DOA-mono: Monocular Depth Estimation Focused On Dynamic Object Perception

Monocular depth estimation, as a fundamental task in the field of computer vision, has demonstrated significant application potential in areas such as autonomous driving and virtual reality (VR). Dynamic object perception refers to the model's ability to identify, locate, and effectively process the depth information of moving objects within a scene. Existing methods have shown excellent results in static scenes; however, when dealing with scenes containing moving objects, the rigid scene assumption is often violated due to the potential movement of fixed reference points in the scene. Existing methods lack the ability to perceive dynamic objects, resulting in reduced accuracy and low stability in depth estimation. To address the challenges posed by motion blur and dynamic scene changes caused by moving objects in monocular depth estimation, this paper proposes a monocular depth estimation network based on dense attention skip connections and a multi-scale feature fusion module. The proposed network enhances the model's dynamic object perception capability, improving depth estimation accuracy and robustness in complex dynamic scenes.