期刊名称：International Journal of Interactive Multimedia and Artificial Intelligence
印刷版ISSN：1989-1660
出版年度：2022
卷号：7
期号：4
页码：112-120
DOI：10.9781/ijimai.2022.01.002
语种：English
出版社：ImaI-Software
摘要：Video surveillance is one of the important state of the art systems to be utilized in order to monitor different areas of modern society surveillance like the general public surveillance system, city traffic monitoring system, and forest monitoring system. Hence, surveillance systems have become especially relevant in the digital era. The needs of the video surveillance systems and its video analytics have become inevitable due to an increase in crimes and unethical behavior. Thus enabling the tracking of individuals object in video surveillance is an essential part of modern society. With the advent of video surveillance, performance measures for such surveillance also need to be improved to keep up with the ever increasing crime rates. So far, many methodologies relating to video surveillance have been introduced ranging from single object detection with a single or multiple cameras to multiple object detection using single or multiple cameras. Despite this, performance benchmarks and metrics need further improvements. While mechanisms exist for single or multiple object detection and prediction on videos or images, none can meet the criteria of detection and tracking of multiple objects in static as well as dynamic environments. Thus, real-world multiple object detection and prediction systems need to be introduced that are both accurate as well as fast and can also be adopted in static and dynamic environments. This paper introduces the Densely Feature selection Convolutional neural Network – Hyper Parameter tuning (DFCNHP) and it is a hybrid protocol with faster prediction time and high accuracy levels. The proposed system has successfully tracked multiple objects from multiple channels and is a combination of dense block, feature selection, background subtraction and Bayesian methods. The results of the experiment conducted demonstrated an accuracy of 98% and 1.11 prediction time and these results have also been compared with existing methods such as Kalman Filtering (KF) and Deep Neural Network (DNN).
