객체 검출과 객체 분할 방법의 무인 감시 데이터셋 적용 결과 비교

Abstract

Visual surveillance aims to detect foreground objects stably under diverse variations caused by weather changes. Traditional visual surveillance algorithms are based on a statistical analysis of pixel variations along the spatio-temporal domain. Recently, deep learning-based algorithms have improved performance compared to conventional algorithms. In this paper, we investigate the performance of object detection and semantic segmentation algorithms on visual surveillance datasets. We use the CDnet dataset, which is widely used in visual surveillance. We adjust labels in the CDnet dataset to be suitable for object detection and semantic segmentation. We investigate the possibility of object detection and semantic segmentation in visual surveillance. Two types of algorithms based on CNN and transformer are used in experiments. Experimental results show that spatio-temporal processing is required to improve performance when we apply object detection and semantic segmentation in visual surveillance.