Vision-based autonomous navigation for quadrotor UAVs in unknown and constrained environments

The demand for autonomous UAV navigation in complex environments is growing rapidly. However, most existing studies focus on simulations, lacking real-world validation under challenging conditions. This study systematically evaluates a vision-based navigation framework combining VINS-Mono for localization and Ego-Planner for path planning in GPS-denied, low-light, high-obstacle-density environments. Unlike traditional approaches that struggle with UAV high-frequency motion and dynamic constraints, VINS-Mono provides accurate high-frequency positioning via visual-inertial fusion. At the same time, Ego-Planner enables smooth, dynamically feasible real-time trajectory generation. Real-world experiments on three randomly generated maps (minimum obstacle spacing: 0.8 m (twice UAV wheelbase); obstacle density: 0.72 obstacles/m 2 ) achieved a minimum 70% success rate, with 87% of trajectories smoother than A* in simulations. Results demonstrate the framework’s robustness, though challenges such as field-of-view constraints and sensor-induced localization errors affecting replanning remain. This study presents a real-world evaluation of vision-based UAV navigation in constrained environments, offering insights into its feasibility and limitations. The findings contribute to improving trajectory optimization and UAV autonomy in constrained GPS-denied scenarios.