Abstract:

Aiming at the obvious scale uncertainty problem of the direct method DSO （Direct Sparse Odometry）, the impact of scale uncertainty on the system positioning accuracy is analyzed, and using a deep learning network that performs depth prediction on a single image is proposed. For the time-consuming problem of the back-end, a Preconditioned Conjugate Gradient（PCG） algorithm is utilized to accelerate the back-end solution. Compared with ORB-SLAM2, DSO, LDSO on the KITTI public dataset, the accuracy of localization of the system is significantly improved.

Key words: depth estimation, visual odometry, scale, preconditioned conjugate gradient, localization

摘要：

针对直接法DSO（Direct Sparse Odometry）存在的明显的尺度不确定性问题，对尺度不确定性给系统定位精度带来的影响进行分析，提出将对单幅图像进行深度估计的深度学习网络和DSO相结合的融合算法；针对DSO后端耗时问题，提出运用预处理共轭梯度（Preconditioned Conjugate gradient，PCG）算法优化后端求解部分。在KITTI公开数据集上与ORB-SLAM2、DSO、LDSO进行对比测试，系统的定位精度得到显著提高。

关键词: 深度估计, 视觉里程计, 尺度, 预处理共轭梯度, 定位