Abstract
Crowd counting has become necessary in the age of information technology
and network development; with many different goals, people have needed
more services to meet their work. Therefore, crowd counting is also
necessary for daily life, such as solid classrooms, vigorous meetings,
conferences, etc. Based on practical needs, we propose counting the
number of people in an image, known as crowd counting. Using several
software techniques, we found that DM-Count outperformed previous
advanced methods on two large-scale counting datasets by a wide margin
in Mean Absolute Error. Furthermore, the VGG16 model for the algorithm
does not certainly use the current
VGG19 and still shows good higher performance. We analyze multiple
people marked with dots in each training image. Existing crowd-counting
methods must either smooth each annotated drop with a Gaussian
distribution or estimate the likelihood of each pixel given the
annotated point. This paper proposes a crowd-counting model using deep
learning and Computer Vision to improve performance by handling
perspective distortion by intelligently using many features generated
during coding. The use of Deep Learning and Computer Vision
experimentally shows that our method can improve up to 34.91% Mean
Absolute Error and 26.25% of Root Mean Squared Error compared to the
referenced methods.