Robust and accurate change detection under sudden illumination variations Luigi Di Stefano Federico Tombari Stefano Mattoccia Errico De Lisi Department of Electronics Computer Science and Systems (DEIS) Viale Risorgimento 2, 40136 - Bologna, Italy Advanced Research Center on Electronic Systems (ARCES) Via Toffano 2/2, 40135 - Bologna, Italy University of Bologna {ldistefano, ftombari, smattoccia}@deis.unibo.it, errico.delisi@studio.unibo.it Abstract A challenging task for most change detection applica- tions is to accurately segment the foreground from the back- ground under the presence of heavy illuminations changes. In fact, photometric changes can be easily misinterpreted as structural changes, thus leading to false positives in the change mask. In this paper we present a novel approach to deal with this problem, which deploys a recently proposed robust visual correspondence measure together with a tonal registration procedure. Furthermore, we provide an exper- imental comparison of our approach with other proposals specifically designed for the task of being robust to sud- den illumination changes, demonstrating the effectiveness of our approach. 1 Introduction Change detection aims at detecting structural changes occurring in time in a scene by analyzing a frame sequence. This is a key task in most advanced video-surveillance applications, for the mask highlighting changed pixels (change mask) typically represents the input data to higher level vision algorithms.This is the case of traditional single- view as well as more recent and advanced multiple-views systems [1, 6]. The most common change detection ap- proach is referred to as background subtraction: given the current frame, F , and a model of the background of the scene, B, the change mask is obtained as the difference between F and B. This approach assumes that the back- ground model is available or can be obtained by processing a short sequence of frames at initialization time (e.g. as shown in [5]). A wide variety of change detection algo- rithms has been proposed in literature, so as to address is- sues such as illumination changes, camouflage and vacillat- ing background. A recent survey providing good coverage of this research area is given by [9]. A major issue for most practical change detection appli- cations is represented by sudden illumination changes oc- curring in the scene. Properly dealing with such a problem is a challenging task for change detection algorithms since the resulting photometric variations can be easily misinter- preted as structural changes, leading to many false positives in the change mask. Algorithms [3, 8, 11] that specifically aim at detecting changes robustly with respect to sudden il- lumination variations typically take the decision of voting a pixel as changed or unchanged based on a given spatial support (e.g. a 3 × 3 or larger window centered at the pixel under evaluation). Typically such algorithms rely on a para- metric (e.g. linear [3, 8]) or non parametric (e.g. order pre- serving [11]) model for the false image changes due to sud- den illumination variations. However, it is well known that such algorithms suffer from an aperture problem, i.e. they cannot detect as changed the pixels belonging to untextured foreground regions. As a result, they typically enable to detect the borders of foreground objects but not accurately their interior parts. Moreover, the use of a spatial support rather than pointwise background subtraction implies inac- curacy as regards localization of the borders of the detected foreground objects. Coarse-to-fine approaches such as [2] can alleviate these problems. In this paper we propose a novel approach aimed at ob- taining robust and accurate foreground segmentation under sudden illumination variations. The paper is structured as follows. Section 2 describes the proposed algorithm along its various stages, then Section 3 discusses the computa- tional requirements dealing with the implementation of the proposed approach. Section 4 shows some qualitative ex- perimental results as well as a comparison with other ap- ACCV'07 Workshop on Multi-dimensional and Multi-view Image Processing, Tokyo, Nov., 2007 MM-P-02 103