Segmenting moving object: the MODEST video object kernel

Segmentation is one of the fundamental problems in image processing. Although human beings and most animals perform this task in a relatively straightforward manner, years of research and developments in machine vision have not yet succeeded to match the same performance. The problem of segmentation is difficult not only because of the complexity of mechanisms involved in it, but also because it is ill posed and in this sense, no unique solution exists to segment a scene. In most situations, a priori knowledge on the nature of the problem (or its solution) is needed, often as a function of the specific application in which the segmentation tool is to be used. A segmentation process leads to a partition of an image or a video sequence into regions according to a given criterion. Many of the above-mentioned applications aim at locating moving objects in the observed scene, thus a change detector can naturally drive the segmentation in a more efficient way. Change detection analysis provides a classification of the pixels in the video sequence into one out of two classes: foreground (moving objects) and background.

To this end, we combine a change detector with a background updating technique. On one hand, the change detector is designed to precisely detect object contours and to be robust to camera noise. On the other hand, the adaptive background scheme accounts for slow environmental light changes. The combination of the two (called the Video Object Kernel) allows to automatically detect multiple moving objects in long video sequences recorded by a monocular static camera. The foreground objects identified by the Video Object Kernel are then tracked along time. A successive step tranforms the 2D tracked shapes in 3D shapes. The description of the 3D shapes is finally given to the content understanding module which derives decisions on the observed scene.

A description of the Video Object Kernel is given in "Segmenting moving object: the MODEST video object kernel". The sequences with the segmented moving objects presented in the paper are available below:

Hall Monitor

Group

Highway 1

Highway 2


A description of the complete MODEST surveillance system is given in "Video-Based Multi-Agent Traffic Surveillance System".


The MODEST official home page