Recently I was largely occupied by
projects and labs from other modules, therefore the progress for FYP is not as
fast as previous a few weeks.
I found the following articles
useful and summarized them as below.
Simultaneous Localization and
Mapping for Forest Harvesters
This article is based on a forest
harvester application that is done in Finland. The environment is extremely
similar to the Out-door project that Jinqiang is doing (I guess).
The main sensor for this mentioned
project is laser scanner. They made use of a graph data structure to store the
measured information and run multiple data association algorithms against the
obtained data on batch basis.
They also brought up a few
interesting points in the article:
1.
They
claimed that the image beacons does not help in the data association problem.
The season is simply that trees in forests looks alike. They claimed that this
is extremely true in well-maintained forests.
2.
The
proposed the idea (not realized though) that measuring the absolute angle would
help in the data association problem.
a.
I
think magnetometer would help in geometry based data association, if the sensor
was able to produce measurements of absolute direction with reasonable
accuracy.
b.
I
am not sure whether the motors from the helicopter would interfere with the
magnetometer.
3.
They
are using GPS in their project. Since their tests are also carried out in
forests, is it the case that GPS data is actually available in forests.
Towards Lazy Data Association in
SLAM
This paper took a second look at a
basic assumption that most maximum likelihood data association algorithm
adopts, which is all pervious decisions made were correct.
Data Association
In
this proposed approach, author maintained a tree-like log-likelihood data
structure. Each time a new data association decision is made, all frontier are
re-examined to see whether modifying one of the pervious decision would result
in greater likelihood. If so, the upper level to the modified leaf would also
be checked recursively, until a maximum is obtained. In the process, all
log-likelihood data are maintained in the tree data structure for later use.
SLAM
The
SLAM part of this algorithm is similar to the Sparse Information Filter SLAM.
All measurements are finally translated in to soft constrains and the mapping
task becomes an optimization problem.
Computation
considerations
If
this idea is implemented on feature bases, the memory and computation usage
would be very expensive. Therefore, I don’t think this method would be able to
run on-board.
The
lazy nature also make this approach hard to meet real-time requirement.
Whenever a modification of previous data association is made, the search might
be substantial. Therefore the computation time variance from frame to frame is
large. In such case it’s easy to hit scheduling deadline, if run in real-time.
In short, the advantage of this
method is the modification solved the root of brittleness for data association
problem, which is the assumption that previous decisions are always right.
The disadvantage of this approach
is that, the lazy nature of the tree search and the large amount of computation
for inverting information matrix make this approach hard to implement for
real-time.
