Improved Eigenbackground method for dynamical background modelling

Abstract

Foreground object segmentation is an essential step in many video understanding and image processing algorithms, especially for video surveillance tasks. Background modelling and subtraction is a popular and basic approach to segment out foreground objects. It can significantly affect the performance of high level video analysis tasks like object tracking and event analysis. The celebrated Eigenbackground model using Principal Component Analysis (PCA) has been proved to be an efficient and effective approach for static background modelling. However, its performance degrades greatly when background model changes due to various reasons. In this project I concentrate on implementing an algorithm to solve the most common background change in surveillance videos --villumination variation. Based on the original PCA method, an adaptive PCA approach is implemented to update background feature space and adapt to background variations. The adaptive algorithm is implemented using C++ with Open Source Computer Vision (OpenCV) library.