Versatile Defacement Detection by Monitoring Video Sequences Using Deep Learning

Arivazhagan S,Newlin Shebiah R

doi:10.24018/ejers.2019.4.7.1396

Abstract

The main objective of this paper is to detect vandal and vandalism by monitoring recorded video sequences. Vandalism is one of the most commonly occurring crimes in the society that indirectly affects the economy of the country. The proposed algorithm takes in the input from the video extracted from surveillance camera which prevails in public places. Further, it is converted into frames and subtracted with the background to detect the foreground object. The background subtracted image contains both human and non-human moving objects. In order to differentiate human pixels and other moving objects in the video sequence, discriminative features are extracted using deep architecture and classified using SVM classifier. Deep features proved to be highly discriminative when compared with the handcrafted Histogram of Oriented Gradients features. By analyzing the dwell time of the person in the restricted scene and his motion pattern with time and significant change in background vandalism act is declared and the person is considered as vandal. The proposed method was evaluated on the videos collected from You Tube with the contents taken during night time, multiple vandals, car vandals etc.

Highlights

In recent days, the tremendous increase in the destruction of government buildings, littering the tourist attractions, damaging the traffic signs, defacing the statues and scribbling on historical monuments by the public has encouraged more research into the field of vandalism detection
It becomes difficult to differentiate graffiti and street art. 2) the underlying difficulty in vandalism detection is due to its application dependency since it changes for every social context. 3) deficient of real vandalism video sequences and Published on July 19, 2019
Datasets to work with. 4) the act of vandalism is complex since it mostly occurs at night and it occurs at a faster rate

Summary

Introduction

The tremendous increase in the destruction of government buildings, littering the tourist attractions, damaging the traffic signs, defacing the statues and scribbling on historical monuments by the public has encouraged more research into the field of vandalism detection. The vandalism detection in video surveillance becomes a difficult task because of the following challenges: 1) the difficulty in distinguishing between the normal behavior and vandalism activity in the restricted region. 2) the underlying difficulty in vandalism detection is due to its application dependency since it changes for every social context. Subtraction is a technique which allows the image’s foreground to be extracted for further processing. It is a widely used approach for detecting moving objects from static cameras. The most basic technique of background subtraction is through basic motion detection via temporal median filter put forward by Q. Kernel Density Estimation (KDE) is used to extract the foreground objects. Current Frame a) Normal behavior (Dwell b) Vandal activity (Dwell time time analysis for the frames analysis for the frames 1543 to 1485 to 1505) 1576)

Objectives

Results

Conclusion