基于局部注意力和位姿迭代优化的自监督单目深度估计算法

A Self-supervised Monocular Depth Estimation Algorithm Based on Local Attention and Iterative Pose Refinement

自监督单目深度估计在自动驾驶、智能制造等领域有着广泛的应用。然而由于自监督训练存在大量训练噪声,其估计精度受到了极大限制。针对自监督单目深度估计算法中深度估计精度有限的问题,本文提出了一种基于局部注意力机制和迭代调优的自监督单目深度估计框架。首先,对于深度估计网络,基于局部像素间深度值的高度相关性,本文设计了一种局部注意力机制来融合高分辨率特征图的局部特征,提升深度估计的准确性;其次,对于位姿估计网络,本文设计了一种迭代调优的位姿估计结构,利用残差优化的方式降低位姿估计难度,提升位姿估计的准确性进而提升深度估计网络的性能。实验表明,本文提出的改进自监督单目深度估计算法有效提升了深度估计的精度。

Self-supervised monocular depth estimation is widely used in many areas,such as autonomous driving and intelligent manufacturing. However,due to the large amount of training noise in self-supervised training,the accuracy of self-supervised monocular depth estimation is limited. To improve the performance of self-supervised monocular depth estimation algorithm,we proposed a modified self-supervised monocular depth estimation algorithm based on local attention mechanism and iterative pose refinement. First,for the depth estimation network,we proposed a local attention mechanism,which is based on the high correlation between the depth of pixels in a local patch,to fuse features of highresolution feature map. Second,for the pose estimation network,we proposed an iterative refinement based architecture,which decreases the pose estimation difficulty with residual optimization and improves the pose estimation accuracy to benefit the depth estimation network. Experiments shown that,the proposed modified self-supervised monocular depth estimation algorithm significantly improves the depth estimation accuracy.