Monocular Depth Estimation with Enhanced Target Awareness for UAV Environment Perception

Depth perception has draw a lot of attention recently as it enables 3D sensing capabilities for robots. Monocular depth estimation offers cost-effective and computationally efficient solutions, making it an ideal choice for deployment on resource-constrained platforms such as unmanned ariel vehicles (UAVs). However, existing work primarily focuses on indoor reconstruction and ground-based autonomous driving - methods that are not directly applicable to aerial scenarios. Moreover, during UAV flight, particular emphasis is placed on specific entities like targets and obstacles, whereas current methods typically perform global depth estimation and thus fail to satisfy this requirement. To tackle these challenges, this work presents a novel approach that jointly learns outdoor depth and target-region information. Target prior knowledge is injected into multi-scale encoder features through a novel mask-guided feature fusion strategy, sharpening object boundaries without adding extra constraints. A target-prioritized loss amplifies the model's attention on regions of interest. Extensive experiments on the DDOS dataset demonstrate that our method achieves state-of-the-art performance.