FYP progress for week 3 & 4

Overview

For the past two weeks, the main focus was on trying to understand the various approach in solving SLAM problem.
The book borrowed from Jinqiang turned out to be extremely useful!

Literature summary:

Particle Filter Based Robust Simultaneous Localization and Map Building for Mobile Robots
This article discribed an approach to detact and fix gound robot kidnap problem in real-time SLAM.
Basically, the author was monitering the residule from odometry readings. When it is likely that fault has occured. The particles are mutated in order to retrack the position of the robot.
The artile is short and does not contain much details.

Noval approach to nonlinear/non-Gaussian batesian state estimation
This article discribes the math and thought behind particle filter and proved particle filter performs better than EKF in certain cases, where it's nonlinear or non-Gaussian.

A Method for Registration of 3-D shapes

This paper discribes in detail the ICP algorithm, however, I have not get the time to finish this paper, before I find the book below is much better.

Probabilistic Robotics

Chapter 5 robot motion

This chapter dealt with motion model of robot. Two approach was given in the book, one was based on controll input, the other was based on odomery.
However, I felt that the two approach given in the book was mainly meant for ground robot, however both could be related to our case.

• The odometry model could be related to the optical flow model, whic is being developed by Wang Fei.
• The control model for a UAV is way more complex than the one for ground vechicle. In order to make the model usable for 2-D SLAM, we have to find a way to approximate the 3-D motion into 2-D. It might also be a problem that the heading of UAV might not be the direction that the UAV is actually propagation. Extensive research in UAV control and modelling need to be done in order to obtain a control-based motion model.

Chpter 6 Robot Perception (measurement model)

This chapter introduced a few types of measurement model that could be used for laser sensor.
In my point of view, establishing an accurate measurement model is criticle for the rest of the task. (especially after seeing Bo Xuan's work).
The model introduced in the book could be summarized as follows:

• Beam model for Range finders is a model that could be justified with rigour and tuned systematically. In the model, there are 4 factors that could contribute to a measurement reading, each of which can be represented by a probability distribution. They are:
• hit: the intended measurement with a narrow gaussian distribution
• short: obstruct by unintented object
• max: max reading returned, due to lighting condition, color of object, reflection etc.
• random: other factor contrbute to a false reading
• The wrights of these four factors and their intrinsic parameters could be tuned.
• The shortcomes of this approach are:
• Expensive computation involved
• Dependency amoung beems, which are assumed to be independent
• Lack of smoothness (trouble caused by objects like table leg, chair etc.)

• Likelihood Fields for Rnage finder is a model only deal with the end point of the measurements(except for the max readings). This method calculate the position of the end point w.r.t. global map using translation matrixes and adjust the probability of encountering an obstacle at the end point. This approach deals with 3 out of 4 factors in the Beam Model
• hit
• max
• random
• The disadvantages of this approaches are:
• It does not model dynamically moving objects (short category)
• The algorithem does not deel with problem of going through obstacles
• It does not takes into account the matching of free space
• While comparing to Beam model, this approach does better in smoothness

• Correlation-based measurement model is not really a measurement model. It's more a method that matches two local maps and produce parameters like distances between maps.
• The advantage of map mapping is that it considers white spaces in the map. But the dis advantage of this approach are there is no physical explanation behind and there is no noise charicteristic of the sensor.

• Feature-Based measurement Model is what Bo Xuan has done (more or less). It has to do feature extraction first before the data could be processed to yield measurements. Being at a higher level, features extracted are way lower in number. Therefore, the computational cost for this approach is much lower.
• The main challenge in feature-based measruement model lies in data association steps. We have to find a way to extract a signature of every feature, so that it could be checked to confirm the association of data. In order to do this, Lidar alone will not be enough.

Chapter 7 Mobile Robot Localization: Markov and Gaussian

This chapter introduced, in detail, two similar approach (EKF and UKF) to realize localization task.
Both EKF and UKF assumes a gaussian distribution for measurement noise. Their differece lies in the different ways to linearize the measurement/ motion model. UKF usually produce better result at the cost of more computation power.
The main draw back of both approaches are that they were subject to global registration and kidnap problems.
Author also briefly mentioned another approach- multi hypothesis tracking algotrithm (MHT), which I think is the most promissing amoung the gaussian filter approaches. It is more robust, while still seems practicle to be implemented in real time.

Chapter 8 Mobile Robot localization: Grid and Monte Carlo

This chapter introduced mainly two approach to achieve the localization problem, and they are grid (histogram filter) and Monte Carlo.
They are similar in a sence that they both try to represent the probability distribution by having a sample of states.
However, Monte Carlo is more efficient as compare to Grid based approach, especially after adopting the KLD-sampling approach.
As compare to Gaussian filter based algorithms, these two approaches solved the global registration and kidnapped problem at the cost of heavier computation. But, if implemented properly, the Monte Corlo approach seems promising to be run in real-time.

Goals for now

Following should be done before CA1:

• Finish the SLAM related chapters in the book
• Look back into the papers that has been found and categorize them according to as the book suggested
• Try finding some more articles through the references in the book