Unsupervised Monocular Depth Estimation Based on Multi-Scale Feature Fusion in Berthing Scenarios

Addressing the challenges of weak texture and detail loss in predicted depth maps during the docking and mooring of unmanned surface vehicles (USVs) in marine environments, we propose a monocular distance estimation algorithm based on multi-scale feature fusion. Asymmetric convolution counters weak texture interference in depth outputs, while multi-scale receptive fields maintain image details. Together they boost monocular distance estimation accuracy for unsupervised models in harbor maneuvering scenarios. Finally, the proposed algorithm is validated and evaluated using a constructed dataset of marine docking and mooring scenes. Experimental results demonstrate that, compared to the baseline model SC_Depth, the relative error is reduced by 0.193, the root-mean-square error is reduced by 4.248, and the logarithmic root-mean-square error is reduced by 0.375. Furthermore, the prediction accuracy shows significant improvement compared to other state-of-the-art algorithms such as SFMLearner and Monodepth2.