走廊场景下辅助视觉里程计初始化的单目深度恢复方法

Monocular depth recovery method assisting visual odometry initialization in corridor environment

单目相机由于缺少尺度信息在视觉里程计等应用场景中性能受限。现有研究大多通过基于深度学习的方法解决这一问题,但其推理速度慢,难以实时运行。针对这一问题,提出了一种走廊场景下基于非线性优化进行快速单目深度恢复的显式方法。采用虚拟相机假设,简化对相机姿态角的求解;通过最小化几何残差,将深度估计问题转换为优化问题;设计一种深度平面构建方法,对空间点深度进行分类,实现走廊等封闭结构场景下的快速深度估计;最后,将所提方法在单目视觉里程计初始化中进行应用,使得单目视觉里程计可以获得真实的尺度信息,并提升其定位精度。实验结果表明:所提方法在走廊场景3m范围内深度估计的相对误差小于8.4%,在Intel Core i5-7300HQCPU处理器中能以20FPS的速度实时运行。

Due to the lack of scale information,the performance of monocular cameras is limited in application scenarios such as visual odometry.Existing researchers mostly address this issue through deep learning-based approaches,yet their inference speed is slow,leading to poor real-time capabilities.To solve this problem,an explicit method based on nonlinear optimization is proposed for rapid monocular depth recovery in corridor environment.The assumption of virtual camera is adopted to simplify the solution of the camera pose angles.The depth estimation problem is transformed into an optimization problem by minimizing the geometric residual.A depth plane construction method is designed to categorize the depth of space points,facilitating swift depth estimation in enclosed structural scenarios,such as corridors.Finally,the proposed method is applied in the initialization process of monocular visual odometry,so that the monocular visual odometry could obtain real scale information and improve its positioning accuracy.The experimental results show that the relative error of depth estimation of the proposed method is less than 8.4%in the corridor scene within 3 m,and can run in real time at a speed of 20 FPS on the Intel Core i5-7300HQCPU processor.