Abstract: Deep learning based visual odometry methods can directly estimate the depth of monocular images and camera movement between adjacent images, which achieve high accuracy and improve running speed. However, this is based on a strong assumption of gray scale invariance, which is often not satisfied in real scenes. As a consequence, a self-supervised method for direct visual odometry based on image alignment is proposed, which gets a robust estimation result with the uncertainty regularization terms estimated from the uncertainty estimation network（uncertainty CNN, U-CNN） and it is called AUDVO（aligned U-CNN deep VO）. Meanwhile, a super resolution network is incorporated into the depth estimation module instead of using a simple interpolation operation for upsampling in order to deal with the holes caused by the inaccurate estimation in the large non-texture area. The evaluation results on the public KITTI datasets demonstrate the effectiveness of AUDVO for robust single-view depth estimation and visual odometry.

Key words: visual odometry, deep learning, uncertainty estimation network

摘要： 基于深度学习的视觉里程计方法（deep visual odometry，DVO）通过神经网络直接估计单目图像的深度和相邻图像之间的相机运动，在保证精度的同时大大提高了运行速度。但这是基于灰度不变假设，作为一个很强的假设，灰度不变假设在现实场景中往往难以满足。为此，提出一种基于图像对齐（image alignment，IA）的直接视觉里程计方法AUDVO（aligned U-CNN deep VO），通过不确定性估计网络（uncertainty CNN，U-CNN）引入正则项进行约束，使得估计的结果更具鲁棒性。为了处理大面积纹理缺失区域上因估计不准确带来的空洞，在设计深度估计模块时通过嵌入超分辨率网络进行上采样。在公开的KITTI数据集上的实验证明了AUDVO在深度和相机位姿估计上的有效性。

关键词: 视觉里程计, 深度学习, 不确定性估计网络