Novel extraction algorithms for the background in video surveillance system

Abstract

Background subtraction plays major role in applications like Video Surveillance System, Optical Motion Capture, Gait Analysis and Multimedia. It fully depends on background modelling in which the background is modelled in order to detect the moving objects in applications. The basic background model is acquiring a background image by removing the moving objects. This is very much possible if the background is static. But many times, the background is dynamic and always changing with situations like changes in illumination, objects introduced or removed from the scene. There are so many background modelling methods studied for last few decades and each method has its own merits and demerits in detecting foreground objects from video streams. Still it is challenge for better background model for foreground separation under undesirable scene constraints in video stream and there is a need for new background modelling algorithm. Three novel algorithms for background subtraction in video stream have been developed and tested for standard video data set in this work. They are background model using Gaussian Mixture Model (GMM) with Ant Colony optimization, background model using Recursive Convolutional Neural Network (RCNN) and background model using Multidimensional Recursive Convolutional Neural Network (MDRNN). Novel background subtraction system that uses a deep Convolutional Neural Network (CNN) to perform the segmentation is recently reported in literature. Vi But the same idea with recurrent connections for better performance is investigated in this work. Multidimensional RNN (MDRNN) is a Modified RNN with multiple recurrent connections for multidimensional data is used for background modelling. Tests results using commercial video datasets demonstrate the effectiveness of our system in real-world situations. newline