Abstract: In response to the issues of low accuracy, poor robustness, slow operation speed, and deployment difficulties faced by traditional monocular depth estimation methods in marine environments, a lightweight depth estimation method for marine robots, named EDepth(Efficient Depth), aimed at enhancing the 3D(Three Dimension) perception capabilities of low-cost marine robots was proposed. Firstly, underwater light attenuation priors was utilized to map input data from the original RGB(Red-Green-Blue) image space to the RBI(Red-Blue-Intensity) input domain through spatial transformation, thereby improving the accuracy of depth estimation. Secondly, the efficient EfficientFormerV2 was employed as the feature extraction module, combined with the visual attention mechanism MiniViT(Mini Vision Transformer) and the light attenuation module to effectively extract and process depth information. Through the design of adaptive bin, the MiniViT module can dynamically adjust depth intervals, enhancing the accuracy of depth estimation. Finally, the network structure of EDepth has been optimized to achieve computational efficiency without sacrificing performance. Experimental results show that EDepth significantly outperforms traditional methods in depth estimation on the RGB-D(Red-Green-Blue Depth) dataset USOD10K. Specifically, EDepth achieves an absolute relative error of 0.587, compared to DenseDepth's 0.518. Although the latter performs better on certain metrics, EDepth only has 4.6 million parameters, reducing the parameter count by 89.69% compared to DenseDepth's 44.61 million parameters, with memory usage decreased to 23.56M, and achieving a frame rate of 14.11 Frames Per Second (FPS) on a single CPU, which is significantly better than DenseDepth's 2.45 FPS. This indicates that EDepth strikes a good balance between depth estimation performance and computational efficiency. The research findings in this paper theoretically provide new perspectives for the further development of marine robot technology and practically endow low-cost marine robots with advanced 3D perception capabilities, offering broad potential application prospects. By introducing EDepth, the application of marine robots in complex underwater environments will become more efficient and precise, promoting in-depth research and exploration in related fields.

Key words: 3D(Three Dimension), perception, adaptive bin, computational efficiency, EfficientFormerV2, marine robots, monocular depth estimation, RGB-D dataset

摘要： 针对传统单目深度估计方法在海洋环境中存在的精度低、鲁棒性差、运行速度慢和难以部署等问题，本文提出了一种轻量化的海洋机器人深度估计方法，命名为 EDepth(Efficient Depth)，旨在提升低成本海洋机器人的三维（3D）感知能力。首先，本文利用水下光衰减先验，通过空间转换将输入数据从原始RGB(Red-Green-Blue)图像空间映射到 RBI(Red-BlueIntensity)输入域，以提高深度估计的准确性。其次，EDepth 采用了高效的 EfficientFormerV2 作为特征提取模块，并结合视觉注意力机制 MiniViT 和光衰减模块，以实现深度信息的有效提取和处理。通过自适应分区的设计，MiniViT(Mini Vision Transformer)模块能够动态调整深度区间，提高深度估计的精度。最后，EDepth 的网络结构经过优化，能够在不牺牲性能的前提下，实现高效的计算。实验结果表明，EDepth 在 RGB-D(Red-Green-Blue Depth)数据集 USOD10K 上的深度估计性能显著优于传统方法。具体而言，EDepth 在绝对相对误差上达到了 0.587，相较于 DenseDepth 的 0.518，尽管后者在某些指标上表现更佳，但相比 DenseDepth 的 4461 万参数和 171.44M 的内存占用，EDepth 仅具备 460 万参数，减少了 89.69%的参数 量，内存占用减少至 23.56M，且在单个 CPU 的每秒帧数(FPS)达到 14.11，明显优于 DenseDepth 的 2.45。这表明，EDepth 在深度估计性能和计算效率之间取得了良好的平衡。本文的研究成果在理论上为海洋机器人技术的进一步发展提供了新的视角，在应用上则为低成本海洋机器人赋予了先进的三维感知能力，具备广泛的潜在应用前景。通过引入 EDepth，海洋机器人在复杂水下环境中的应用将变得更加高效与精准，推动相关领域的深入研究与探索。

关键词: 三维(3D)感知, 自适应分区, 计算效率, EfficientFormerV2, 海洋机器人, 单目深度估计, RGB-D数据集