Go Straight, Turn Right:

Pose Graph Reduction through Trajectory Segmentation using Line Segments

Yasir Latif and Jose Neira

SLAM : A mature problem

SLAM is a mature problem
- Solved (NO!)
- Least Square formulation
Current Challenges
- Do it Faster!
- Robust SLAM
Where does speed up come from?
- Sparse structure of the problem
- Architecture specific instructions (SSE2/SSE3)
- Problem specific data-structures (SLAM++)
- Very good open-source libraries for SLAM (iSAM/g2o/GTSAM etc)

Problem: Lifelong (t → ∞) SLAM

SLAM over days, weeks, years, decades(?)
- Increasing amount of data to deal with
- other problems (dynamic objects/loop closing)
Sensor dependent map / Information sampling
- When to add new information to the Pose Graph?
- Are all robot poses equally informative?
- How much information is needed to generate a good map estimate?

What is interesting?

Go Straight, Turn Right
- Turns are interesting (more informative)
- Going straight is boring
SLAM in general is a non-linear problem
- unless you are travelling along a straight line
Local information is reliable
- local errors are small but accumulate over time and cause drifts
- local information is correct

Main Idea

Reduce the pose-graph by discarding poses that lie (approximately) on a line
- (x,y,t) : A line segment starting at (x,y) extending in the direction t
- As long as the motion is in this direction, ignore the incoming sensor data
- when sufficient deviation has occurred, introduce new poses
- What is sufficient deviation?
- Which new poses get added?

Algorithm

      For every new pose being added
	Update line to end at this pose
	make sure that all poses are with a distance Dmax
	if not:
	  find the pose (i) with the largest distance
	  terminate line at pose (i)
	  keep the two poses (start,end)
	  current start at pose (i+1), current end at this pose

Results

Computationally inexpensive
- works with just local information
- compares poses since start of last line
- online operation
Great reduction in number of poses
- Bicocca: ~ 86% (at 5cm)
- New College: ~ 76% (at 5cm)
Benefits
- ~10x speedup compared to full graph optimization
- Nearly as accurate (RMSE error 3 cm)
- Consistent Uncertainity Estimates

Where can it be used

Resource limited devices (Raspberry Pi/Cell phones)
Computing the initial guess
General speed up

Go Straight, Turn Right:

SLAM : A mature problem

Problem: Lifelong (t → ∞) SLAM

What is interesting?

Main Idea

Algorithm

Results

Where can it be used

Demos

Thank you