Video Crowd Counting Using Developed Deep Neural Networks

Video Crowd Counting Using Developed Deep Neural Networks


Video Crowd Counting Using Developed Deep Neural Networks

نوع: Type: thesis

مقطع: Segment: masters

عنوان: Title: Video Crowd Counting Using Developed Deep Neural Networks

ارائه دهنده: Provider: Nasrin Ranjbaran

اساتید راهنما: Supervisors: Mr. Dr. Moharram Mansourizadeh - Mr. Dr. Hassan Khotan Lo

اساتید مشاور: Advisory Professors:

اساتید ممتحن یا داور: Examining professors or referees: Mr. Dr. Mirhossein Dezfulian and Mrs. Dr. Mah Lagha Afrasiabi

زمان و تاریخ ارائه: Time and date of presentation: Saturday, 27 JAN 2024. 2 P.M

مکان ارائه: Place of presentation: Khanmohammadi Engineer Seminar (Electrical Department)

چکیده: Abstract: Nowadays, due to the expansion of deep learning methods for data processing and particularly using it on machine vision field, many of artificial intelligence researchers intend to develop them for designing more efficient algorithms. In the other side development of 3d acquisition technologies including Lidars and RGB-D cameras provide a good opportunity for processing 3d data. Numerous applications such as autonomous driving, remote sensing and robotic increase attention to 3d data processing and make this field attractive. One of the data type that could represent 3d sphere is point cloud. As it can preserves the original geometric information, is going to be more popular these days. Recently, deep learning methods on point cloud drastically grow. However due to some challenges such as irregularity, sparsity and also being unorder, point cloud processing has its difficulty too. Many methods try to address problems of point cloud processing but one of the most familiar method is PointNet . PointNet, as a pioneer method, achieve remarkable accuracy but it can't extract local features from point clouds. The proposed method is to address not only the impact of local features but also the relative features. The results show it approves the accuracy and is comparable with the state-of-art methods