Instance segmentation of crowd detection in the camera images

Abstract

Artificial Intelligence (AI) is the new era in Remote Sensing (RS) applications like many other research areas. Convolutional Neural Network (CNN) structures are widely used in supervised and unsupervised classification within deep learning methods. One of the main fields of these deep learning methods is grouped under semantic segmentation applications. Instance segmentation according to the classes that are detected in the images is a type of them. In this study, in the images and real-time frames different deep learning methods are conducted and person class is detected. As data, public camera images, and Unmanned Aerial Vehicles (UAV) images are used. Each object that detected is segmented and the results are shown quantitatively. In the experiments, Mask-RCNN and Yolact++ architectures are used with selected backbones such as ResNet. Also, the time durations of each model's applications are calculated. As total time consuming for each frame, Yolact++ is faster, but the scores are yielded better in the Mask-R CNN model in the experiments. Security, target detection, and metropolitan city vision systems and many other industries are using such as crowd and person detecting applications. It is also expected to increase by the time more and more shortly, as well. © 2021 Elsevier B.V., All rights reserved.