Foreground segmentation using GMM combined temporal differencing

Abstract

Various computer vision applications like biometric identification, analysis of traffic, face detection techniques, video analysis, and surveillance require the use of moving object identification as a fundamental step. A lot of efforts have been made in the past to find approaches which can detect motion but most methods are limited to particular situations and are not applicable everywhere. This paper proposes another, more robust approach towards object detection using Gaussian Mixture Model for background subtraction and temporal differencing for foreground segmentation that provides a promising result with the application of morphological operations and filtering. The GMM approach is a multimodal approach that faces the constraint of outlier pixels because of sudden illumination changes and commonage pixels among object and background. Other limitations of GMM include its learning rate and the number of models. This paper includes use of GMM and information about temporal gradient using temporal differencing for object detection. By use of adaptive GMM along with temporal differencing methods and filtering in post processing results in successful and robust object detection. This paper also compares the approach we have taken to other approaches by comparing the results obtained using a standard video dataset.