Paper Key : IRJ************850
Author: Chand Fathima
Date Published: 04 Mar 2024
Abstract
Crowd counting remains a critical research domain in computer vision, where the accuracy of existing methods, particularly multi-column convolution neural networks (MCNN), encounters challenges in scenarios with uneven crowd distributions. Addressing this, our paper integrates crowd global density features into the MCNN framework using cascaded learning. Additionally, we propose an enhanced MCNN architecture featuring max-ave pooling and deconvolutional layers to preserve detailed features lost during down-sampling. Experimental validations on UCF_CC_50 and ShanghaiTech datasets demonstrate superior accuracy and stability of our approach.Furthermore, we review the evolution of crowd-counting techniques, emphasizing the transition from traditional handcrafted methods to intelligent machine-learning-based approaches, particularly convolutional neural networks (CNNs). Despite facing challenges like occlusion and clutter, CNNs offer promising solutions for intelligent crowd counting and analysis, facilitating adaptive monitoring and management of dynamic crowd gatheringsOur study also discusses the current challenges faced by crowd counting systems, especially those employing density estimation, such as perspective distortion and ineffective population density determination methods. To address these issues, we propose leveraging CNNs for crowd counting through detection, clustering, and regression techniques, promising more robust results by leveraging CNN's advanced image processing capabilitiesGiven the increasing importance of video surveillance for safety and traffic control, our findings underscore the significance of advancing crowd counting methodologies, particularly leveraging CNNs, to meet evolving surveillance needs and ensure accurate monitoring in dynamic environments.