Object Segmentation Using Background Modelling and Cascaded Change Detection

Abstract

The automatic extraction and analysis of visual information is becoming generalised. The first step in this processing chain is usually separating or segmenting the captured visual scene in individual objects. Obtaining a perceptually correct segmentation is however a cumbersome task. Moreover, typical applications relying on object segmentation, such as visual surveillance, introduce two additional requirements: (1) it should represent only a small fraction of the total amount of processing time and (2) realtime overall processing. We propose a technique that tackles these problems using a cascade of change detection tests, including noise-induced, illumination variation and structural changes. An objective comparison of common pixelwise modelling methods is first done. A cost-based partition distance between segmentation masks is introduced and used to evaluate the methods. Both the mixture of Gaussians and the kernel density estimation are used as a base to detect structural changes in the proposed algorithm. Experimental results show that the cascade technique consistently outperforms the base methods, without additional post-processing and without additional processing overheads.