command Tree로 쉽게 package에 대한 tree를 만들 수 있다.

tree

https://blog.csdn.net/qq_18649781/article/details/86438300

super clears https://www.youtube.com/watch?v=mY3G_hjrt7A&list=PLubUquiqNQdP_H6uUmU-9f0y_LheA3Hil&index=2

https://darkpgmr.tistory.com/56

https://hellbell.tistory.com/entry/Iteratively-Reweighted-Least-Square-IRLS

http://www.cv-learn.com/20210313-ba/

센서는 모두 에러를 가지고 있기 때문에 infomration matrix 필요

error * information matrix * error

Eq{1} [e1,e2,e3][][e1,e2,e3]^T <- e1,e2,e3 are sensors

skew-matrix on information matrix meant relationship between each sensors in covariance matrix.

if we assume each sensor doesn’t impact on other, then, we can define skew matrix’s elements to “0”

diagonal elements are 1/sigma^2 which is simga is standard distribution of sensor.

from {1} we got, e(x) = (e1^2)/(sigma)^2 + (e1^2)/(sigma)^2 + (e1^2)/(sigma)^2 with that, we only care about biggest value.

to find armin(x)

1) Talyor expension of e 2) Linearization -> independent of delta x + linear dependent + quadratic dependency 3) iteration(delta x differentiation) -> H * delta x = -b 4) delta x = -H^-1 & b 5) x = x + delta x

this process is gaussian newton

**case if H is m x n matrix, we can not inverse it

using trick of linear algebra

H^t * H * dletax = H^t(-b) // pseudo inverse(when inverse matrix not possible)

delta x = H*b

**case very expensive inverse matrix method that’s why

many way to matrix decomposition is

• cholesky(matrix size small)
• conjugate gradient(matrix size big)

probabilistic notation : minimizing the error squared term is the equivalent to maximising the log-likelihood that our state vector x is on the measured point in our map.

Least Square는 즉 미지 상수 abc 를 구하는 것이다

Factor graph optimization is an optimization of any generic factor graph with nodes (states) and edges (constraints), for example you can have IMU preintegration constraints between two poses that you wish to minimize the error of based on the covariance matrix of the measurements.

Bundle adjustment is a special case of factor graph optimization where the only states are camera poses and landmark position, and the only constraints are the reprojection constraints from the landmarks to the cameras.

I’m not sure about g2o, but you can definitely use ceres to do either, as its just the case of defining the correct cost functions

https://arxiv.org/pdf/1704.05959.pdf

Bundle adjustment http://jinyongjeong.github.io/2020/03/01/Jacobian_of_BA/

https://jungwoong.tistory.com/92

즉 효율적으로 쓰레드 관리를 할 수 있다.

std::condition_variable를 사용해서

이는 공유 변수가 변경이 되면 condition variable을 통해 통지를 하게 되고,

condition variable을 가지고 있는 local {}은 통지가 올때까지는 쓰레드 중단, 사용 하지 않다가.

condtion variable에 의해 공유 변수 변경이 통지가 되었을때, 깨어나면서 쓰레드가 돈다.

즉, 쓰레드를 계속 돌리도록 방치하는 것이 아니라, condition variable을 통해서 공유 변수의 변경을 통해 쓰레드를 깨우고 잠재운다.