VIO

BA

定义:通过优化相机位姿特征点的空间位置,使得每条光束都能打到光心(都能打到光心的意思就是不断迭代减少打到光心的误差。

 

一个SLAM问题求解的整体思路

1.确定所有的约束关系(因子图)

2.针对每个约束,确定三个要素:误差项、优化变量、协方差

3.计算雅可比矩阵,调用GN、LM等,求解BA

 

高斯牛顿法

假设e(x+\Deltax)是n*1维,那么P^{-1}就是n*n维,得到的结果就是1*1        

假设优化变量x的维度是m*1

 

EKF滤波和优化

1.EKF滤波相当于只迭代一次的优化

2.多次优化精度比滤波高,但效率低于滤波

 

IMU预积分

PVQ:位置位移旋转

1.积分下一个时刻的PVQ作为视觉初始值

2.预积分相邻帧的PVQ变化量,作为IMU约束

3.计算IMU误差的协方差和雅可比                                                                                                                                                                                                                                  

VIO分类

将视觉约束加入到联合优化是紧耦合,

将视觉约束后的位姿加入到联合优化是松耦合 。

 

将当前帧看不见的路标点(空间点,3维)太老的路标点(能被所有帧看到的点)边缘化(即删掉),然后对删掉的路标点所连着的位姿之间两两加上约束, 之后删掉最近的路标点看不到的帧。                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        

 

你可能感兴趣的:(VIO)