(a) (b) (c)
Figure 11. Parameter Coordinate Clustering Analysis of Lines in Hough
Space
2.3 Determine the optimal adaptive threshold by counting the
number of
votes
To ensure that there is no missed detection during Hough transform, it
is necessary to detect straight lines at low voting thresholds. However,
the detection results at low voting thresholds are often affected by
interfering pixel points, resulting in the detection of many interfering
straight lines. Under this premise, the straight line fitting results
obtained by k-means clustering often have significant errors, and in
extreme cases, the detection method may even fail. Based on the above
situation, this article proposes a method for determining the optimal
adaptive threshold . After the clustering analysis is completed, this
method calculates the distribution of voting numbers, degree of
dispersion, and the relationship between the mean and standard deviation
of the calculated voting numbers of coordinate points in the parameter
space. By observing the distribution of these data points, the optimal
adaptive threshold is set. Specifically, the low threshold ensures that
Hough transform does not miss detection, and k-means clustering is
performed on the data points in the parameter space detected by the low
threshold, ensuring that there are four straight lines in the fitting
result. Calculate the mean and standard deviation of the number of
straight line votes in the i-th cluster clustering results, and set. The
value of is generally determined based on experimental conditions and is
generally within the range of [0,2]. Using the piecewise function to
place the element in the i-th cluster that satisfies into a new set.
After setting an adaptive threshold for each cluster result, 4 new
clustering results will be generated,The number of votes in these four
new clustering results is all high, which means that the interference
lines detected under low thresholds have been filtered out. Figure 12a
shows that when m=1, there are many lines in the first cluster that are
lower than the green dashed line, all of which are interference lines,
Figure 12b shows the distribution of votes when m=2. Based on the
experimental analysis, this article takes m=1. From the experimental
results, it can be seen that the method proposed in this article for
determining adaptive thresholds based on the number of votes is to
identify the lines with high voting numbers in each cluster, preparing
for the improved clustering centroid calculation and line fitting method
proposed in Section 2.4.