3 questions

1.How is the covariance of an imaginary line formulated? can the cross co-relation of points be used to help it?

2. The value of H determines K, the correction. Can we prevent certain updates that might hamper the consistency of EKF?

3. Observability Analysis

mandate

make new case
plot X vs X estimate, and P estimates in all cases

EKF

in EKF,
dP/dk is actually dPprior/dk

1. why is covariance not dependent on z?
2. determinant of P matrix decreasing monotonically, what does it mean?
3. If I see a new landmark whilst also seeing old landmarks, how can uncertainty in the new one be very high? it should be top-limited by the relative uncertainty of other landmarks

To do

1. Complete box demo
2. Talk to Dr.PRK Rao about EKF SLAM (Thursday)
3. EKF SLAM C program complete (all variables and prediction function)
4. Complete reading GPSLAM/ 2 dissa papers

by weekend

010909

Deadline: 04/09/09
---------------
1. Making Visibility work completely.(no code flaws)
2. Making Clustering program work in the lab environment.
-Near features might cause problems with gate values, hence need to be discarded/an intelligent solution is required.(Discarding also needs to be consistent)
3. Understand all work and core variables and how they change and jacobians.
---------------------
Deadline:16/09/09

1. One SLAM run start-end-start(w or w/o loop closing)
2. Drawing like in CoreSLAM(Saving all scan values and position offline and making a plot) and comparing it with only odometry based mapping results.
-----------------

180809

robot error in movement is causing some non-visible lines to be called pakka lines.

z(2)/association is working fine.
Also needed,
if a point landmark is associated flag

150809

New Bug discovered in code :

Whilst the lines are getting associated well, the value of the line in the state and the value of z(2) of a line which gets associated with it are quite different surprisingly.

150809 + Mahalanobis Distance

In the current visibility based prediction, if a line p is termed to be visible, it means that the line p is visible from the mean of the position of the robot. However, the position of the robot is not necessarily at the mean. Hence, the prediction of lines might go haywire when a few lines are visible from the mean but not from the current position of the robot which is not at the mean. In other words, we would forcibly be associating lines as we estimated them to be visible.

To prevent this, we need to test visibility from extrema of the error ellipse for x, y and theta. Upon finding the line visible from both extrema only, we should add it to pakka lines.

Excerpts from RS:

evaluating an outlier from a distribution.

if two points are at the same distance from the mean of a 2-d distribution, a euclidean distance measure from the mean would term them to be at equal distance from the distribution. However, if the density of the distribution is low in one direction and high in another, then the distance of the two points should not be shown as equal. However, a euclidean distance measure does not respect that. Hence the need for Mahalanobis distance.

Useful links to udnerstand :
http://www.aiaccess.net/English/Glossaries/GlosMod/e_gm_mahalanobis.htm

070809

j in association of lines used for zf should be changed to jbest as j will always give a max value.

zf_lines = [zf_lines z(:,pointlm_count+curr_permline(jbest))]
from:
zf_lines = [zf_lines z(:,pointlm_count+curr_permline(j))]

In Progress: code for lines association first, points next->l1p2

Satish was right. if a line is associated, then its ep_lines value need not be edited.

060809

Code for EKF SLAM based on prediction and data association of lines using endpoints( points associate first and then lines) is done

edit from last code:
to:
zf = [zf z(:,pointlm_count+curr_permline(j))]
from:
zf = [zf z(:,pointlm_count+j)]

Remarks:

association vector zf needs index of line from all features in scan(i.e. z) but j is an index of curr_permline. curr_permline(j) is an index to the actual line in z

SLAM :

Feature extraction from scans

1.Intuition
2.Neural Networks / Training
3.SIFT
4.sensor fusion.
------------------------------