Preventing Cross Border Infiltration using CNN

Abstract: Motion detection and the subsequent identification of the cause of motion plays a pivotal role in advanced surveillance systems. Cross border infiltrations are increasing rapidly and an autonomous system that identifies and classifies the cause of motion has a tremendous potential to augment the surveillance process. Deep learning methods can be used to smarten such systems to perform under any illumination setting with high accuracy. In our study we have developed a system that identifies intruders in well illuminated as well as in low or no illumination conditions. The system identifies motion in a live video feed by background subtraction and identifies the regions of interest (ROI). Subsequently, a Convolutional Neural Network (CNN) model trained using regular and thermal images classifies the regions of interest and detects if motion is caused by a human intruder. Our system detects human intruders with an accuracy of 96.13%.

Keywords: visual surveillance, background subtraction, motion detection, static background, convolutional neural networks