Deep-Init:基于深度学习的视觉惯性里程计非联合初始化方法  

Deep-Init:Non Joint Initialization Method for Visual Inertial Odometry Based on Deep Learning

作　　者：史殿习 高云琦 宋林娜 刘哲 周晨磊 陈莹 SHI Dianxi;GAO Yunqi;SONG Linna;LIU Zhe;ZHOU Chenlei;CHEN Ying(Intelligent Game and Decision Lab(IGDL),Beijing 100091,China;National Innovation Institute of Defense Technology,Beijing 100071,China;College of Computer,National University of Defense Technology,Changsha 410073,China;Tianjin(Binhai)Artificial Intelligence Innovation Center,Tianjin 300450,China)

机构地区：[1]智能博弈与决策实验室,北京100091 [2]国防科技创新研究院,北京100071 [3]国防科技大学计算机学院,长沙410073 [4]天津(滨海)人工智能创新中心,天津300457

出　　处：《计算机科学》2024年第7期327-336,共10页Computer Science

基　　金：科技部科技创新2030-重大项目(2020AAA0104802);自然科学基金(91948303)。

摘　　要：对于非线性的单目VIO系统来说,其初始化过程至关重要,初始化结果的好坏直接影响整个系统运行过程中状态估计的精度。为此,将深度学习方法引入单目VIO系统的初始化过程中,提出了一种高效的非联合初始化方法(简称Deep-Init),其核心是使用深度神经网络对IMU中陀螺仪的偏置和噪声等随机误差项进行准确估计,得到初始化过程中的关键参数,即陀螺仪的bias;同时,将IMU预积分与SfM进行松耦合,通过位置和旋转对齐,使用最小二乘法对绝对尺度、速度以及重力矢量等进行快速恢复,并将其作为初始值来引导非线性紧密耦合的优化框架。由于深度神经网络对陀螺仪数据进行补偿,从而大大提高了IMU中旋转估计量的准确性,有效提高了IMU数据的信噪比,同时减少了最小二乘方程失效的次数,因此进一步减少了计算量。使用去除误差项的陀螺仪数据的预积分量替换SfM中的旋转量,将IMU的旋转量作为真值,不仅避免了将不准确的SfM值作为真值进行初始化时所带来的误差,有效提升了系统状态估计的精度,而且能够有效地适应高速运动、光照变换剧烈和纹理重复等SfM估计效果差的场景。在EuRoC数据集上,对所提方法的有效性了进行实验验证,实验结果表明,所提出的初始化方法Deep-Init无论是精度还是耗时均取得了良好的效果。For a non-linear monocular VIO system,its initialization process is crucial,and the initialization result directly affects the accuracy of the state estimation during the whole system operation.To this end,this paper introduces a deep learning method into the initialization process of the monocular VIO system and proposes an efficient non-joint initialization method(referred to as Deep-Init).The core of this method is to use a deep neural network to accurately estimate the random error terms such as bias and noise of the gyroscope in the IMU,to obtain the key parameter in the initialization process,i.e.the bias of the gyroscope.At the same time,we loosely couple the IMU pre-integration to the SfM.The absolute scale,velocity and gravity vector are quickly recovered by position and rotation alignment using least squares,which are used as initial values to guide the non-linear tightly coupled optimization framework.The accuracy of the rotation estimates in the IMU is greatly increased due to the compensation of the gyroscope data by the deep neural network,which effectively improves the signal-to-noise ratio of the IMU data.This also reduces the number of least squares equation failures,further reducing the computational effort.Using the pre-integrated amount of gyroscope data with the error term removed to replace the rotation amount in the SfM and using the IMU rotation amount as the true value,not only avoids the errors associated with initializing inaccurate SfM values as the true value but slao effectively improves the accuracy of system state estimation.Moreover,it enables effective adaptation to scenarios where SfM estimation is poor,such as high-speed motion,drastic lighting changes and texture repetition.The validity of the proposed method is verified on the EuRoC dataset,and the experimental results show that the proposed Deep-Init initialization method achieves good results in terms of both accuracy and time consumption.

关 键 词：视觉惯性里程计 深度学习 初始化 惯性测量单元 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]