Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Poland.
For the problem of real-time crowd monitoring there is Images of crowds with different densities tend to
an established practice of using extensive closed- present distinct texture patterns. High density crowded
circuit television systems. As routine crowd areas are often made up of fine patterns (which
monitoring is tedious, human observers, responsible correspond to high frequencies in the frequency
“Automatic estimation of crowd density using texture”, International Workshop on Systems and Image Processing, IWSIP’97, May 28-30,
Poland.
given by θ . Similarly, for a fixed value of r , S ( r , θ ) of the crowd density estimation based on texture
measures. The rows of the tables show the distribution
may be considered an 1-D function S r (θ ) , which of the estimation of crowd densities for each group of
gives the behaviour of the spectrum along the circle of crowd density test set. The values of the diagonal
radius r centred on the origin. (printed in bold face) indicate the correct classification
percentages for each group.
Global texture descriptors can be obtained by
integrating such 1-D functions [6]. Therefore, in a As the estimation of crowd density is supervised, a set
discrete case, these descriptors are: of 151 images was used to train the neural network,
π R summing up 299 images.
S( r) = ∑ Sθ ( r ) and S (θ ) = ∑ Sr (θ ) ,
θ =0 r =0
Crowd densities of images of the training and test sets
were manually estimated in advance, in order to
where R is the radius of a circle centred at the origin. establish a comparison standard. Using the manual
For an NxN spectrum, R is typically is chosen as N/2. estimation, the images were separated in groups of
very low density (0-15 people), low density (16-30
For the problem of automatic estimation of crowd people), moderate density (31-45 people), high density
density by using frequency information, spectrum (46-60 people) and very high density (more than 60
values have been used in the new technique to obtain people). Examples of images of such groups are
texture descriptors from the Fourier spectrum (actually, presented in Figure 1.
only 1/3 of the spectrum around the origin is
considered): The results of the estimation based on the statistical
r2 π
GLDM method, presented in Table 1, reached a mean
S ( r1 , r2 ) = ∑ ∑ Sθ ( r )
r = r1 θ = 0
of 82% correct estimation. It is possible to verify that
the results were quite good for all groups except for
θ2 N 3
S (θ1 , θ2 ) = ∑ ∑ S r (θ ) . the one made up of low density crowd images, which
θ =θ 1 r =1 reached only 54% of correct estimation.
Acknowledgements
References