The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bib. Beijing 2008 
direction parameters. The aspects of accuracy and efficiency 
have been improved greatly. Operating steps can see the 
following chart: 
Figure 3 Flowchart for Road Extraction 
3.1 Edge Detection(Rafael C., 2004) 
In order to apply the method of phase classification effectively, 
the step of edge detection is very important. Enough edge 
information can be useful to the work of road information 
extraction. 
Till now, there have been many operators(Chen Y. 2003) for 
edge detection, such as Robert, Log, Sobel, Canny and so on. 
Analyzing these existed methods, we decide to use Canny to go 
on with the work of edge detection as well as the method of 
grey morphology. 
3.3 Determination of Initial Point on Road 
According to the method described in chapter 2, calculate the 
phase information of the edge image, then combine this 
information with the gradient scope, group the edge points, 
finally fit the edge line using the method least square, which 
can we get the straight lines and the curve lines, those are all 
road edge lines. Taking into account the shadow shielding near 
the road, we can judged the initial point of the road by the 
following criteria,. 
1) The width of the road at this point should be consistent with 
the given width of the road. 
2) The grey value at this point should be consistent with the 
given value more or less. 
3) Search for eight neighbourhoods of this point. Except for this 
point, if pixels in its eight neighbourhoods contain one edge 
point, this point can be seen the starting point of the road. It can 
be interpreted by the following chart: 
Figure 4 Search the Eight Neighbourhood of the Starting Point 
3.4 Track binary image to acquire supported region for 
road edge 
From the lower left of the image, search binary edge image line 
by line. From the starting point, calculate its gradient direction, 
mark the gradient direction of this pixel values 
as *(* = 0,1,2,— ,7) ^ and t hen search its eight 
neighbourhoods. If there exit pixels that belong to the same area 
named ^ , put this neighbour point into the same road line with 
current pixel. In the computer, we can note it into an array to 
store safely. Arrange the new point as the current one, do the 
same search work in the way described above until all the 
pixels in the marked area have been detected. All the points 
searched belong to one array. Similarly, search the total edge 
image to account points of the same type in each area of road 
into a linked list in computer. 
Firstly, with the knowledge of grey morphology (Zhu et al. 
2004), the image will be corroded and inflated once. In this way 
the roads that are fuzzy can be improved. Meanwhile the 
information of the road’s edge can be strengthened. Secondly, 
improve the Canny operator, take advantage of the double 
thresholds to detect the light edge parts and the dark parts. 
Based on the principle of non- local maximum suppression, we 
can get the edge information in detail. Finally, thin the edge line 
with the method of skeleton extraction in the area of 
morphology, which can we get the road edge line of single pixel. 
By this step of edge detection, the edge information could be 
preserved most which is useful to do the following work. 
3.2 Grouping Road Edge and Re-fitting Line in Image 
According to the phase and amplitude information accounted, 
group the edge of the road edge lines, and then fit them into 
straight lines and curves. 
3.5 Improve the method of phase classification to fit the 
road line 
In the 3.2.2, we described the traditional method of phase 
classification to get the edge line of the road. Now some points 
of the method have been improved in order to get various kinds 
of road line in image. That is, fit the three adjacent points into a 
line by least squares method respectively. Calculate each value 
of slop from every three adjacent points, namely a ' > a 2 > a 3 . 
Set a threshold, namely: Tl. From many examinations, here we 
can define Tl to be 2 .if the calculated slopes meet the 
1 
condition 
0 < a , 
2 at the same time, and we define 
them as the same road points. 
In the corresponding position in original image, draw the red 
line to connect detected three points. If the point does not 
conform to the rule, remove it. Take the same steps to the 
following groups until it reaches the end. 
511