S. Singh et al. (Eds.): ICAPR 2005, LNCS 3687, pp. 653 – 662, 2005. © Springer-Verlag Berlin Heidelberg 2005 Meeting the Application Requirements of Intelligent Video Surveillance Systems in Moving Object Detection Donatello Conte 1 , Pasquale Foggia 2 , Michele Petretta 1 , Francesco Tufano 1 , and Mario Vento 1 1 Dipartimento di Ingegneria dell’Informazione ed Ingegneria Elettrica, Università di Salerno Via P.te Don Melillo 1 I-84084 Fisciano (SA), Italy {dconte, mpetretta, ftufano, mvento}@unisa.it 2 Dipartimento di Informatica e Sistemistica, Università di Napoli “Federico II”, Via Claudio 21 I-80125 Napoli, Italy foggiapa@unina.it Abstract. In a video surveillance system, moving object detection is the most challenging problem especially if the system is applied to complex environ- ments with variable lighting, dynamic and articulate scenes, etc. Furthermore, a video surveillance system is a real-time application, so discouraging the use of good, but computationally expensive, solutions. This paper presents a set of im- provements of a basic background subtraction algorithm that are suitable for video surveillance applications. Besides we present a new performance evalua- tion scheme never used in the context of moving object detection algorithms. 1 Introduction Video surveillance applications need to work in the absence of detailed a priori knowledge about the objects of interest, and this reason makes preferable the use of segmentation algorithms working without models. These algorithms, usually, try to segment the frame of the video into two regions: foreground (pixels belonging to the objects of interest) and background. In a second phase the foreground pixels are grouped to determine the blobs representing the objects. In video surveillance sys- tems, background subtraction is the most used approach for the object detection step. Frequently in literature background and reference image are synonymous. The basic idea is to obtain the foreground region comparing the current image to a reference image. The background pixels can be either represented by a single color value [9] or by a probabilistic distribution. In [6] the authors use a uniform distribution; this choice is effective only if the background model is always perfectly synchronized with scene changes. Alternatively, in order to reduce the sensitivity to the variation of the light conditions or to mitigate waving tree problems (they occurs when part of the background of the scene is detected as object of interest because it is performing little movements), a simple statistical model is used introducing a Gaussian description of the background pixels [15]. Although this solution mitigates errors due to a not per- fectly synchronized reference image, on the other side it produces a system less sensi- tive in the regions where a great variance of colors has been calculated (also for the detection of the objects of interest). To avoid this loss of sensitivity, a more