Intelligent target tracking on multi video images

ABSTRACT dummy Intelligent Target Tracking on Multi Video Images Sevinç AY Ph.D. Thesis FIRAT UNIVERSITY Graduate School of Natural and Applied Sciences Department of Software Engineering July 2023, Pages: xii + 96 In recent years, with the developments in the field of technology, the use of visual surveillance systems consisting of camera, data storage devices and software systems has become widespread. These systems, which are used for visual surveillance, are widely preferred in observing the movements and behaviors of objects or people. Thanks to various algorithms in the surveillance systems used in traffic, it is possible to detect the abnormal behavior of suspicious persons or vehicles among the crowds. Accurate detection of these movements helps to detect and prevent many dangers that may occur. The need to use visual tracking systems in many areas of daily life has increased the interest of researchers in object tracking in recent years. Object detection and object tracking is increasingly used for the interpretation of visual surveillance systems. Within the scope of the thesis study, multiple video images taken from more than one camera located on a road route were used. In the first stage of this thesis, a vehicle, which is considered suspicious, was monitored using the vehicle tracking method developed on multiple traffic video images. In the second stage, it was ensured that the cameras on which the vehicle could be viewed were determined. In the last stage, the suspicious vehicle was detected by using deep learning-based object recognition algorithms. In the thesis study, Region-Based Convolutional Neural Network (CNN), Faster Region-Based Convolutional Neural Network (Faster CNN) and You Only Look Once (YOLO) were used. The mean average precision (mAP) obtained from the experimental results were compared and it was stated that the 89% value obtained by training the Faster CNN deep learning model with the ResNet101 backbone network was the highest mean. Keywords: Deep Learning, Vechile Tracking, Vechile Detection, Fast RCNN, Faster RCNN, YOLO