems: limited robustness in extreme conditions, the need to integrate perception with flight control hardware efficiently, and the lack of standard benchmarks and certification for AI perception. The paper calls for full‑cycle datasets, research on low‑power inference and hardware acceleration, and regulatory paths that include explainable AI and continuous monitoring. Overall, this study provides an exhaustive classification and performance comparison of computer vision solutions for autonomous aerial vehicles, clarifies the strengths and limitations of existing approaches across vehicle categories.