一种结合等变滤波和视觉观测网络的混合视觉惯性里程计

A hybrid visual-inertial odometry combining equivariant filter and visual observation network

关键词:  视觉惯性里程计  等变滤波  光流估计  预训练  深度学习

Key words:  Visual-inertial odometry  Equivariant filter  Optical flow estimation  Pretraining  Deep learning

摘要:

传统的面向惯导的滤波器在设计时通常没有考虑零偏状态的几何性质，针对这一问题，提出了新的群以及群作用，并以此为基础推导了一种新的等变滤波方法。为了验证该滤波算法的性能，构建了一个基于学习的结合等变滤波和视觉观测网络的视觉惯性里程计系统，名为Deep-EqF-VIO。首先，构建了流形上的导航状态动力学系统，其次根据等变滤波原理推导了等变滤波方法，最后结合视觉观测网络构建了视觉惯性里程计系统。为了进一步提高系统性能，提出了一种融合稠密光流伪标签的预训练方法，通过光流估计任务引导视觉观测网络从输入图像中提取更加通用的几何运动特征。实验结果表明，在大多数场景中，Deep-EqF-VIO相较于同类方法有更好的表现，并且在使用预训练方法进行重新训练后，性能得到了进一步的改善，误差下降率最大达到了49.80%，具有一定的应用前景。

Abstract:

Traditional geometric filter algorithms are usually designed without considering the geometric properties of bias states. To deal with this problem, a group and a group action are proposed, and a new equivariant filter method is derived. To evaluate the performance of this filter algorithm, a visual-inertial odometry system based on an equivariant filter and a visual observation network, named Deep-EqF-VIO, is developed. Firstly, a navigation state dynamics system is constructed on the manifold, then an equivariant filter method is derived by the equivariant filter principle, and finally a visual-inertial odometry system is constructed by combining with a visual observation network. In order to further improve the performance of the system, a pretraining method incorporating with dense optical flow pseudo-labels is proposed. This method guides the visual observation network to extract more generalized geometric motion features from input images via the optical flow estimation task. Experimental results show that Deep-EqF-VIO achieves the best accuracy in most scenes compared to similar algorithms. After retraining the VIO system by using the proposed pre-training method, the VIO performance can be further improved, with the maximum error reduction rate reaching 49.80%. This method has a certain degree of application potential.