解耦动态区域的自监督单目深度估计模型

Self-Supervised Monocular Depth Estimation Model with Decoupled Dynamic Regions

摘要: 近年来，自监督单目深度估计因其无需深度标签的优势，在计算机视觉领域获得了广泛关注。然而，传统自监督单目深度预测方法通常基于静态场景假设，这导致在相邻帧中出现动态对象时，深度预测的精度会显著下降。为了解决这一问题，本文提出了一种多帧自监督单目深度估计模型。该模型通过分割网络预先识别图像中的运动物体，并利用多帧图像之间的光流信息来重构图像。通过将静态场景与动态物体分开处理，该方法有效提高了动态物体深度估计的准确性。此外，本文设计了动态物体重构损失(Dynamic Object Reconstruction Loss, DRL)和深度一致损失(Depth Consistency Loss, DCL)，以监督动态重构图和重构深度图的生成。实验结果表明，在三个公共数据集上，该方法优于现有的主流方法，能够在动态场景中准确预测深度图。

Abstract: Recent years, self-supervised monocular depth estimation has garnered extensive attention in the field of computer vision due to its advantage of not requiring depth labels. However, traditional self-supervised monocular depth prediction methods are typically based on the assumption of static scenes, which leads to a significant decrease in depth prediction accuracy when dynamic objects appear in consecutive frames. To address this issue, this paper proposes a multi-frame self-supervised monocular depth estimation model. The model identifies moving objects in images through a segmentation network and reconstructs images using optical flow information between multiple frames. By separating static scenes from dynamic objects, this approach effectively improves the accuracy of depth estimation for dynamic objects. Additionally, this paper have designed the Dynamic Object Reconstruction Loss (DRL) and Depth Consistency Loss (DCL) to supervise the generation of dynamic reconstruction images and reconstructed depth maps. Experimental results demonstrate that this method outperforms existing mainstream approaches on three public datasets, enabling accurate depth prediction in dynamic scenes.