Changeset 694

Timestamp:

02/21/10 23:50:46 (2 years ago)

Author:

mosttw

Message:

Report checkpoint

Files:

• branches/2010-image-rec/docs/report.tex (modified) (3 diffs)

branches/2010-image-rec/docs/report.tex

@@ -48 +48 @@
 \begin{figure}[htb]
         \centering
-                \includegraphics[width=0.25\textwidth]{figures/robot.jpg}
-        \caption{Moxom's Master, the Rose-Hulman Robotics Team's IGVC entry.}
+                \includegraphics[height=2in]{figures/}
+                \includegraphics[height=2in]{figures/robot.jpg}
+        \caption{Examples of IGVC terrain, and Moxom's Master, the Rose-Hulman Robotics Team's IGVC entry.}
         \label{fig:moxoms-master}
 \end{figure}
@@ -71 +72 @@
 \section{Process}
 % Also hit on key challenges here.
-Our process to get from camera images to information useful for navigation requires several steps
-of processing and reprocessing. Our first step is to find an aproximation of texture using a sobel 
-filter. We then use a color histograming classifier, based on an assumed safe area of grass, to 
-classify pixels in the image as grass and not grass. The texture data from the sobel is combined 
-with the data of which pixels are not grass to determine which are line and which are obstacle. 
-The line pixels are further processed using a hough transform to find the actual lines of the course.
-The objects are then tranformed into real world coordinates by performing a perspective 
-transform on the bottom pixels of the objects.
+Our process to get from camera images to information useful for navigation
+requires several steps of processing and reprocessing. Our first step is to
+find an approximation of texture using a Sobel filter.  We then use a color
+histogramming classifier, based on an assumed safe area of grass, to classify
+pixels in the image as grass and not grass. The texture data from the Sobel is
+combined with the data of which pixels are not grass to determine which are
+line and which are obstacle.  The line pixels are further processed using a
+hough transform to find the actual lines of the course.  The objects are then
+transformed into real world coordinates by performing a perspective transform
+on the bottom pixels of the objects.
+
+\subsection{Generation of HSV histograms}
+The first step in classifying pixels is to convert the image to the HSV color space.  The image is then scaled down to quarter size, as at full five-megapixel resolution colors are more noisy than when scaled down.  
 
 \subsection{Sobel filter for texture}
+A Sobel filter with an aperture size of 5 is then applied to the full-resolution value band of the image.  The filter is applied at full resolution to avoid throwing away any edge information while resizing.  The operation takes approximately 0.1 seconds in OpenCV, so it appears to be viable to do in real time.  The full-resolution filtered image is then downscaled to quarter-size and blurred, producing an image that indicates the ``texture'' in each pixel's region, as illustrated in Figure \ref{fig:texture}.
 
-\subsection{Classification on Color analysis}
+\begin{figure}[htb]
+        \centering
+                \includegraphics[width=0.5\textwidth]{figures/texture.png}
+        \caption{The original and blurred Sobelized value band, used as a measure of texture.}
+        \label{fig:texture}
+\end{figure}
 
 \subsection{Final Pixel Classification}
+Finally, 
 
 \subsection{Hough Line Transform}
@@ -107 +120 @@
 
 \section{Future work}
-The next step with our algorithm is to make it work on cases where it fails, those with lines 
-appearing within our safe zone. There is also work to be done for it run on the robot. The 
-algorithm must be able to run on images at a speed for live decision making. The algorithm
-must be moved from our testing matlab to opencv.
+The next step with our algorithm is to make it work on cases where it fails,
+those with lines appearing within our safe zone.  It is possible that this
+could be done by excluding portions of the ``safe zone'' which exceed a
+threshold in brightness, or to apply the previous image's histogram to the safe
+zone and exclude those pixels that fall far outside it. There is also work to
+be done for it to be run on the robot. The algorithm must be able to run in
+real time to support live decision-making, which means that it must be ported
+from our Matlab proof-of-concept to OpenCV.
+
+\section{Conclusion}