Depth Map Generation for Underwater Platforms Using Monocular Cameras

Depth maps represent the 3D geometry of a scene, which can help underwater unmanned vehicles to better perceive their surrounding environments. Monocular camera is an economical and effective sensor that are suitable for depth map generation underwater. Different from existing end-to-end monocular depth map generation model, in this paper, we introduce a three-stage framework for depth map generation in underwater environments, using solely a monocular camera. The key idea is that depth maps are a composite of a series of sub-regions with known depths, and the clarity variations in the digital images can indicate the depth information. We firstly employ underwater image enhancement model to improve imaging quality, followed by a segmentation model to obtain the regions of interest from the enhanced image. By designing a composite clarity feature indicator and deriving the optical empirical formula of the monocular camera, we propose an effective underwater monocular depth estimation method and analyze its error. Real-world experiments are presented to evaluate the proposed framework. Compared to recent underwater depth map generation models, our model not only can obtain absolute depth but also achieves better depth imaging performance.