A rear end Optimization Overview
1. The back end module role
distal estimated from:
k moment the camera pose, xk
Map signpost j jj world coordinates, yj
Front-end estimate is based only on the previous time k or the first few moments. Back-end optimization done, it is based on all the information so far observed, estimated before the results of the optimization.
Back-end optimization, the pose of x and y are seen as signs a probability distribution on random variables. Known camera motion data u, observation data z, the problem is converted to determine the probability distribution, maximum likelihood estimation.
Back-end optimization There are two main methods:
Suppose Markov property, k time state only with the time k-1. That this lecture EKF method.
State at time k is assumed that all relevant previous state, based on a method of nonlinear optimization. The talk explained in detail in the next.
2. Description of the problem
variables:
XK: all the time k unknowns, comprising a camera and this point in time the m visible signs.
A known:
estimate:
According to all known distribution, now estimated pose and landmark status.
Second, the Kalman filter
1. linear system and a Kalman filter (KF) derived little specific
Kalman filter based on linear systems, there is no approximation is optimal linear unbiased estimate.
2. nonlinear system and an extended Kalman filter
The actual problem SLAM system:
- The equations of motion and nonlinear observation equation
- Gaussian distribution after nonlinear transformation, is no longer Gaussian distribution
It is necessary to approximate the non-Gaussian distribution is approximated to a Gaussian distribution. KF nonlinear system is the EKF. This is done by: a point in the vicinity of the motion equation and the observation equation to consider first-order Taylor expansion, i.e. the linear portion, and then follow KF
3). Limitations
- Markov of limitations
- A first order approximation of limitations
- Need to store the mean and variance, because a large amount of road signs, so a lot covariance matrix
Three BA optimization and map
BA (Bundle Adjustment): projection by heavy continuous optimization of parameter estimation. Because of its sparse properties can be achieved online operation.
Derivation slightly
After the October changed the original
Copyright: This article is the original article CSDN bloggers "KylinQAQ", and follow CC 4.0 BY-SA copyright agreement, reproduced, please attach the original source link and this statement.
Original link: https: //blog.csdn.net/pikachu_777/article/details/85041582