Comparison of Pretrained Monocular Depth Estimation Models for Autonomous Drone Obstacle Avoidance

In recent years, the industrial use of drones has been attracting significant attention and is being utilized in various fields. In particular, small drones are highly valued for their high maneuverability, affordability, and ease of use, expanding their range of applications. However, small drones are subject to weight restrictions, which limit the number and types of sensors they can carry. As a result, they may not be able to equip the depth sensors required for autonomous flight, or their measurable range may be short. Additionally, advancements in deep learning technology have led to the development of monocular depth estimation models that can estimate depth without the use of depth sensors. This study aims to achieve autonomous drone flight in a simulation environment by combining a path planning method using depth images with a depth estimation model that estimates distance from RGB images, constructing an algorithm that can also be applied in real-world scenarios. In the experiments, multiple depth estimation models were compared and their obstacle avoidance performance was quantitatively evaluated.