“gq OO «4D (OD FD (D DD DO (D D uc 8 
"*» (D (D US 
m © ao» mp m 7 
= 5 (D UO BO c+ ™ v0 € 0 2 7 
L:-] 
me LB DD un er 
  
  
  
  
  
  
algorithms proposed by (  Fishler 1981, 
Zamperoni 1982, Shu 1989, Ballard 1982 ) all 
can be used to follow lines by some 
modifications. 
Theoritical analysis and experimental 
results indicate that each algorithm has its 
advantages and disadvantages. Segmentation 
algorithm can easily leap over the break 
point (Fig, 3a), but it needs too many 
manual operations when following a multi 
-curved line (Fig. 3b). As for recursive 
following method, it is easy to follow 
lines like Fig. 3b shows, but it is 
difficult to deal with the break point in 
Fig, 3a, Variable radius circular seaching 
method can deal with both situations in Fig. 
3a and Fig. 3b (Fig. 3¢), but it will cause 
some errors when lines occur desentily (Fig. 
3d). On the other hand, the same seaching 
scheme with different criterion ( such as 
grey value criterion, difference criterion, 
local sum of grey values criterion) will 
lead to different results, Therefore, 
there is no ideal following algorithm 
which is suitable for any situation. 
A 
A 
= 
/ 
A 
7 ras IAN 
/ 
» 
a) b) 
e) d) 
Fig.3 Different algorithms 
3.2.3 Multi-criteria line following In view 
of the differences of various algorithms, we 
think that the main way to improve line 
following procedure is to extract more 
information and apply multi-criteria, As for 
  
line following, there are a lot of 
information can be used, such as pixel grey 
value, first order difference, second 
239 
order difference, local sum of grey 
values, curvature and so on. In addition, 
the knowledge about the line shaped 
objects can be used in the mean time. 
The organization mode for multi-criteria can 
be divided into two kinds. One is to 
arrange the criteria to a decision tree 
according to their  strengthenness and 
operating speed. The other is to make a 
unique criterion by a linear function which 
is similar to the consuming function of 
heuristic searching. Suppose that Ci, Ca," 
C. are criteria collected, the unique 
eriterion can be defined as 
C=k,C,+kaC2a+ 358 22 +knCn 
where k;, ks, ..., k are weighted 
coefficients, Desicion tree and unique 
criterion are constructed from a lot of 
experiments and can be adaptively modified 
when the program operates. 
3.2.4 The line following algorithms in the 
system at present A, Fully automatic line 
following (FALF) That is to say, start 
point location and line following are 
automatically performed without any manual 
operation. Two kinds of searching scheme 
are adopted. The first is arc searching (Fig. 
48) , where the angle « and radius R are 
automatically modified in the process, The 
second is recursive following(Fig. 4b). Multi 
-eriteria and decision tree form are adopted 
in each method. FALF is performed in the 
working window defined by the orientation 
points,also it can be peformed in & 
manually defined window, The two 
following methods are used individually 
or integrately. B. Line following with 
manual operation If some lines lose 
through FALF, Line following with some 
manual operation is performed. Manually 
pointing a start point on a line, the 
whole line will be automatically followed by