Changeset 642

Timestamp:

01/31/10 22:10:15 (2 years ago)

Author:

mosttw

Message:

Outline of project plan

Files:

• branches/2010-image-rec/docs/project-plan.tex (copied) (copied from branches/2010-image-rec/docs/lit-review.tex) (1 diff)

branches/2010-image-rec/docs/project-plan.tex

@@ -19 +19 @@
 \end{flushright}
 
-\section*{Project Literature Review}
+\section*{Project Plan}
 
 %\begin{multicols}{2}
 
-\subsection*{Monocular Vision based Indoor Mobile Robot}
-
-\paragraph{Paper:} Takeshi Saitoh Tomoyuki~Osaki Naoya~Tada, Keisuke~Murata and Ryosuke Konishi.
-\newblock {Monocular Vision based Indoor Mobile Robot}.
-\newblock In {\em In The 23rd International Technical Conference on
-  Circuits/Systems, Computers and Communications}, 2008.
-
-\paragraph{Summary:} The paper discusses problems with and methods of navigation applied to an indoor setting. It demonstrates problems with brightness classification within an indoor setting --- how arbitrary lighting is difficult to compensate for. It also examined specific types of data unique to indoor environments that may be useful, such using the color of baseboards to locate the bases of walls to avoid. The authors then take this further, finding interruptions in these lines that represent obstacles, and determining obstacles using sobel edge detectors. They then examines these techniques in practice, discussing experiments performed from a single camera robot.
- 
-\paragraph{Applicability:} As our project involves guiding a robot based on lines and obstacles, their work with baseboards is very similar to how our robot will detect lines painted on the ground. The work where the paper extends these lines to determine a corridor for movement will be applicable to similar instances where our robot must navigate dashed or hidden lines. The work with obstacle avoidance and with high level navigation decision is also very near what our robot will need to do with image data.
-
-\paragraph{Issues:} The work was done with indoor robots, whereas ours must function outdoors.  In our application we can expect a homogeneity of brightness within an image, but not through a sequence of images (as the robot's rotation relative to the sun changes and the camera adjusts exposure), while an indoor robot may find the opposite (reflective surfaces cause hot spots in the image).  This robot also has prior knowledge of its environment, while one following the rules we're given has none.  Finally, the authors primarily consider straight lines, while we must navigate a curved course.
+\subsection*{Report}
 
 
-\subsection*{Appearance-Based Obstacle Detection with Monocular Color Vision}
+\subsection*{Iterative enhancement plan}
 
-\paragraph{Paper:} I.~Ulrich and I.~Nourbakhsh.
-\newblock {Appearance-based obstacle detection with monocular color vision}.
-\newblock In {\em Proceedings of the National Conference on Artificial
-  Intelligence}, pages 866--871. Menlo Park, CA; Cambridge, MA; London; AAAI
-  Press; MIT Press; 1999, 2000.
 
-\paragraph{Summary:} The authors describe a color-based method for detecting obstacles in a robot's field of view.  They propose taking a histogram of the area in front of the robot and using it to classify the remainder of the image as ground or not ground on a pixel-by-pixel basis.  They then augment this technique by integrating multiple histograms over time, permitting the robot to adapt to changing surfaces.  They test their system using a small mobile robot indoors and outdoors.
+\subsection*{Timeline}
 
-\paragraph{Applicability:}  The techniques described are extremely applicable to navigation in the IGVC.  Indeed, the downsides they mention --- such as misclassification of obstacles with color similar to the ground --- don't apply, since the color of all obstacles at the IGVC are very different from the green of the grass.  The use of histograms over time to adapt to changing surfaces will be useful to us when adapting to changing lighting conditions, whether from rotation of the robot or time of day.  The ability to train the robot by briefly driving it around is far superior to manually capturing and classifying exemplars to train the vision system, as is currently done.
 
-\paragraph{Issues:} They only demonstrate their algorithm on relatively smooth surfaces, which is significantly different from the grassy terrain our robot must traverse.  It seems likely that introducing a measure of texture to the algorithm would better-enable it to differentiate between grass and non-grass, as the obstacles at the IGVC are bright, colorful plastic.
 
 %\end{multicols}